test(evals): measure Java code intelligence as the user sees it

`pnpm eval` drives the production workbench in Chromium against a release `openidea-server` and jdtls, and judges each scenario on what the user sees: time to feedback, to the first result and to the final one, and the results against `evals/java/golden/`. Scenarios cover navigation, usages, implementations, gutter markers, code lenses, diagnostics, completion, quick documentation and rename, in `fixtures/java-app` and a pinned spring-petclinic, warm and cold. The report (`target/eval/report.json` / `.md`) names each failure's kind, a repro command and the server commands in the slowest run's window; `--save-baseline` compares later runs. - The frontend gets `data-*` hooks the scenarios read: usage / chooser rows carry their location, the usages popup its loading state and scope, a hint whether it is progress and what it answers. - `smoke/tauriShim.ts`: the in-page Tauri runtime shared by the jsdom smoke tests and the evals; `perf/lib/bundle.ts` builds the perf bundle for both. - `.github/tool-versions.env` pins the language-server versions for Language smoke and the evals. - `.agents/skills/eval`: the measure → fix → prove loop for agents.
fix(linux): keep the first-map hint under Mutter's auto-maximize share (#2520 )
2026-11-11 20:25:41 +02:00 · 2026-11-11 22:39:59 +08:00 · 2026-11-11 22:29:05 +08:00 · 2026-11-11 22:18:48 +08:00 · 2026-11-11 22:18:48 +08:00 · 2026-11-11 22:13:38 +08:00
@@ -0,0 +1,83 @@
+---
+name: eval
+description: Run the code-intelligence evals (`pnpm eval`), pick the worst failing scenario, diagnose it from the attached timeline, fix it, prove the fix with before / after numbers measured on this machine without regressing other scenarios, and open one PR. Use for "跑 eval", "优化 Java 代码智能", "eval 失败了", an issue labelled `eval`, or any change to Java navigation / usages / gutter markers that needs numbers.
+---
+
+# Eval loop: measure → fix → prove
+
+`pnpm eval` drives the production workbench in Chromium against a release `openidea-server`
+with real language servers (Java / jdtls for now) and judges every scenario on what the user
+sees: time to feedback, to the first result, to the final one, and the results themselves vs
+`evals/<language>/golden/`. One machine, before and after: numbers from different machines or
+from the nightly host don't judge a fix.
+
+## 0. Preconditions
+
+- `jdtls` on PATH and `JAVA_HOME` pointing at a JDK of the pinned `JAVA` version (projects
+  compile against it, and the JDK scenarios' results are in its sources; jdtls itself runs on
+  a newer installed JDK when it needs one). Pinned versions: `.github/tool-versions.env`. A
+  mismatch prints a warning and the report records both: say so in the PR when it applies.
+- First run builds `cargo build --release -p ide-host --bin openidea-server` and the perf
+  bundle; later iterations of a frontend-only change can use `--no-build` after a
+  `pnpm exec vite build --mode perf --outDir target/eval/dist`, a backend-only one after the
+  cargo build alone.
+- Run evals one at a time on this machine, with nothing heavy (builds, other evals) beside them.
+
+## 1. Baseline on the base commit
+
+```sh
+git switch --detach origin/main     # or the fix branch's merge base
+pnpm eval --save-baseline           # target/eval/baseline.json, this machine only
+```
+
+Read `target/eval/report.json`. Each verdict has `status`, `failures` (`kind`: `error`,
+`timeout`, `degraded`, `correctness`, `budget`, `golden`), `regressions`, `repro` (the command
+that reruns it) and, for failures, `diagnostics`: every command sent to the server in the
+slowest run's window (`cmd`, `start`, `ms`, `ok`, `bytes`, `items`), background tasks, server
+status changes and the server's log tail.
+
+## 2. Pick one problem
+
+Worst first: wrong or missing results and silent degradation, then timeouts, then the
+feedback budget (100 ms), then the slowest `completeMs`. One problem per PR. Say which
+scenario and metric, with its value vs budget.
+
+## 3. Diagnose from the timeline
+
+- Commands in the window ordered by `start`: what ran, what waited on what, what's serial
+  that could overlap, which request returned many `items`/`bytes`, which failed.
+- `feedbackMs` far above 16 ms with a fast `firstResultMs`: the UI shows nothing while it
+  waits (a hint or popup that only appears after a delay).
+- A task still running across the window: background work competing with the request.
+- For the frontend side rerun with `--trace` and open `target/eval/traces/<id>@<cond>.json`
+  in Chrome's Performance panel.
+- Name the cause at class level (as in `issue-fix`): it usually affects sibling scenarios.
+
+## 4. Fix, then rerun that scenario until it passes
+
+```sh
+pnpm eval --only <scenario> --condition <warm|cold>   # the verdict's `repro`
+```
+
+Correctness is enforced, so a faster answer that drops, truncates or serves stale results
+fails. Don't trade a budget miss for a degraded answer.
+
+## 5. Full suite, then PR
+
+```sh
+pnpm eval                           # vs the baseline from step 1: no new failures or regressions
+```
+
+PR per `issue-fix` §6, plus in **How it was verified** the before / after table from
+`target/eval/report.md` (both on this machine, commits named) and the tool versions line.
+
+## Guard rails
+
+- Goldens (`evals/*/golden/`) and budgets (`evals/budgets.json`) are human-owned. Don't edit
+  them, and don't run `--update-golden`, to make a scenario pass. When a golden looks wrong
+  (a tool bump, a deliberate behaviour change), stop and say so: a PR that touches them gets
+  `needs-decision`.
+- Don't weaken a scenario's checks or timeouts in `evals/<language>/scenarios.ts` to pass.
+  Adding a scenario (for a case the sweep or a user found) is fine; record its golden with
+  `--update-golden --only <id>` and ask a person to review the file.
+- A flaky timing (passes on rerun) is not a regression: rerun once before acting on one.
@@ -0,0 +1 @@
+../../.agents/skills/eval
@@ -0,0 +1,2 @@
+custom:
+  - https://github.com/l0ng-ai/openidea#sponsor
@@ -0,0 +1,14 @@
+# Language server and tool versions for Language smoke (.github/workflows/language-smoke.yml)
+# and the code-intelligence evals (pnpm eval). Pinned so a release can't fail an unrelated
+# PR or move a golden result; bump them deliberately (and re-record goldens: pnpm eval --update-golden).
+VTSLS=0.3.0
+TYPESCRIPT=6.0.3
+BASEDPYRIGHT=1.40.2
+RUFF=0.17.3
+GOPLS=v0.24.0
+JDTLS=1.63.0-202610191322
+KOTLIN_LSP=263.7084.0
+GRADLE=8.14.3
+INTELEPHENSE=1.18.5
+# The JDK projects compile against (JAVA_HOME): library navigation opens its sources.
+JAVA=21
@@ -24,6 +24,7 @@ on:
      - "package.json"
      - "pnpm-lock.yaml"
      - ".github/workflows/language-smoke.yml"
+      - ".github/tool-versions.env"
  pull_request:
    branches: [main]
    types: [opened, synchronize, reopened]
@@ -38,16 +39,6 @@ concurrency:
 env:
  CARGO_TERM_COLOR: always
  CARGO_INCREMENTAL: 0
-  # Pinned so a server release can't fail an unrelated PR; bump them deliberately.
-  VTSLS: 0.3.0
-  TYPESCRIPT: 6.0.3
-  BASEDPYRIGHT: 1.40.2
-  RUFF: 0.17.3
-  GOPLS: v0.24.0
-  JDTLS: 1.63.0-202610191322
-  KOTLIN_LSP: 263.7084.0
-  GRADLE: 8.14.3
-  INTELEPHENSE: 1.18.5

 jobs:
  smoke:
@@ -61,6 +52,8 @@ jobs:
    timeout-minutes: 40
    steps:
      - uses: actions/checkout@v4
+      # Tool versions, shared with the evals (pnpm eval).
+      - run: grep -E '^[A-Z0-9_]+=' .github/tool-versions.env >> "$GITHUB_ENV"
      - uses: dtolnay/rust-toolchain@stable
        with: { components: "rust-analyzer, rustfmt" }
      # One build cache for every language; main saves it.
@@ -96,7 +89,7 @@ jobs:

      - if: matrix.language == 'java' || matrix.language == 'kotlin'
        uses: actions/setup-java@v4
-        with: { distribution: temurin, java-version: "21" }
+        with: { distribution: temurin, java-version: "${{ env.JAVA }}" }
      - if: matrix.language == 'java'
        run: |
          mkdir -p "$RUNNER_TEMP/jdtls"
@@ -251,7 +251,28 @@ jobs:
              core.setFailed(`latest.json would miss ${missing.join(", ")}`);
              return;
            }
-            const manifest = { version: tag.replace(/^v/, ""), notes: "", pub_date: new Date().toISOString(), platforms };
+            // The update's notes: the release body without its Download section. That
+            // section only matters on the release page, and What's new follows it.
+            const { data: release } = await github.rest.repos.getRelease({ owner, repo, release_id });
+            const withoutDownloadSection = (body) => {
+              const kept = [];
+              let skip = 0;
+              for (const line of body.split(/\r?\n/)) {
+                const heading = /^(#{1,6})\s+(.*)$/.exec(line);
+                if (heading) {
+                  const level = heading[1].length;
+                  if (!skip && level <= 3 && /^download$/i.test(heading[2].trim())) {
+                    skip = level;
+                    continue;
+                  }
+                  if (skip && level <= skip) skip = 0;
+                }
+                if (!skip) kept.push(line);
+              }
+              return kept.join("\n").trim();
+            };
+            const notes = withoutDownloadSection(String(release.body ?? ""));
+            const manifest = { version: tag.replace(/^v/, ""), notes, pub_date: new Date().toISOString(), platforms };
            const old = assets.find((a) => a.name === "latest.json");
            if (old) await github.rest.repos.deleteReleaseAsset({ owner, repo, asset_id: old.id });
            await github.rest.repos.uploadReleaseAsset({
@@ -40,6 +40,7 @@ CLAUDE.local.md
 .claude/skills/*
 !.claude/skills/verify
 !.claude/skills/issue-fix
+!.claude/skills/eval

 # Local reference screenshots
 pycharm-screenshots/
@@ -9,6 +9,7 @@
    "dist",
    "target",
    "src/bindings.ts",
+    "evals/*/golden",
    "src/editor/core/cm/angular/parser*.ts",
    "src/editor/core/cm/jsp/parser*.ts",
    "src/editor/core/cm/astro/parser*.ts",
@@ -24,8 +24,8 @@ Visual style: dark, low-contrast greys, one cool accent
  Plugin docs: `plugins/README.md`, `plugins/themes.md`.
 - `java/` — JDT LS extension bundles (`java-usages`, `kotlin-bridge`) and the patched
  Lombok agent, embedded into the app by `ide-sdk` / `ide-tools`.
- `e2e/` — Playwright UI tests (`pnpm e2e`). `fixtures/` — sample projects per language
-  for manual and automated testing.
+- `e2e/` — Playwright UI tests (`pnpm e2e`). `evals/` — code-intelligence evals (`pnpm eval`).
+  `fixtures/` — sample projects per language for manual and automated testing.
 - `docs/` — the Mintlify documentation site (`docs.json` navigation, `en/` and `zh/` pages);
  preview with `npx mint dev` in `docs/`.

@@ -59,9 +59,12 @@ Unclear product decisions → `needs-decision` and stop.

 ## Skills

+Do not use Impeccable or its related skills in this repository, including globally installed copies.
+
 Shared agent skills live in `.agents/skills/` (`.claude/skills/` links to them):
 `verify` drives the running app through Tauri MCP and captures screenshot evidence;
-`issue-fix` takes an issue from reproduction to one PR.
+`issue-fix` takes an issue from reproduction to one PR;
+`eval` takes a failing code-intelligence eval to a fix with before / after numbers.

 ## Commands

@@ -81,6 +84,11 @@ Shared agent skills live in `.agents/skills/` (`.claude/skills/` links to them):
  `pnpm gen:bindings`.
 - `pnpm perf` (UI interaction latency / frames / leaks), `pnpm perf:rust` (criterion
  benches vs budgets). Budgets: `perf/budgets.json`.
+- `pnpm eval [--only <id>] [--runs n] [--save-baseline]` — Java code intelligence (navigation,
+  usages, gutter markers) in the production workbench against a release `openidea-server` and
+  jdtls: time to feedback / first result / complete, results vs `evals/<language>/golden/`.
+  Report: `target/eval/report.json`. Goldens and `evals/budgets.json` are changed by people only.
+  Tool versions: `.github/tool-versions.env` (shared with Language smoke).
 - `pnpm release:prepare <version>` — the release commit's changes: app version everywhere,
  `Cargo.lock`, and the plugin API snapshot when `wit/` moved to a new version.

@@ -4145,7 +4145,7 @@ checksum = "3d3067d79b975e8844ca9eb072e16b31c3c1c36928edf9c6789548c524d0d954"

 [[package]]
 name = "ide-agent"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "dunce",
 "encoding_rs",
@@ -4169,7 +4169,7 @@ dependencies = [

 [[package]]
 name = "ide-analysis"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "ast-grep-core",
 "ide-core",
@@ -4183,7 +4183,7 @@ dependencies = [

 [[package]]
 name = "ide-android"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "ide-core",
 "serde",
@@ -4194,7 +4194,7 @@ dependencies = [

 [[package]]
 name = "ide-build"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "blake3",
 "chardetng",
@@ -4212,7 +4212,7 @@ dependencies = [

 [[package]]
 name = "ide-cli"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "dirs 6.0.0",
 "serde",
@@ -4226,7 +4226,7 @@ dependencies = [

 [[package]]
 name = "ide-clipboard"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "gtk",
 "objc2 0.6.5",
@@ -4236,7 +4236,7 @@ dependencies = [

 [[package]]
 name = "ide-core"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "dunce",
 "encoding_rs",
@@ -4253,7 +4253,7 @@ dependencies = [

 [[package]]
 name = "ide-coverage"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "ide-core",
 "ide-run",
@@ -4267,15 +4267,20 @@ dependencies = [

 [[package]]
 name = "ide-dap"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "futures",
 "ide-core",
+ "ide-heap",
+ "ide-jdwp",
 "ide-jsonrpc",
+ "ide-jvm-attach",
+ "ide-jvm-inspect",
 "ide-language",
 "libc",
 "memchr",
 "parking_lot",
+ "percent-encoding",
 "portable-pty",
 "serde",
 "serde_json",
@@ -4288,7 +4293,7 @@ dependencies = [

 [[package]]
 name = "ide-db"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "bb8",
 "blake3",
@@ -4318,7 +4323,7 @@ dependencies = [

 [[package]]
 name = "ide-docker"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "ide-core",
 "serde",
@@ -4331,7 +4336,7 @@ dependencies = [

 [[package]]
 name = "ide-extension"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "base64 0.22.1",
 "blake3",
@@ -4352,7 +4357,7 @@ dependencies = [

 [[package]]
 name = "ide-fonts"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "objc2-core-foundation",
 "objc2-core-text",
@@ -4364,7 +4369,7 @@ dependencies = [

 [[package]]
 name = "ide-github"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "ide-core",
 "ide-vcs",
@@ -4375,9 +4380,27 @@ dependencies = [
 "which",
 ]

+[[package]]
+name = "ide-heap"
+version = "0.1.42"
+dependencies = [
+ "criterion",
+ "ide-core",
+ "libc",
+ "memmap2",
+ "proptest",
+ "serde",
+ "serde_json",
+ "specta",
+ "tempfile",
+ "thiserror 2.0.21",
+ "tracing",
+ "uuid",
+]
+
 [[package]]
 name = "ide-host"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "base64 0.22.1",
 "blake3",
@@ -4397,9 +4420,11 @@ dependencies = [
 "ide-db",
 "ide-docker",
 "ide-extension",
+ "ide-heap",
 "ide-http",
 "ide-index",
 "ide-jsonrpc",
+ "ide-jvm-attach",
 "ide-language",
 "ide-lsp",
 "ide-mcp",
@@ -4407,6 +4432,7 @@ dependencies = [
 "ide-plugin",
 "ide-remote",
 "ide-run",
+ "ide-scaffold",
 "ide-sdk",
 "ide-search",
 "ide-spell",
@@ -4443,7 +4469,7 @@ dependencies = [

 [[package]]
 name = "ide-http"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "ide-core",
 "ide-vfs",
@@ -4462,7 +4488,7 @@ dependencies = [

 [[package]]
 name = "ide-index"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "blake3",
 "criterion",
@@ -4482,9 +4508,21 @@ dependencies = [
 "tracing",
 ]

+[[package]]
+name = "ide-jdwp"
+version = "0.1.42"
+dependencies = [
+ "ide-jvm-inspect",
+ "proptest",
+ "tempfile",
+ "thiserror 2.0.21",
+ "tokio",
+ "tracing",
+]
+
 [[package]]
 name = "ide-jsonrpc"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "encoding_rs",
 "futures",
@@ -4499,9 +4537,29 @@ dependencies = [
 "tracing",
 ]

+[[package]]
+name = "ide-jvm-attach"
+version = "0.1.42"
+dependencies = [
+ "libc",
+ "tempfile",
+ "thiserror 2.0.21",
+]
+
+[[package]]
+name = "ide-jvm-inspect"
+version = "0.1.42"
+dependencies = [
+ "futures",
+ "serde",
+ "serde_json",
+ "specta",
+ "thiserror 2.0.21",
+]
+
 [[package]]
 name = "ide-language"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "blake3",
 "criterion",
@@ -4557,13 +4615,14 @@ dependencies = [

 [[package]]
 name = "ide-lsp"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "dunce",
 "futures",
 "globset",
 "ide-core",
 "ide-jsonrpc",
+ "ide-jvm-inspect",
 "ide-language",
 "ide-text",
 "ide-vfs",
@@ -4583,7 +4642,7 @@ dependencies = [

 [[package]]
 name = "ide-mcp"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "bytes",
 "dunce",
@@ -4602,7 +4661,7 @@ dependencies = [

 [[package]]
 name = "ide-notebook"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "futures",
 "ide-core",
@@ -4625,7 +4684,7 @@ dependencies = [

 [[package]]
 name = "ide-plugin"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "blake3",
 "dunce",
@@ -4648,9 +4707,10 @@ dependencies = [

 [[package]]
 name = "ide-profiler"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "ide-core",
+ "ide-jvm-attach",
 "ide-run",
 "serde",
 "serde_json",
@@ -4661,7 +4721,7 @@ dependencies = [

 [[package]]
 name = "ide-remote"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "dunce",
 "globset",
@@ -4680,7 +4740,7 @@ dependencies = [

 [[package]]
 name = "ide-run"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "dunce",
 "encoding_rs",
@@ -4702,9 +4762,25 @@ dependencies = [
 "zip",
 ]

+[[package]]
+name = "ide-scaffold"
+version = "0.1.42"
+dependencies = [
+ "ide-core",
+ "ide-tools",
+ "ide-vcs",
+ "insta",
+ "serde",
+ "serde_json",
+ "specta",
+ "tempfile",
+ "toml 0.9.12+spec-1.1.0",
+ "tracing",
+]
+
 [[package]]
 name = "ide-sdk"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "dunce",
 "ide-core",
@@ -4724,9 +4800,10 @@ dependencies = [

 [[package]]
 name = "ide-search"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "criterion",
+ "encoding_rs",
 "globset",
 "grep-matcher",
 "grep-regex",
@@ -4744,7 +4821,7 @@ dependencies = [

 [[package]]
 name = "ide-spell"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "ide-core",
 "parking_lot",
@@ -4757,7 +4834,7 @@ dependencies = [

 [[package]]
 name = "ide-ssh"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "dirs 6.0.0",
 "flate2",
@@ -4778,7 +4855,7 @@ dependencies = [

 [[package]]
 name = "ide-syntax"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "ide-core",
 "ide-language",
@@ -4793,7 +4870,7 @@ dependencies = [

 [[package]]
 name = "ide-terminal"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "ide-core",
 "libc",
@@ -4806,7 +4883,7 @@ dependencies = [

 [[package]]
 name = "ide-text"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "criterion",
 "ide-core",
@@ -4819,7 +4896,7 @@ dependencies = [

 [[package]]
 name = "ide-tools"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "flate2",
 "hex",
@@ -4840,7 +4917,7 @@ dependencies = [

 [[package]]
 name = "ide-vcs"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "blake3",
 "dunce",
@@ -4861,7 +4938,7 @@ dependencies = [

 [[package]]
 name = "ide-vfs"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "chardetng",
 "dunce",
@@ -4885,11 +4962,11 @@ dependencies = [

 [[package]]
 name = "ide-web"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
+ "dirs 6.0.0",
 "ide-core",
 "open",
- "parking_lot",
 "percent-encoding",
 "serde",
 "specta",
@@ -4901,7 +4978,7 @@ dependencies = [

 [[package]]
 name = "ide-workspace"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "blake3",
 "dunce",
@@ -6479,7 +6556,7 @@ dependencies = [

 [[package]]
 name = "openidea"
-version = "0.1.41"
+version = "0.1.42"
 dependencies = [
 "base64 0.22.1",
 "blake3",
@@ -6513,6 +6590,7 @@ dependencies = [
 "ide-profiler",
 "ide-remote",
 "ide-run",
+ "ide-scaffold",
 "ide-sdk",
 "ide-search",
 "ide-spell",
@@ -7719,9 +7797,9 @@ dependencies = [

 [[package]]
 name = "rustix"
-version = "1.1.5"
+version = "1.1.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "891efababe418670775f199f0d233d84843c227a0949a883ce15b37c78d6629d"
+checksum = "e5e76742e5294153d4dde1207dbec5bdd81eef6fabdc0d2118babadabfd858aa"
 dependencies = [
 "bitflags 2.13.2",
 "errno",
@@ -5,7 +5,7 @@ members = ["crates/*", "src-tauri"]
 exclude = ["sdk", "plugins"]

 [workspace.package]
-version = "0.1.41"
+version = "0.1.42"
 edition = "2024"
 rust-version = "1.96"
 license = "Apache-2.0"
@@ -43,10 +43,15 @@ ide-ssh = { path = "crates/ide-ssh" }
 ide-remote = { path = "crates/ide-remote" }
 ide-host = { path = "crates/ide-host" }
 ide-mcp = { path = "crates/ide-mcp" }
+ide-scaffold = { path = "crates/ide-scaffold" }
 ide-notebook = { path = "crates/ide-notebook" }
 ide-extension = { path = "crates/ide-extension" }
 ide-plugin = { path = "crates/ide-plugin" }
 ide-cli = { path = "crates/ide-cli" }
+ide-jvm-attach = { path = "crates/ide-jvm-attach" }
+ide-jvm-inspect = { path = "crates/ide-jvm-inspect" }
+ide-jdwp = { path = "crates/ide-jdwp" }
+ide-heap = { path = "crates/ide-heap" }

 # serialization & IPC typing
 serde = { version = "1", features = ["derive"] }
@@ -82,6 +87,7 @@ trash = "5"
 blake3 = "1"
 dunce = "1"
 dirs = "6"
+memmap2 = "0.9"
 zip = { version = "9.2.3", default-features = false, features = ["deflate"] }

 # search, index, language
@@ -19,6 +19,8 @@
  <a href="#plugins">Plugins</a>
  &nbsp;·&nbsp;
  <a href="#community">Community</a>
+  &nbsp;·&nbsp;
+  <a href="#sponsor">Sponsor</a>
 </p>

 https://github.com/user-attachments/assets/234dd54c-1255-4b3a-b35c-533d78d737c0
@@ -68,6 +70,17 @@ the [official plugins](plugins/official). The plugin API is still a preview.

 - Contributing: see [CONTRIBUTING.md](CONTRIBUTING.md) and the [contributor docs](https://openidea.sh/en/contribute/setup).

+## Sponsor
+
+If OpenIDEA saves you time, you can support its development with Alipay or WeChat Pay.
+
+<table>
+  <tr>
+    <td align="center"><img src=".github/assets/sponsor-alipay.png" alt="Alipay QR code" width="200"><br>Alipay</td>
+    <td align="center"><img src=".github/assets/sponsor-wechat.png" alt="WeChat Pay QR code" width="200"><br>WeChat Pay</td>
+  </tr>
+</table>
+
 ## License

 [Apache 2.0](LICENSE). Bundled third-party assets are listed in [`THIRD_PARTY_LICENSES.md`](THIRD_PARTY_LICENSES.md).
@@ -1,11 +1,10 @@
 # Third-party licenses

-## Flutter and Vite project icons
+## Project type icons

-`src/ui/brands/flutter.svg` and `vite.svg` use the unmodified logo paths from
+`src/ui/brands/*.svg` use the unmodified logo paths from
 [Simple Icons](https://github.com/simple-icons/simple-icons), retrieved through
-Iconify (`simple-icons:flutter`, `simple-icons:vite`) and filled with the project
-icon colors. Licensed under [CC0 1.0 Universal](https://github.com/simple-icons/simple-icons/blob/develop/LICENSE.md).
+Iconify (`simple-icons:angular`, `simple-icons:astro`, `simple-icons:cplusplus`, `simple-icons:dart`, `simple-icons:django`, `simple-icons:dotnet`, `simple-icons:elixir`, `simple-icons:express`, `simple-icons:fastapi`, `simple-icons:flask`, `simple-icons:flutter`, `simple-icons:go`, `simple-icons:haskell`, `simple-icons:html5`, `simple-icons:kotlin`, `simple-icons:ktor`, `simple-icons:laravel`, `simple-icons:nestjs`, `simple-icons:nextdotjs`, `simple-icons:nodedotjs`, `simple-icons:nuxt`, `simple-icons:openjdk`, `simple-icons:phoenixframework`, `simple-icons:php`, `simple-icons:python`, `simple-icons:quarkus`, `simple-icons:ruby`, `simple-icons:rubyonrails`, `simple-icons:rust`, `simple-icons:scala`, `simple-icons:springboot`, `simple-icons:svelte`, `simple-icons:swift`, `simple-icons:symfony`, `simple-icons:vite`, `simple-icons:zig`) and filled with the brand colors (the standard icon grey for black logos). Licensed under [CC0 1.0 Universal](https://github.com/simple-icons/simple-icons/blob/develop/LICENSE.md).

 ## Project Lombok

@@ -8,6 +8,7 @@
 use super::native::claude::history;
 use serde::Serialize;
 use specta::Type;
+use std::io::BufRead;
 use std::path::{Path, PathBuf};

 /// A session on this machine that no task holds yet.
@@ -62,11 +63,28 @@ fn message_text(e: &serde_json::Value) -> Option<String> {
    }
 }

-/// Claude Code transcripts: `<config>/projects/*/<id>.jsonl`.
-pub fn claude_sessions(config_dir: &Path) -> Vec<ExternalSession> {
+/// Whether a session that ran in `cwd` belongs to the project at `root` (in it or below it).
+fn in_project(cwd: Option<&str>, root: &Path) -> bool {
+    cwd.is_some_and(|cwd| Path::new(cwd).starts_with(root))
+}
+
+/// Claude Code transcripts of the project at `root`: `<config>/projects/<escaped cwd>/<id>.jsonl`.
+/// Only the folders named after `root` or a folder below it are read, so other projects'
+/// transcripts (often gigabytes) are never opened.
+pub fn claude_sessions(config_dir: &Path, root: &Path) -> Vec<ExternalSession> {
    let mut out = Vec::new();
    let projects = config_dir.join("projects");
-    let dirs: Vec<PathBuf> = std::fs::read_dir(&projects).into_iter().flatten().flatten().map(|d| d.path()).collect();
+    let escaped = history::escape(root);
+    let dirs: Vec<PathBuf> = std::fs::read_dir(&projects)
+        .into_iter()
+        .flatten()
+        .flatten()
+        .filter(|d| {
+            let name = d.file_name().to_string_lossy().into_owned();
+            name == escaped || name.starts_with(&format!("{escaped}-"))
+        })
+        .map(|d| d.path())
+        .collect();
    for dir in dirs {
        let files: Vec<PathBuf> = std::fs::read_dir(&dir)
            .into_iter()
@@ -93,6 +111,10 @@ pub fn claude_sessions(config_dir: &Path) -> Vec<ExternalSession> {
                .filter_map(|e| e.get("cwd").and_then(serde_json::Value::as_str))
                .next()
                .map(str::to_owned);
+            // Escaping is lossy (`a/b-c` and `a/b/c` share a folder): the transcript decides.
+            if !in_project(cwd.as_deref(), root) {
+                continue;
+            }
            out.push(ExternalSession {
                provider: "claude".to_owned(),
                session_id: id,
@@ -138,12 +160,13 @@ fn codex_user_text(line: &serde_json::Value) -> Option<String> {

 /// One rollout file's session, or `None` when it names no `session_meta`.
 fn codex_file(path: &Path) -> Option<ExternalSession> {
-    let text = std::fs::read_to_string(path).ok()?;
+    let file = std::fs::File::open(path).ok()?;
    let mut id: Option<(String, Option<String>)> = None;
    let mut title: Option<String> = None;
    // `session_meta` is first, but cap the scan: giant rollouts must not stall listing.
-    for line in text.lines().take(400) {
-        let Ok(v) = serde_json::from_str::<serde_json::Value>(line) else { continue };
+    for line in std::io::BufReader::new(file).lines().take(400) {
+        let Ok(line) = line else { break };
+        let Ok(v) = serde_json::from_str::<serde_json::Value>(&line) else { continue };
        if id.is_none()
            && let Some(meta) = codex_meta(&v)
        {
@@ -183,11 +206,11 @@ fn rollout_files(dir: &Path, depth: usize, out: &mut Vec<PathBuf>) {
    }
 }

-/// Codex sessions: `<home>/.codex/sessions/**`.
-pub fn codex_sessions(codex_home: &Path) -> Vec<ExternalSession> {
+/// Codex sessions of the project at `root`: `<home>/.codex/sessions/**`.
+pub fn codex_sessions(codex_home: &Path, root: &Path) -> Vec<ExternalSession> {
    let mut files = Vec::new();
    rollout_files(&codex_home.join("sessions"), 4, &mut files);
-    files.iter().filter_map(|p| codex_file(p)).collect()
+    files.iter().filter_map(|p| codex_file(p)).filter(|s| in_project(s.cwd.as_deref(), root)).collect()
 }

 #[cfg(test)]
@@ -222,7 +245,7 @@ mod tests {
            ],
        );
        write(&tmp.path().join("projects/-work-app/empty.jsonl"), &[]);
-        let sessions = claude_sessions(tmp.path());
+        let sessions = claude_sessions(tmp.path(), Path::new("/work/app"));
        assert_eq!(sessions.len(), 1);
        let s = &sessions[0];
        assert_eq!((s.provider.as_str(), s.session_id.as_str()), ("claude", "abc123"));
@@ -234,8 +257,8 @@ mod tests {
    #[test]
    fn claude_title_falls_back_to_the_first_prompt() {
        let tmp = tempfile::tempdir().unwrap();
-        write(&tmp.path().join("projects/-x/def456.jsonl"), &[claude_user("u1", "hello there")]);
-        let sessions = claude_sessions(tmp.path());
+        write(&tmp.path().join("projects/-work-app/def456.jsonl"), &[claude_user("u1", "hello there")]);
+        let sessions = claude_sessions(tmp.path(), Path::new("/work/app"));
        assert_eq!(sessions.len(), 1);
        assert_eq!(sessions[0].title, "hello there");
    }
@@ -253,7 +276,7 @@ mod tests {
            ],
        );
        write(&tmp.path().join("sessions/notes.jsonl"), &[json!({"nope": true})]);
-        let sessions = codex_sessions(tmp.path());
+        let sessions = codex_sessions(tmp.path(), Path::new("/work"));
        assert_eq!(sessions.len(), 1);
        let s = &sessions[0];
        assert_eq!((s.provider.as_str(), s.session_id.as_str()), ("codex", "thread-1"));
@@ -264,7 +287,39 @@ mod tests {
    #[test]
    fn missing_directories_list_nothing() {
        let tmp = tempfile::tempdir().unwrap();
-        assert!(claude_sessions(&tmp.path().join("nope")).is_empty());
-        assert!(codex_sessions(&tmp.path().join("nope")).is_empty());
+        assert!(claude_sessions(&tmp.path().join("nope"), Path::new("/work")).is_empty());
+        assert!(codex_sessions(&tmp.path().join("nope"), Path::new("/work")).is_empty());
+    }
+
+    #[test]
+    fn only_sessions_of_the_project_are_listed() {
+        let tmp = tempfile::tempdir().unwrap();
+        let user = |cwd: &str| {
+            json!({"type": "user", "uuid": "u1", "parentUuid": null,
+                "message": {"role": "user", "content": "hi"}, "cwd": cwd})
+        };
+        write(&tmp.path().join("projects/-work-app/in-root.jsonl"), &[user("/work/app")]);
+        write(&tmp.path().join("projects/-work-app-web/below.jsonl"), &[user("/work/app/web")]);
+        // `/work/app-old` escapes like a folder below `/work/app`; its transcript says otherwise.
+        write(&tmp.path().join("projects/-work-app-old/sibling.jsonl"), &[user("/work/app-old")]);
+        write(&tmp.path().join("projects/-other/elsewhere.jsonl"), &[user("/other")]);
+        let mut ids: Vec<String> =
+            claude_sessions(tmp.path(), Path::new("/work/app")).into_iter().map(|s| s.session_id).collect();
+        ids.sort();
+        assert_eq!(ids, ["below", "in-root"]);
+
+        let rollout = |name: &str, cwd: &str| {
+            write(
+                &tmp.path().join(format!("codex/sessions/2026/01/02/{name}.jsonl")),
+                &[json!({"payload": {"type": "session_meta", "payload": {"id": name, "cwd": cwd}}})],
+            )
+        };
+        rollout("mine", "/work/app/api");
+        rollout("theirs", "/other");
+        let codex: Vec<String> = codex_sessions(&tmp.path().join("codex"), Path::new("/work/app"))
+            .into_iter()
+            .map(|s| s.session_id)
+            .collect();
+        assert_eq!(codex, ["mine"]);
    }
 }
@@ -592,17 +592,19 @@ impl AgentManager {
        self.tasks.read().values().filter_map(|t| t.state.lock().session_id.clone()).collect()
    }

-    /// Sessions on this machine that no task holds yet, newest first (#312).
+    /// Sessions of this project on this machine that no task holds yet, newest first. Reads
+    /// transcripts from disk: call it off the UI and command threads.
    pub fn external_sessions(&self) -> Vec<ExternalSession> {
        let mut out = Vec::new();
+        let root = self.shared.paths.project_root.clone();
        if let Some(spec) = self.shared.spec("claude")
            && let Some(dir) = native::claude::cli::config_dir(&spec.env)
        {
-            out.extend(external::claude_sessions(&dir));
+            out.extend(external::claude_sessions(&dir, &root));
        }
        let home = self.shared.paths.home_dir.clone().or_else(home_dir);
        if let Some(home) = home {
-            out.extend(external::codex_sessions(&home.join(".codex")));
+            out.extend(external::codex_sessions(&home.join(".codex"), &root));
        }
        let known = self.imported_session_ids();
        out.retain(|s| !known.contains(&s.session_id));
@@ -1232,15 +1234,6 @@ impl AgentManager {
        };
        // A copy of the task (a fork) may work in the same worktree.
        let shared = self.tasks.read().values().any(|t| t.id != id && t.state.lock().task.cwd.0 == cwd);
-        if remove_worktree && !shared && isolation == Isolation::Worktree && cwd.exists() {
-            // On Windows the agent's processes lock the folder (their working directory,
-            // open files) until they have exited.
-            if let Some(peer) = live {
-                let _ = tokio::time::timeout(EXIT_WAIT, peer.exited()).await;
-            }
-            let worktrees = self.shared.paths.worktrees_dir.clone();
-            blocking(move || worktree::remove(&root, &worktrees, &cwd)).await?;
-        }
        self.tasks.write().remove(id);
        *task.deleted.lock() = true;
        let _ = std::fs::remove_file(self.shared.paths.store_dir.join(format!("{id}.json")));
@@ -1250,7 +1243,17 @@ impl AgentManager {
        {
            let _ = std::fs::remove_dir_all(self.shared.paths.store_dir.join(saved));
        }
+        // The list drops the task now; removing a large worktree's folder can take a while.
        self.shared.emit(AgentEvent::Removed { id: id.to_owned() });
+        if remove_worktree && !shared && isolation == Isolation::Worktree && cwd.exists() {
+            // On Windows the agent's processes lock the folder (their working directory,
+            // open files) until they have exited.
+            if let Some(peer) = live {
+                let _ = tokio::time::timeout(EXIT_WAIT, peer.exited()).await;
+            }
+            let worktrees = self.shared.paths.worktrees_dir.clone();
+            blocking(move || worktree::remove(&root, &worktrees, &cwd)).await?;
+        }
        Ok(())
    }

@@ -31,7 +31,7 @@ pub(super) fn transcript(config_dir: &Path, cwd: &Path, id: &str) -> PathBuf {
 }

 /// The project directory name Claude Code uses for `cwd`.
-pub(super) fn escape(cwd: &Path) -> String {
+pub(crate) fn escape(cwd: &Path) -> String {
    cwd.to_string_lossy().chars().map(|c| if c.is_ascii_alphanumeric() { c } else { '-' }).collect()
 }

@@ -59,6 +59,8 @@ const BUILTIN: &[(&str, &str, Driver, &[&str], &str)] = &[
    ("cursor", "Cursor", Driver::Acp, &["cursor-agent", "acp"], "Run `cursor-agent login` in a terminal."),
    ("opencode", "OpenCode", Driver::Acp, &["opencode", "acp"], "Run `opencode auth login` in a terminal."),
    ("qwen", "Qwen Code", Driver::Acp, &["qwen", "--experimental-acp"], "Run `qwen` in a terminal and sign in."),
+    ("qoder", "Qoder", Driver::Acp, &["qodercli", "--acp"], "Run `qodercli login` in a terminal."),
+    ("qodercn", "Qoder CN", Driver::Acp, &["qoderclicn", "--acp"], "Run `qoderclicn login` in a terminal."),
    ("kimi", "Kimi CLI", Driver::Acp, &["kimi", "acp"], "Run `kimi` in a terminal and sign in."),
    ("goose", "Goose", Driver::Acp, &["goose", "acp"], "Run `goose configure` in a terminal."),
    ("auggie", "Auggie", Driver::Acp, &["auggie", "--acp"], "Run `auggie login` in a terminal."),
@@ -409,6 +411,45 @@ mod tests {
    use super::*;
    use serde_json::json;

+    #[test]
+    fn qoder_global_and_cn_editions_are_both_built_in() {
+        let info = |id: &str| describe(&load(None).into_iter().find(|s| s.id == id).unwrap());
+        let (global, cn) = (info("qoder"), info("qodercn"));
+        assert_eq!((global.name.as_str(), global.command.as_str()), ("Qoder", "qodercli --acp"));
+        assert_eq!((cn.name.as_str(), cn.command.as_str()), ("Qoder CN", "qoderclicn --acp"));
+        assert_eq!(global.login_hint.as_deref(), Some("Run `qodercli login` in a terminal."));
+    }
+
+    #[test]
+    fn qodercn_is_available_without_a_file_and_preserves_existing_customizations() {
+        let dir = tempfile::tempdir().unwrap();
+        let qodercn = load(None).into_iter().find(|s| s.id == "qodercn").unwrap();
+        assert_eq!(qodercn.driver, Driver::Acp);
+        assert!(!is_unlisted(&qodercn));
+        let info = describe(&qodercn);
+        assert_eq!(info.name, "Qoder CN");
+        assert_eq!(info.command, "qoderclicn --acp");
+        assert_eq!(info.source, AgentSource::Builtin);
+        assert!(info.command_editable);
+        assert_eq!(info.login_hint.as_deref(), Some("Run `qoderclicn login` in a terminal."));
+
+        std::fs::write(
+            dir.path().join("qodercn.toml"),
+            "command = \"/custom/qoderclicn\"\nargs = [\"--acp\"]\n[env]\nK = \"v\"\n",
+        )
+        .unwrap();
+        let specs = load(Some(dir.path()));
+        assert_eq!(specs.iter().filter(|s| s.id == "qodercn").count(), 1);
+        let custom = specs.iter().find(|s| s.id == "qodercn").unwrap();
+        assert_eq!(custom.command, "/custom/qoderclicn");
+        assert_eq!(custom.env, [("K".into(), "v".into())]);
+        set_builtin_command(dir.path(), "qodercn", None).unwrap();
+        let reset = load(Some(dir.path())).into_iter().find(|s| s.id == "qodercn").unwrap();
+        assert_eq!(reset.command, "qoderclicn");
+        assert_eq!(reset.args, ["--acp"]);
+        assert_eq!(reset.env, custom.env);
+    }
+
    #[test]
    fn login_hints_cover_all_builtin_commands_and_preserve_separate_login_tools() {
        for spec in builtins() {
@@ -14,8 +14,18 @@ pub(crate) struct Created {
    pub base_commit: String,
 }

+/// The repository of `path`, free to execute: the host runs every agent command that
+/// reaches a repository only in a trusted project.
+fn discover(path: &Path) -> IdeResult<Option<Repo>> {
+    let repo = Repo::discover(path)?;
+    if let Some(repo) = &repo {
+        repo.set_execution_allowed(true);
+    }
+    Ok(repo)
+}
+
 fn repo_at(path: &Path) -> IdeResult<Repo> {
-    Repo::discover(path)?.ok_or_else(|| IdeError::unavailable(format!("{} is not in a git repository", path.display())))
+    discover(path)?.ok_or_else(|| IdeError::unavailable(format!("{} is not in a git repository", path.display())))
 }

 /// The current branch (or commit when detached) and its commit.
@@ -91,10 +101,11 @@ fn retrying(attempts: u32, pause: Duration, mut op: impl FnMut() -> std::io::Res
 /// Files under `cwd` that differ from `base` (untracked files included), with line counts.
 /// Outside a repository there is nothing to compare against: no changes.
 pub(crate) fn changes(cwd: &Path, base: &str) -> IdeResult<Vec<AgentChange>> {
-    let Some(repo) = Repo::discover(cwd)? else { return Ok(Vec::new()) };
+    let Some(repo) = discover(cwd)? else { return Ok(Vec::new()) };
    let root = repo.root().to_path_buf();
-    let stats = repo.numstat_with_worktree(base)?;
-    let stat = |rel: &str| stats.iter().find(|(p, ..)| p == rel).map_or((0, 0), |&(_, a, r)| (a, r));
+    let stats: std::collections::HashMap<String, (u32, u32)> =
+        repo.numstat_with_worktree(base)?.into_iter().map(|(p, a, r)| (p, (a, r))).collect();
+    let stat = |rel: &str| stats.get(rel).copied().unwrap_or((0, 0));
    let mut out: Vec<AgentChange> = repo
        .diff_with_worktree(base)?
        .into_iter()
@@ -103,8 +114,9 @@ pub(crate) fn changes(cwd: &Path, base: &str) -> IdeResult<Vec<AgentChange>> {
            AgentChange { path: IpcPath(root.join(&f.path)), rel_path: f.path, kind: f.kind, added, removed }
        })
        .collect();
+    let tracked: std::collections::HashSet<String> = out.iter().map(|c| c.rel_path.clone()).collect();
    for s in repo.status()? {
-        if s.unstaged == Some(ChangeKind::Untracked) && !out.iter().any(|c| c.rel_path == s.rel_path) {
+        if s.unstaged == Some(ChangeKind::Untracked) && !tracked.contains(&s.rel_path) {
            let added = std::fs::read_to_string(&*s.path).map_or(0, |t| t.lines().count() as u32);
            out.push(AgentChange { path: s.path, rel_path: s.rel_path, kind: ChangeKind::Added, added, removed: 0 });
        }
@@ -124,7 +136,10 @@ pub(crate) fn stamp(path: &Path) -> Stamp {

 /// The files already changed in `cwd` against `HEAD`, stamped so later edits to them show.
 pub(crate) fn snapshot(cwd: &Path) -> IdeResult<BTreeMap<String, Stamp>> {
-    Ok(changes(cwd, "HEAD")?.into_iter().map(|c| (c.rel_path, stamp(&c.path))).collect())
+    // Paths and stamps only: no diff and no line counts, which read every untracked file
+    // and held up starting a task in a checkout with many of them.
+    let Some(repo) = Repo::discover(cwd)? else { return Ok(BTreeMap::new()) };
+    Ok(repo.status()?.into_iter().map(|s| (s.rel_path, stamp(&s.path))).collect())
 }

 /// Drops the changes that were there before the task and haven't been touched since.
@@ -204,7 +219,7 @@ pub(crate) fn restore(
 }

 pub(crate) fn file_at(cwd: &Path, path: &Path, rev: &str) -> IdeResult<Option<String>> {
-    let Some(repo) = Repo::discover(cwd)? else { return Ok(None) };
+    let Some(repo) = discover(cwd)? else { return Ok(None) };
    let encoding = match ide_vfs::load_text(path) {
        Ok(loaded) => encoding_rs::Encoding::for_label(loaded.encoding.as_bytes()).unwrap_or(encoding_rs::UTF_8),
        Err(_) => ide_vfs::detect_encoding(&repo.file_bytes_at(path, Some(rev))?.unwrap_or_default(), None),
@@ -79,23 +79,26 @@ fn scan(
    Ok(())
 }

-/// `None` means an input cannot be resolved safely: build normally without reuse.
-/// All filesystem work belongs on the caller's blocking thread.
-pub fn launch_fingerprint(
+/// One planned step as fingerprinted: resolved program, args, cwd and env.
+type StepFingerprint = (PathBuf, Vec<String>, PathBuf, BTreeMap<String, String>);
+
+/// Environment, planned steps and input paths shared by every Maven build
+/// fingerprint: the filtered process environment plus the wrapper, settings
+/// and JDK files the planned invocation reads.
+struct Collected {
+    inputs: Vec<PathBuf>,
+    environment: BTreeMap<String, String>,
+    steps: Vec<StepFingerprint>,
+    inherited_hash: u64,
+}
+
+fn collect(
    root: &Path,
-    model: &BuildModel,
-    module: &Path,
-    test: bool,
+    projects: &[&maven::MavenProject],
+    plan: &BuildPlan,
    local_repo: Option<PathBuf>,
-    cancel: &CancellationToken,
-) -> IdeResult<Option<LaunchFingerprint>> {
-    let Some(m) = &model.maven else { return Ok(None) };
-    let projects = maven_launch_projects(root, m, module, test);
-    if projects.is_empty() {
-        return Ok(None);
-    }
-    let Some(mut inputs) = maven::launch_inputs(root, &projects, local_repo) else { return Ok(None) };
-    let plan = launch_plan(root, model, module, test, BuildKind::Build)?;
+) -> IdeResult<Option<Collected>> {
+    let Some(mut inputs) = maven::launch_inputs(root, projects, local_repo) else { return Ok(None) };
    let mut environment: BTreeMap<String, String> = std::env::vars()
        .map(|(key, value)| (key.to_ascii_uppercase(), value))
        .filter(|(key, _)| {
@@ -165,26 +168,111 @@ pub fn launch_fingerprint(
            Path::new(home).join(if cfg!(windows) { "bin/java.exe" } else { "bin/java" }),
        ]);
    }
-    let steps: Vec<_> =
-        plan.steps.iter().map(|s| (process::program_path(&s.program, &s.env), &s.args, &s.cwd, &s.env)).collect();
+    let steps = plan
+        .steps
+        .iter()
+        .map(|s| (process::program_path(&s.program, &s.env), s.args.clone(), s.cwd.clone(), s.env.clone()))
+        .collect();
    // Profiles and plugins can read arbitrary environment variables. Keep a signature
    // of the inherited environment without persisting unrelated credentials.
    let inherited: BTreeMap<_, _> = std::env::vars_os().collect();
    let mut inherited_hash = std::hash::DefaultHasher::new();
    inherited.hash(&mut inherited_hash);
-    let parameters =
-        serde_json::to_string(&(root, module, test, steps, environment, inherited_hash.finish(), projects))
-            .map_err(|e| IdeError::internal(e.to_string()))?;
+    Ok(Some(Collected { inputs, environment, steps, inherited_hash: inherited_hash.finish() }))
+}
+
+/// Scans every input, storing project inputs relative to the project.
+fn scan_all(
+    root: &Path,
+    inputs: Vec<PathBuf>,
+    cancel: &CancellationToken,
+) -> IdeResult<BTreeMap<PathBuf, Option<Stamp>>> {
    let mut files = BTreeMap::new();
    for path in inputs {
        scan(root, &path, &mut files, &mut HashSet::new(), cancel)?;
    }
    // Store project inputs relative to the project, including each file's relative path.
-    let files = files.into_iter().map(|(p, stamp)| (p.strip_prefix(root).unwrap_or(&p).to_path_buf(), stamp)).collect();
-    Ok(Some(LaunchFingerprint { version: 2, parameters, files }))
+    Ok(files.into_iter().map(|(p, stamp)| (p.strip_prefix(root).unwrap_or(&p).to_path_buf(), stamp)).collect())
+}
+
+/// `None` means an input cannot be resolved safely: build normally without reuse.
+/// All filesystem work belongs on the caller's blocking thread.
+pub fn launch_fingerprint(
+    root: &Path,
+    model: &BuildModel,
+    module: &Path,
+    test: bool,
+    local_repo: Option<PathBuf>,
+    cancel: &CancellationToken,
+) -> IdeResult<Option<LaunchFingerprint>> {
+    let Some(m) = &model.maven else { return Ok(None) };
+    let projects = maven_launch_projects(root, m, module, test);
+    if projects.is_empty() {
+        return Ok(None);
+    }
+    let plan = launch_plan(root, model, module, test, BuildKind::Build)?;
+    let Some(collected) = collect(root, &projects, &plan, local_repo)? else { return Ok(None) };
+    let Collected { inputs, environment, steps, inherited_hash } = collected;
+    let parameters = serde_json::to_string(&(root, module, test, steps, environment, inherited_hash, projects))
+        .map_err(|e| IdeError::internal(e.to_string()))?;
+    Ok(Some(LaunchFingerprint { version: 2, parameters, files: scan_all(root, inputs, cancel)? }))
+}
+
+/// Inputs fingerprint of a menu build (`Build Project` / `Build Module`): the
+/// built modules' sources, resources and POM chain plus the exact planned
+/// steps, so an unchanged project skips Maven entirely (#2476). Only `Build`
+/// callers reuse it, never `Rebuild` or `Artifacts`; `None` disables reuse.
+/// All filesystem work belongs on the caller's blocking thread.
+pub fn menu_fingerprint(
+    root: &Path,
+    model: &BuildModel,
+    file: Option<&Path>,
+    cancel: &CancellationToken,
+) -> IdeResult<Option<LaunchFingerprint>> {
+    let plan = plan(root, model, BuildKind::Build, file, None)?;
+    if plan.system != BuildSystem::Maven {
+        return Ok(None);
+    }
+    let Some(m) = &model.maven else { return Ok(None) };
+    let projects: Vec<&maven::MavenProject> = match file {
+        Some(file) => {
+            // The innermost linked project owning the file, as in single-file compilation.
+            let rel = rel_str(root, file);
+            let module = m
+                .projects
+                .iter()
+                .filter(|p| rel == p.dir || rel.starts_with(&format!("{}/", p.dir)) || p.dir.is_empty())
+                .max_by_key(|p| p.dir.len());
+            match module {
+                Some(project) if project.packaging != "pom" => {
+                    maven_launch_projects(root, m, &subdir(root, &project.dir), true)
+                }
+                _ => return Ok(None),
+            }
+        }
+        // A project build walks every linked project.
+        None => m.projects.iter().collect(),
+    };
+    if projects.is_empty() {
+        return Ok(None);
+    }
+    let Some(collected) = collect(root, &projects, &plan, None)? else { return Ok(None) };
+    let Collected { inputs, environment, steps, inherited_hash } = collected;
+    let scope = file.map(|f| rel_str(root, f));
+    let parameters = serde_json::to_string(&("menu-build", scope, steps, environment, inherited_hash))
+        .map_err(|e| IdeError::internal(e.to_string()))?;
+    Ok(Some(LaunchFingerprint { version: 2, parameters, files: scan_all(root, inputs, cancel)? }))
 }

 impl LaunchFingerprint {
+    /// Folders (relative to the project) of the Maven projects whose POM is an input.
+    pub(super) fn project_dirs(&self) -> impl Iterator<Item = &Path> {
+        self.files
+            .keys()
+            .filter(|p| p.is_relative() && p.file_name().is_some_and(|n| n == "pom.xml"))
+            .map(|pom| pom.parent().unwrap_or(Path::new("")))
+    }
+
    /// Relative paths whose content changed while `(length, mtime)` stayed the same
    /// compared to `previous` (timestamp-preserving external edits, coarse-granularity
    /// filesystems). Timestamp-incremental builders (Maven, the Kotlin plugin) would
@@ -0,0 +1,253 @@
+//! Resources a Gradle build changes while copying them (filters, `expand`, renames, extra
+//! sources). The Java language server copies them as they are into Buildship's
+//! `bin/<source set>` folders, which launches run from; IntelliJ IDEA runs Gradle's copy.
+//! Before a launch, Gradle processes them and its output replaces those plain copies, so
+//! each resource is on the classpath once, as the build produces it.
+use super::{BuildPlan, BuildStep, BuildSystem, Format, run, subdir};
+use crate::BuildModel;
+use crate::gradle::{self, GradleModel, ReportedProject};
+use ide_core::{CancellationToken, IdeError, IdeResult, Severity};
+use std::fs;
+use std::path::{Path, PathBuf};
+
+/// A source set whose resources Gradle processes, with the folder the launch reads them from.
+#[derive(Debug, PartialEq)]
+struct Target {
+    /// Task path from the root build (`:app:processResources`).
+    task: String,
+    /// Where the task writes the resources.
+    output: PathBuf,
+    /// The Java language server's output folder of the source set, on the launch classpath.
+    jdtls: PathBuf,
+}
+
+fn targets(reported: &[ReportedProject], class_paths: &[String]) -> Vec<Target> {
+    // buildSrc is built before the build, never addressable as `:buildSrc:<task>`.
+    let addressable = |p: &&ReportedProject| !p.build.as_deref().is_some_and(|b| b.ends_with(":buildSrc"));
+    let sets: Vec<_> = reported
+        .iter()
+        .filter(addressable)
+        .flat_map(|p| p.source_sets.iter().filter_map(move |s| Some((p, s, s.processed.as_ref()?))))
+        .collect();
+    if sets.is_empty() {
+        return Vec::new();
+    }
+    // Canonical, as Gradle reports its folders: jdtls may spell them differently.
+    let entries: Vec<PathBuf> = class_paths.iter().filter_map(|p| dunce::canonicalize(p).ok()).collect();
+    sets.into_iter()
+        .filter_map(|(p, set, resources)| {
+            // Buildship's output folder of the source set.
+            let jdtls = dunce::canonicalize(Path::new(&p.dir).join("bin").join(&set.name)).ok()?;
+            let project = if p.path == ":" { "" } else { p.path.as_str() };
+            entries.contains(&jdtls).then(|| Target {
+                task: format!("{project}:{}", resources.task),
+                output: PathBuf::from(&resources.dir),
+                jdtls,
+            })
+        })
+        .collect()
+}
+
+/// Before a JVM launch from the Java language server's output folders on `class_paths`:
+/// runs the Gradle tasks that process resources of those folders' source sets (see the
+/// module docs) and copies their output over the plain copies. `cache_dir` holds the
+/// linked builds' last syncs. Returns the tasks run, none for builds that copy their
+/// resources as they are.
+pub fn process_launch_resources(
+    root: &Path,
+    model: &BuildModel,
+    cache_dir: &Path,
+    class_paths: &[String],
+    cancel: &CancellationToken,
+) -> IdeResult<Vec<String>> {
+    // A composite also reports the builds it includes: the deepest build owning a folder runs its task.
+    let mut builds: Vec<&GradleModel> = model.gradle.iter().collect();
+    builds.sort_by_key(|g| std::cmp::Reverse(subdir(root, &g.dir).components().count()));
+    let mut done: Vec<PathBuf> = Vec::new();
+    let mut tasks = Vec::new();
+    for g in builds {
+        let reported = crate::sync::reported_gradle_projects(&gradle::cache_dir(cache_dir, &g.dir));
+        let targets: Vec<Target> =
+            targets(&reported, class_paths).into_iter().filter(|t| !done.contains(&t.jdtls)).collect();
+        if targets.is_empty() {
+            continue;
+        }
+        let mut step = BuildStep::new(&g.command, &["--console=plain"], &subdir(root, &g.dir), Format::Gradle);
+        step.args.extend(targets.iter().map(|t| t.task.clone()));
+        step.env = g.env.clone();
+        let plan = BuildPlan { system: BuildSystem::Gradle, title: "Process resources".into(), steps: vec![step] };
+        let result = run(root, &plan, cancel, &|_| {})?;
+        if !result.success {
+            let errors: Vec<&str> = result
+                .messages
+                .iter()
+                .filter(|m| m.severity == Severity::Error)
+                .take(5)
+                .map(|m| m.text.as_str())
+                .collect();
+            return Err(IdeError::internal(errors.join("\n")));
+        }
+        for t in targets {
+            if cancel.is_cancelled() {
+                return Err(IdeError::cancelled());
+            }
+            overlay(&t.output, &t.output, &t.jdtls)?;
+            done.push(t.jdtls);
+            tasks.push(t.task);
+        }
+    }
+    Ok(tasks)
+}
+
+/// Copies the files under `dir` to the same place under `to` (relative to `base`) where
+/// they differ, so unchanged resources are not rewritten on every launch.
+fn overlay(base: &Path, dir: &Path, to: &Path) -> IdeResult<()> {
+    let entries = match fs::read_dir(dir) {
+        // No resources: the task had nothing to copy.
+        Err(e) if e.kind() == std::io::ErrorKind::NotFound => return Ok(()),
+        entries => entries?,
+    };
+    for entry in entries {
+        let entry = entry?;
+        let path = entry.path();
+        let kind = entry.file_type()?;
+        if kind.is_dir() {
+            overlay(base, &path, to)?;
+        } else if kind.is_file() {
+            let dest = to.join(path.strip_prefix(base).map_err(|e| IdeError::internal(e.to_string()))?);
+            if !same_content(&path, &dest) {
+                if let Some(parent) = dest.parent() {
+                    fs::create_dir_all(parent)?;
+                }
+                fs::copy(&path, &dest)?;
+            }
+        }
+    }
+    Ok(())
+}
+
+fn same_content(a: &Path, b: &Path) -> bool {
+    match (fs::metadata(a), fs::metadata(b)) {
+        (Ok(m), Ok(n)) if m.len() == n.len() => matches!((fs::read(a), fs::read(b)), (Ok(x), Ok(y)) if x == y),
+        _ => false,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::gradle::{ProcessedResources, ReportedSourceSet};
+
+    fn write(path: &Path, text: &str) {
+        fs::create_dir_all(path.parent().unwrap()).unwrap();
+        fs::write(path, text).unwrap();
+    }
+
+    fn set(name: &str, processed: Option<&Path>) -> ReportedSourceSet {
+        ReportedSourceSet {
+            name: name.into(),
+            test: name == "test",
+            processed: processed.map(|dir| ProcessedResources {
+                task: if name == "main" { "processResources".into() } else { "processTestResources".into() },
+                dir: dir.to_string_lossy().into_owned(),
+            }),
+            ..Default::default()
+        }
+    }
+
+    fn project(path: &str, dir: &Path, sets: Vec<ReportedSourceSet>) -> ReportedProject {
+        ReportedProject {
+            path: path.into(),
+            dir: dir.to_string_lossy().into_owned(),
+            source_sets: sets,
+            ..Default::default()
+        }
+    }
+
+    /// Only processed source sets whose jdtls output is on the classpath: the launched
+    /// module, upstream modules of the build, tests only for a test classpath; never buildSrc.
+    #[test]
+    fn targets_are_processed_sets_on_the_classpath() {
+        let tmp = tempfile::tempdir().unwrap();
+        let root = dunce::canonicalize(tmp.path()).unwrap();
+        let (app, lib, plain, src) = (root.join("app"), root.join("lib"), root.join("plain"), root.join("buildSrc"));
+        for dir in [&root, &app, &lib, &plain, &src] {
+            for set in ["main", "test"] {
+                fs::create_dir_all(dir.join("bin").join(set)).unwrap();
+            }
+        }
+        let out = |dir: &Path, set: &str| dir.join("build/resources").join(set);
+        let mut build_src = project(":buildSrc", &src, vec![set("main", Some(&out(&src, "main")))]);
+        build_src.build = Some(":buildSrc".into());
+        let reported = [
+            project(":", &root, vec![set("main", Some(&out(&root, "main")))]),
+            project(":app", &app, vec![set("main", Some(&out(&app, "main"))), set("test", Some(&out(&app, "test")))]),
+            project(":lib", &lib, vec![set("main", Some(&out(&lib, "main")))]),
+            project(":plain", &plain, vec![set("main", None)]),
+            build_src,
+        ];
+        let entry = |dir: &Path, set: &str| dir.join("bin").join(set).to_string_lossy().into_owned();
+        let runtime = [
+            entry(&app, "main"),
+            // Spelled differently from Gradle's canonical folders.
+            format!("{}/", entry(&lib, "main")),
+            entry(&plain, "main"),
+            entry(&src, "main"),
+            "/m2/spring-core.jar".into(),
+        ];
+        let tasks = |cp: &[String]| targets(&reported, cp).into_iter().map(|t| t.task).collect::<Vec<_>>();
+        assert_eq!(tasks(&runtime), [":app:processResources", ":lib:processResources"]);
+        let mut test = runtime.to_vec();
+        test.insert(0, entry(&app, "test"));
+        assert_eq!(tasks(&test), [":app:processResources", ":app:processTestResources", ":lib:processResources"]);
+        assert_eq!(tasks(&[entry(&root, "main")]), [":processResources"], "the root project");
+        let app_targets = targets(&reported, &runtime);
+        let t = &app_targets[0];
+        assert_eq!(
+            (t.output.as_path(), t.jdtls.as_path()),
+            (out(&app, "main").as_path(), app.join("bin/main").as_path())
+        );
+
+        let unprocessed = [project(":app", &app, vec![set("main", None)])];
+        assert!(targets(&unprocessed, &runtime).is_empty());
+    }
+
+    /// The processed copy replaces jdtls's plain one (the reported `@version@` case), new
+    /// files are added, identical ones are left alone and files Gradle doesn't produce stay.
+    #[test]
+    fn overlay_replaces_plain_copies_with_gradle_output() {
+        let tmp = tempfile::tempdir().unwrap();
+        let (output, jdtls) = (tmp.path().join("build/resources/main"), tmp.path().join("bin/main"));
+        write(&output.join("application.yaml"), "spring:\n  application:\n    version: 0.0.1\n");
+        write(&output.join("static/extra.txt"), "from an extra source");
+        write(&output.join("migrations/V1__init.sql"), "create table t (id int);");
+        write(&jdtls.join("application.yaml"), "spring:\n  application:\n    version: @version@\n");
+        write(&jdtls.join("migrations/V1__init.sql"), "create table t (id int);");
+        write(&jdtls.join("com/example/App.class"), "bytecode");
+        let unchanged = fs::metadata(jdtls.join("migrations/V1__init.sql")).unwrap().modified().unwrap();
+        std::thread::sleep(std::time::Duration::from_millis(20));
+
+        overlay(&output, &output, &jdtls).unwrap();
+        let read = |rel: &str| fs::read_to_string(jdtls.join(rel)).unwrap();
+        assert_eq!(read("application.yaml"), "spring:\n  application:\n    version: 0.0.1\n");
+        assert_eq!(read("static/extra.txt"), "from an extra source");
+        assert_eq!(read("com/example/App.class"), "bytecode");
+        let modified = fs::metadata(jdtls.join("migrations/V1__init.sql")).unwrap().modified().unwrap();
+        assert_eq!(modified, unchanged, "identical files are not rewritten");
+
+        overlay(&tmp.path().join("missing"), &tmp.path().join("missing"), &jdtls).unwrap();
+    }
+
+    #[test]
+    fn builds_without_processed_resources_run_nothing() {
+        let tmp = tempfile::tempdir().unwrap();
+        // Running this command would fail the call.
+        let missing = tmp.path().join("missing-gradle").to_string_lossy().into_owned();
+        let model =
+            BuildModel { gradle: vec![GradleModel { command: missing, ..Default::default() }], ..Default::default() };
+        let class_paths = [tmp.path().join("bin/main").to_string_lossy().into_owned()];
+        let tasks =
+            process_launch_resources(tmp.path(), &model, tmp.path(), &class_paths, &CancellationToken::new()).unwrap();
+        assert!(tasks.is_empty());
+    }
+}
@@ -4,9 +4,17 @@
 //! for the Build tool window, the Problems view and the editor.

 mod fingerprint;
+mod launch_resources;
 mod parse;
+mod single;
+mod up_to_date;

-pub use fingerprint::{LaunchFingerprint, launch_fingerprint};
+pub use fingerprint::{LaunchFingerprint, launch_fingerprint, menu_fingerprint};
+pub use launch_resources::process_launch_resources;
+pub use single::{
+    SingleFileTarget, compile_file_plan, find_javac, javac_step, single_file_classpath, single_file_target,
+};
+pub use up_to_date::{menu_up_to_date, record_menu_build};

 use crate::process::{self, Stream};
 use crate::{BuildModel, maven};
@@ -7,6 +7,7 @@
 //! [`super::resolve`] turns them into IDE coordinates.

 use ide_core::Severity;
+pub(crate) use ide_core::output_encoding::strip_ansi;
 use serde_json::Value;

 /// Output dialect of a build step.
@@ -98,32 +99,6 @@ pub(crate) struct OutputParser {
    pending: Option<Pending>,
 }

-/// Removes ANSI escape sequences (colour, cursor movement).
-pub(crate) fn strip_ansi(s: &str) -> String {
-    if !s.contains('\u{1b}') {
-        return s.to_owned();
-    }
-    let mut out = String::with_capacity(s.len());
-    let mut chars = s.chars().peekable();
-    while let Some(c) = chars.next() {
-        if c != '\u{1b}' {
-            out.push(c);
-            continue;
-        }
-        if chars.peek() == Some(&'[') {
-            chars.next();
-            for c in chars.by_ref() {
-                if c.is_ascii_alphabetic() || c == '~' {
-                    break;
-                }
-            }
-        } else {
-            chars.next();
-        }
-    }
-    out
-}
-
 fn num(s: &str) -> Option<u32> {
    s.trim().parse().ok().filter(|n| *n > 0)
 }
@@ -0,0 +1,331 @@
+//! Single-file Java compilation with `javac` for the Compile File action (#2476).
+//!
+//! A module build (`-pl … -am`) still walks every upstream module, and packaging
+//! walks the whole reactor; compiling one file only needs that file's module
+//! outputs plus its dependency classpath, which is much faster to assemble.
+use super::*;
+use std::fs;
+use std::process::Command as ProcCommand;
+
+/// Output file of the `dependency:build-classpath` probe, relative to each module.
+const CP_FILE: &str = "target/ide-compile-cp.txt";
+
+/// A `.java` source and the module build that compiles it.
+pub struct SingleFileTarget {
+    /// Absolute module directory.
+    pub module: PathBuf,
+    /// The source is under `src/test`: compile to test classes with the test classpath.
+    pub test: bool,
+    /// Absolute `target/classes` (or `target/test-classes`).
+    pub out_dir: PathBuf,
+    /// Reactor root building the module (owns the Maven command and options).
+    pub top: maven::MavenProject,
+    /// `-pl` value of the module relative to the reactor root; empty when the
+    /// module is the root itself.
+    pub pl: String,
+    /// `maven.compiler.release` (or equivalent) of the module's project.
+    pub java_version: Option<u32>,
+}
+
+/// Locates the linked Maven module owning a `.java` source file.
+pub fn single_file_target(root: &Path, model: &BuildModel, file: &Path) -> IdeResult<SingleFileTarget> {
+    if file.extension().is_none_or(|ext| ext != "java") {
+        return Err(IdeError::invalid("only Java sources can be compiled as a single file"));
+    }
+    if !file.is_file() {
+        return Err(IdeError::not_found("no Java source found for single-file compilation"));
+    }
+    let Some(m) = &model.maven else {
+        return Err(IdeError::not_found("no linked Maven project found for module"));
+    };
+    // The innermost project whose directory contains the file owns it; the
+    // aggregator root (`""`) is a prefix of everything, so take the longest match.
+    let rel = rel_str(root, file);
+    let project = m
+        .projects
+        .iter()
+        .filter(|p| rel == p.dir || rel.starts_with(&format!("{}/", p.dir)) || p.dir.is_empty())
+        .max_by_key(|p| p.dir.len())
+        .ok_or_else(|| IdeError::not_found("no linked Maven project found for module"))?;
+    if project.packaging == "pom" {
+        // An aggregator owns no sources itself: a file that only matches it
+        // belongs to no linked module.
+        return Err(IdeError::not_found("no linked Maven project found for module"));
+    }
+    let module = subdir(root, &project.dir);
+    let top = maven_reactor(m, &project.dir)
+        .ok_or_else(|| IdeError::not_found("no linked Maven project found for module"))?;
+    let top_dir = subdir(root, &top.dir);
+    let pl = module.strip_prefix(&top_dir).map(|r| r.to_string_lossy().replace('\\', "/")).unwrap_or_default();
+    // Test sources see the test classpath and compile into test classes.
+    let parts: Vec<&str> = rel.split('/').collect();
+    let test = parts.windows(2).any(|w| w == ["src", "test"]);
+    let out_dir = module.join("target").join(if test { "test-classes" } else { "classes" });
+    Ok(SingleFileTarget { module, test, out_dir, top: top.clone(), pl, java_version: project.java_version })
+}
+
+/// `javac` for the module: its `JAVA_HOME`, else `PATH`.
+pub fn find_javac(env: &BTreeMap<String, String>) -> String {
+    let home = env.get("JAVA_HOME").map(Path::new);
+    let exe = if cfg!(windows) { "javac.exe" } else { "javac" };
+    if let Some(javac) = home.map(|h| h.join("bin").join(exe)).filter(|p| p.is_file()) {
+        return javac.to_string_lossy().into_owned();
+    }
+    exe.to_owned()
+}
+
+/// Dependency classpath of the module: existing reactor class directories plus
+/// `mvn dependency:build-classpath` (which resolves without compiling anything,
+/// so `-am` stays cheap). Entries are platform-separated (`;` on Windows).
+pub fn single_file_classpath(root: &Path, model: &BuildModel, target: &SingleFileTarget) -> IdeResult<String> {
+    let mut entries: Vec<String> = Vec::new();
+    // The module's own classes first: the file almost always uses classes next to it.
+    let own: &[&str] = if target.test { &["target/test-classes", "target/classes"] } else { &["target/classes"] };
+    for out in own {
+        let path = target.module.join(out);
+        if path.is_dir() {
+            entries.push(path.to_string_lossy().into_owned());
+        }
+    }
+    if let Some(m) = &model.maven {
+        // Upstream reactor modules first: their freshly built classes shadow
+        // installed snapshots of the same artifacts.
+        for project in maven_launch_projects(root, m, &target.module, target.test) {
+            let dir = subdir(root, &project.dir);
+            if dir == target.module {
+                continue;
+            }
+            for classes in ["target/classes", "target/test-classes"] {
+                let path = dir.join(classes);
+                if path.is_dir() {
+                    entries.push(path.to_string_lossy().into_owned());
+                }
+            }
+        }
+    }
+    let probe = CP_FILE;
+    let mut args = vec!["-B", "-Dstyle.color=never", maven::LOCAL_REPO_TRACKING_ARG];
+    args.extend(target.top.options.iter().map(String::as_str));
+    if !target.pl.is_empty() {
+        args.extend(["-pl", target.pl.as_str(), "-am"]);
+    }
+    let output_arg = format!("-Dmdep.outputFile={probe}");
+    args.extend(["dependency:build-classpath", output_arg.as_str(), "-Dmdep.includeScope=test", "-q"]);
+    let top_dir = subdir(root, &target.top.dir);
+    let status = ProcCommand::new(&target.top.command)
+        .args(&args)
+        .envs(&target.top.env)
+        .current_dir(&top_dir)
+        .output()
+        .map_err(|e| IdeError::internal(format!("cannot resolve the compile classpath: {e}")))?;
+    if !status.status.success() {
+        let stderr = String::from_utf8_lossy(&status.stderr);
+        let tail: Vec<&str> = stderr.lines().filter(|l| !l.trim().is_empty()).collect();
+        let tail = tail.iter().rev().take(5).rev().cloned().collect::<Vec<_>>().join("\n");
+        return Err(IdeError::internal(format!("cannot resolve the compile classpath:\n{tail}")));
+    }
+    // With `-am` every visited module writes its own file; only this module's counts.
+    let cp_file = target.module.join(probe);
+    let content = fs::read_to_string(&cp_file).map_err(|e| {
+        IdeError::internal(format!("cannot resolve the compile classpath: {} ({e})", cp_file.display()))
+    })?;
+    let separator = if cfg!(windows) { ';' } else { ':' };
+    entries.extend(content.split([separator, '\n', '\r']).map(str::trim).filter(|e| !e.is_empty()).map(str::to_owned));
+    entries.dedup();
+    Ok(entries.join(&separator.to_string()))
+}
+
+/// The `javac` step compiling `file` into `target`'s output directory.
+pub fn javac_step(target: &SingleFileTarget, file: &Path, classpath: &str, javac: &str) -> IdeResult<BuildStep> {
+    fs::create_dir_all(&target.out_dir)
+        .map_err(|e| IdeError::internal(format!("cannot create {} ({e})", target.out_dir.display())))?;
+    let mut step = BuildStep::new(javac, &[], &target.module, Format::Gradle);
+    // Raw `javac` prints `path:line: error: …`, the same form Gradle forwards.
+    step.args.push("-encoding".into());
+    step.args.push("UTF-8".into());
+    step.args.push("-d".into());
+    step.args.push(target.out_dir.to_string_lossy().into_owned());
+    step.args.push("-cp".into());
+    step.args.push(classpath.to_owned());
+    // Classes of the module that were never built (or are older than their source) are read
+    // from its sources; only `file` is written, so a test compile never puts main classes
+    // into test-classes.
+    let sources = source_roots(target);
+    if !sources.is_empty() {
+        let separator = if cfg!(windows) { ";" } else { ":" };
+        step.args.push("-sourcepath".into());
+        step.args.push(sources.iter().map(|p| p.to_string_lossy()).collect::<Vec<_>>().join(separator));
+        step.args.push("-implicit:none".into());
+    }
+    if let Some(version) = target.java_version {
+        step.args.push("--release".into());
+        step.args.push(version.to_string());
+    }
+    step.args.push(file.to_string_lossy().into_owned());
+    step.env = target.top.env.clone();
+    Ok(step)
+}
+
+/// The module's Java source folders that exist, generated ones included; test sources see
+/// the main ones too.
+fn source_roots(target: &SingleFileTarget) -> Vec<PathBuf> {
+    let scopes: &[(&str, &str)] = if target.test {
+        &[("src/test/java", "target/generated-test-sources"), ("src/main/java", "target/generated-sources")]
+    } else {
+        &[("src/main/java", "target/generated-sources")]
+    };
+    let mut roots = Vec::new();
+    for (src, generated) in scopes {
+        roots.push(target.module.join(src));
+        let generated = target.module.join(generated);
+        let mut dirs: Vec<PathBuf> =
+            fs::read_dir(&generated).into_iter().flatten().flatten().map(|d| d.path()).filter(|p| p.is_dir()).collect();
+        dirs.sort();
+        roots.extend(dirs);
+    }
+    roots.retain(|p| p.is_dir());
+    roots
+}
+
+/// Resolves the classpath and plans the `javac` invocation for one Java source.
+pub fn compile_file_plan(root: &Path, model: &BuildModel, file: &Path) -> IdeResult<BuildPlan> {
+    let target = single_file_target(root, model, file)?;
+    let classpath = single_file_classpath(root, model, &target)?;
+    let javac = find_javac(&target.top.env);
+    let name = file.file_name().map(|n| n.to_string_lossy().into_owned()).unwrap_or_default();
+    Ok(BuildPlan {
+        system: BuildSystem::Maven,
+        title: format!("Compile {name}"),
+        steps: vec![javac_step(&target, file, &classpath, &javac)?],
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn write(root: &Path, rel: &str, value: &str) {
+        let path = root.join(rel);
+        fs::create_dir_all(path.parent().unwrap()).unwrap();
+        fs::write(path, value).unwrap();
+    }
+
+    fn fixture(root: &Path) -> BuildModel {
+        write(
+            root,
+            "pom.xml",
+            r#"<project><groupId>demo</groupId><artifactId>parent</artifactId><version>1</version>
+            <packaging>pom</packaging><modules><module>lib</module><module>app</module></modules></project>"#,
+        );
+        write(
+            root,
+            "lib/pom.xml",
+            r#"<project><parent><groupId>demo</groupId><artifactId>parent</artifactId><version>1</version></parent>
+            <artifactId>lib</artifactId></project>"#,
+        );
+        write(
+            root,
+            "app/pom.xml",
+            r#"<project><parent><groupId>demo</groupId><artifactId>parent</artifactId><version>1</version></parent>
+            <artifactId>app</artifactId><dependencies><dependency><groupId>demo</groupId><artifactId>lib</artifactId><version>1</version></dependency></dependencies></project>"#,
+        );
+        write(root, "app/src/main/java/demo/Main.java", "class Main {}");
+        write(root, "app/src/test/java/demo/MainTest.java", "class MainTest {}");
+        write(root, "lib/target/classes/Lib.class", "built");
+        crate::load(root, &root.join("cache"), None)
+    }
+
+    #[test]
+    fn target_finds_the_owning_module_and_the_test_scope() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path();
+        let model = fixture(root);
+        let main = single_file_target(root, &model, &root.join("app/src/main/java/demo/Main.java")).unwrap();
+        assert_eq!(main.module, root.join("app"));
+        assert!(!main.test);
+        assert_eq!(main.out_dir, root.join("app/target/classes"));
+        let test = single_file_target(root, &model, &root.join("app/src/test/java/demo/MainTest.java")).unwrap();
+        assert!(test.test);
+        assert_eq!(test.out_dir, root.join("app/target/test-classes"));
+    }
+
+    #[test]
+    fn target_rejects_non_java_missing_and_unlinked_files() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path();
+        let model = fixture(root);
+        write(root, "app/pom.xml", "<project/>");
+        assert!(single_file_target(root, &model, &root.join("app/pom.xml")).is_err());
+        assert!(single_file_target(root, &model, &root.join("app/src/main/java/demo/Missing.java")).is_err());
+        write(root, "other/Main.java", "class Main {}");
+        assert!(single_file_target(root, &model, &root.join("other/Main.java")).is_err());
+    }
+
+    #[test]
+    fn javac_prefers_the_module_jdk_and_scopes_test_sources() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path();
+        let model = fixture(root);
+        let file = root.join("app/src/test/java/demo/MainTest.java");
+        let target = single_file_target(root, &model, &file).unwrap();
+        let jdk = root.join("jdk");
+        fs::create_dir_all(jdk.join("bin")).unwrap();
+        let exe = if cfg!(windows) { "javac.exe" } else { "javac" };
+        fs::write(jdk.join("bin").join(exe), "fake").unwrap();
+        let mut target = target;
+        target.top.env.insert("JAVA_HOME".into(), jdk.to_string_lossy().into_owned());
+        let step = javac_step(&target, &file, "/repo/a.jar", &find_javac(&target.top.env)).unwrap();
+        assert!(step.program.ends_with(exe));
+        assert!(target.out_dir.is_dir());
+        let args = step.args.join(" ");
+        assert!(args.contains("-encoding UTF-8"), "{args}");
+        assert!(args.contains("-d ") && args.contains("test-classes"), "{args}");
+        assert!(args.contains("-cp /repo/a.jar"), "{args}");
+        assert!(args.contains("MainTest.java"), "{args}");
+        // Test sources resolve their own and the main sources, without writing them.
+        let sourcepath = step.args.iter().position(|a| a == "-sourcepath").map(|i| step.args[i + 1].clone());
+        let roots = sourcepath.expect("a -sourcepath");
+        assert!(roots.contains("src/test/java") && roots.contains("src/main/java"), "{roots}");
+        assert!(step.args.iter().any(|a| a == "-implicit:none"), "{args}");
+        // No `JAVA_HOME`: plain `javac` from `PATH`.
+        assert_eq!(find_javac(&BTreeMap::new()), exe);
+    }
+
+    /// The classpath probe runs the project's Maven command with `-pl … -am`
+    /// `dependency:build-classpath`; the fake `mvnw` below only records its
+    /// arguments and writes a canned classpath file for the `-pl` module.
+    #[cfg(unix)]
+    #[test]
+    fn classpath_probe_writes_one_file_per_module_and_reads_the_compiled_one() {
+        use std::os::unix::fs::PermissionsExt;
+
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path();
+        fixture(root);
+        let script = root.join("mvnw");
+        fs::write(
+            &script,
+            "#!/bin/sh\npl='.'\nwhile [ $# -gt 0 ]; do\nif [ \"$1\" = \"-pl\" ]; then pl=\"$2\"; shift 2; else shift; fi\ndone\nmkdir -p \"$pl/target\"\nprintf '/repo/a.jar:/repo/b.jar' > \"$pl/target/ide-compile-cp.txt\"\n",
+        )
+        .unwrap();
+        fs::set_permissions(&script, fs::Permissions::from_mode(0o755)).unwrap();
+        let model = crate::load(root, &root.join("cache"), None);
+        let file = root.join("app/src/main/java/demo/Main.java");
+        let target = single_file_target(root, &model, &file).unwrap();
+        assert!(target.top.command.ends_with("mvnw"));
+        assert_eq!(target.pl, "app");
+        let cp = single_file_classpath(root, &model, &target).unwrap();
+        let separator = ':';
+        // Reactor class directories first, so fresh outputs shadow installed snapshots.
+        let lib = root.join("lib/target/classes").to_string_lossy().into_owned();
+        assert!(cp.starts_with(&lib), "{cp}");
+        // The module's own classes come first once they exist.
+        fs::create_dir_all(root.join("app/target/classes")).unwrap();
+        let cp = single_file_classpath(root, &model, &target).unwrap();
+        let own = root.join("app/target/classes").to_string_lossy().into_owned();
+        assert!(cp.starts_with(&format!("{own}:{lib}")), "{cp}");
+        assert!(cp.contains("/repo/a.jar"), "{cp}");
+        assert!(cp.contains("/repo/b.jar"), "{cp}");
+        assert!(cp.split(separator).count() >= 3, "{cp}");
+    }
+}
@@ -0,0 +1,253 @@
+//! Up-to-date checks for menu builds (`Build Project` / `Build Module`, #2476).
+//!
+//! Menu builds have no fingerprint record on the launch path, so a second
+//! click always pays a full offline Maven run. This module stores one JSON
+//! record per scope (project, or one module file's build) after a successful
+//! build and compares it before the next build; on a match the build is
+//! skipped entirely. Like launch reuse, timestamp-preserving external edits
+//! are re-stamped first so Maven actually recompiles them on the next run.
+
+use super::*;
+use std::fs;
+
+/// Record of the last successful menu build of one scope.
+#[derive(serde::Serialize, serde::Deserialize)]
+struct Record {
+    fingerprint: LaunchFingerprint,
+    /// Output folders of the built projects that existed after the build, relative to the
+    /// project. Deleting one outside the IDE (`mvn clean` in a terminal) leaves every input
+    /// unchanged, so a build is skipped only while they are all still there.
+    #[serde(default)]
+    outputs: Vec<PathBuf>,
+}
+
+/// The build's output folders that exist now, relative to `root`.
+fn existing_outputs(root: &Path, fingerprint: &LaunchFingerprint) -> Vec<PathBuf> {
+    fingerprint
+        .project_dirs()
+        .flat_map(|dir| ["target/classes", "target/test-classes"].map(|out| dir.join(out)))
+        .filter(|rel| root.join(rel).is_dir())
+        .collect()
+}
+
+fn record_path(cache: &Path, root: &Path, file: Option<&Path>) -> PathBuf {
+    let base = cache.join("menu-build").join("records");
+    match file {
+        Some(f) => base.join(f.strip_prefix(root).unwrap_or(f)).with_extension("json"),
+        None => base.join("project.json"),
+    }
+}
+
+/// Re-stamps files whose content changed without advancing mtime (git
+/// checkout, external sync): timestamp-incremental Maven would otherwise skip
+/// them as "unchanged" and the rebuild would compile nothing.
+fn restamp(root: &Path, edits: &[PathBuf]) {
+    let now = std::time::SystemTime::now();
+    for rel in edits {
+        let path = root.join(rel);
+        let _ = fs::File::options().write(true).open(&path).and_then(|f| f.set_modified(now)).or_else(|_| {
+            fs::OpenOptions::new()
+                .create(true)
+                .truncate(false)
+                .write(true)
+                .open(&path)
+                .and_then(|f| f.set_modified(now))
+        });
+    }
+}
+
+/// Whether the last successful menu build of this scope is still fresh: its
+/// record exists and the fingerprint matches. On a mismatch caused by
+/// timestamp-preserving external edits the files are re-stamped first, so the
+/// rebuild Maven runs actually recompiles them.
+pub fn menu_up_to_date(
+    cache: &Path,
+    root: &Path,
+    model: &BuildModel,
+    file: Option<&Path>,
+    cancel: &CancellationToken,
+) -> bool {
+    let current = match menu_fingerprint(root, model, file, cancel) {
+        Ok(Some(current)) => current,
+        _ => return false,
+    };
+    let bytes = match fs::read(record_path(cache, root, file)) {
+        Ok(bytes) => bytes,
+        _ => return false,
+    };
+    let record: Record = match serde_json::from_slice(&bytes) {
+        Ok(record) => record,
+        _ => return false,
+    };
+    if record.fingerprint != current {
+        restamp(root, &record.fingerprint.mtime_preserved_edits(&current));
+        return false;
+    }
+    record.outputs.iter().all(|out| root.join(out).is_dir())
+}
+
+/// Records a successful menu build as the baseline for future up-to-date
+/// checks. Failures are swallowed: a missing record only costs one Maven run.
+pub fn record_menu_build(
+    cache: &Path,
+    root: &Path,
+    model: &BuildModel,
+    file: Option<&Path>,
+    cancel: &CancellationToken,
+) {
+    let current = match menu_fingerprint(root, model, file, cancel) {
+        Ok(Some(current)) => current,
+        _ => return,
+    };
+    // A torn write fails to parse later and is treated as "no record", so a
+    // plain write is enough.
+    let path = record_path(cache, root, file);
+    if path.parent().is_some_and(|p| fs::create_dir_all(p).is_err()) {
+        return;
+    }
+    let outputs = existing_outputs(root, &current);
+    let bytes = match serde_json::to_vec(&Record { fingerprint: current, outputs }) {
+        Ok(bytes) => bytes,
+        _ => return,
+    };
+    let _ = fs::write(&path, bytes);
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn write(root: &Path, rel: &str, value: &str) {
+        let path = root.join(rel);
+        fs::create_dir_all(path.parent().unwrap()).unwrap();
+        fs::write(path, value).unwrap();
+    }
+
+    fn fixture(root: &Path) -> BuildModel {
+        maven_two_module(root, "int x;");
+        BuildModel { maven: Some(maven::load(root)), ..Default::default() }
+    }
+
+    fn maven_two_module(root: &Path, extra_app: &str) {
+        write(
+            root,
+            "pom.xml",
+            r#"<project><modelVersion>4.0.0</modelVersion><groupId>t</groupId><artifactId>root</artifactId><version>1</version><packaging>pom</packaging><modules><module>app</module><module>lib</module></modules></project>"#,
+        );
+        write(
+            root,
+            "app/pom.xml",
+            r#"<project><modelVersion>4.0.0</modelVersion><parent><groupId>t</groupId><artifactId>root</artifactId><version>1</version></parent><artifactId>app</artifactId><dependencies><dependency><groupId>t</groupId><artifactId>lib</artifactId><version>1</version></dependency></dependencies></project>"#,
+        );
+        write(
+            root,
+            "lib/pom.xml",
+            r#"<project><modelVersion>4.0.0</modelVersion><parent><groupId>t</groupId><artifactId>root</artifactId><version>1</version></parent><artifactId>lib</artifactId></project>"#,
+        );
+        write(root, "app/src/main/java/t/A.java", &format!("package t; public class A {{ {extra_app} }}"));
+        write(root, "lib/src/main/java/t/L.java", "package t; public class L {}");
+    }
+
+    #[test]
+    fn record_then_skip_without_changes() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path();
+        let model = fixture(root);
+        let cache = root.join("cache");
+        let file = root.join("app/src/main/java/t/A.java");
+        let cancel = CancellationToken::new();
+
+        assert!(!menu_up_to_date(&cache, root, &model, Some(&file), &cancel));
+        record_menu_build(&cache, root, &model, Some(&file), &cancel);
+        assert!(menu_up_to_date(&cache, root, &model, Some(&file), &cancel));
+        // The project scope is independent of the module scope.
+        assert!(!menu_up_to_date(&cache, root, &model, None, &cancel));
+        record_menu_build(&cache, root, &model, None, &cancel);
+        assert!(menu_up_to_date(&cache, root, &model, None, &cancel));
+    }
+
+    #[test]
+    fn deleted_outputs_invalidate_the_record() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path();
+        let model = fixture(root);
+        let cache = root.join("cache");
+        let cancel = CancellationToken::new();
+        for out in ["app/target/classes/t", "lib/target/classes/t", "app/target/test-classes"] {
+            fs::create_dir_all(root.join(out)).unwrap();
+        }
+        record_menu_build(&cache, root, &model, None, &cancel);
+        assert!(menu_up_to_date(&cache, root, &model, None, &cancel));
+        // `mvn clean` outside the IDE: no input changed, but the classes are gone.
+        fs::remove_dir_all(root.join("lib/target")).unwrap();
+        assert!(!menu_up_to_date(&cache, root, &model, None, &cancel));
+        fs::create_dir_all(root.join("lib/target/classes")).unwrap();
+        assert!(menu_up_to_date(&cache, root, &model, None, &cancel));
+    }
+
+    #[test]
+    fn edit_invalidates_and_rerecord_restores_skip() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path();
+        let model = fixture(root);
+        let cache = root.join("cache");
+        let file = root.join("app/src/main/java/t/A.java");
+        let cancel = CancellationToken::new();
+
+        record_menu_build(&cache, root, &model, Some(&file), &cancel);
+        assert!(menu_up_to_date(&cache, root, &model, Some(&file), &cancel));
+        write(root, "app/src/main/java/t/A.java", "package t; public class A { int y; }");
+        assert!(!menu_up_to_date(&cache, root, &model, Some(&file), &cancel));
+        record_menu_build(&cache, root, &model, Some(&file), &cancel);
+        assert!(menu_up_to_date(&cache, root, &model, Some(&file), &cancel));
+    }
+
+    #[test]
+    fn mtime_preserved_edit_forces_rebuild_and_restamps() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path();
+        let model = fixture(root);
+        let cache = root.join("cache");
+        let rel = "app/src/main/java/t/A.java";
+        let file = root.join(rel);
+        let cancel = CancellationToken::new();
+
+        record_menu_build(&cache, root, &model, Some(&file), &cancel);
+        let before = fs::metadata(&file).unwrap().modified().unwrap();
+        // Same length, same mtime, different content: an external edit that
+        // preserved the timestamp.
+        write(root, rel, "package t; public class A { int z; }");
+        fs::File::options().write(true).open(&file).and_then(|f| f.set_modified(before)).unwrap();
+        assert!(!menu_up_to_date(&cache, root, &model, Some(&file), &cancel));
+        assert!(fs::metadata(&file).unwrap().modified().unwrap() > before);
+    }
+
+    #[test]
+    fn corrupt_record_never_skips() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path();
+        let model = fixture(root);
+        let cache = root.join("cache");
+        let file = root.join("app/src/main/java/t/A.java");
+        let cancel = CancellationToken::new();
+
+        fs::create_dir_all(cache.join("menu-build").join("records").join("app/src/main/java/t")).unwrap();
+        fs::write(cache.join("menu-build").join("records").join("app/src/main/java/t/A.json"), b"{not json").unwrap();
+        assert!(!menu_up_to_date(&cache, root, &model, Some(&file), &cancel));
+    }
+
+    #[test]
+    fn unsupported_builds_never_skip() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path();
+        let model = fixture(root);
+        write(root, "other/Main.java", "public class Main {}");
+        let cache = root.join("cache");
+        let cancel = CancellationToken::new();
+
+        // A file outside any non-aggregator module has no fingerprint.
+        let stray = root.join("other/Main.java");
+        record_menu_build(&cache, root, &model, Some(&stray), &cancel);
+        assert!(!menu_up_to_date(&cache, root, &model, Some(&stray), &cancel));
+    }
+}
@@ -8,6 +8,24 @@ def canon = { File f -> f.canonicalPath }
 // the identity path (`:lib:core`, `:buildSrc`) tells them apart and addresses their tasks
 // from the root build. Internal API, hence the fallback.
 def identity = { Project p -> try { p.identityPath.toString() } catch (ignored) { p.path } }
+// The task copying a source set's resources and its output folder, when it changes them on the
+// way (filters, `expand`, renames, includes / excludes, extra sources): jdtls copies resources as
+// they are, so launches take them from Gradle. Internal API, hence the fallback.
+def processedResources = { Project p, s ->
+    try {
+        def task = p.tasks.findByName(s.processResourcesTaskName)
+        if (task == null) return null
+        def specs = 0
+        def changes = task.mainSpec.sourcePaths.size() > 1
+        task.rootSpec.buildRootResolver().walk { r ->
+            specs++
+            changes |= !r.allCopyActions.isEmpty() || !r.allIncludes.isEmpty() || !r.allExcludes.isEmpty() ||
+                !r.allIncludeSpecs.isEmpty() || !r.allExcludeSpecs.isEmpty() || r.destPath.segments.length > 0
+        }
+        // The root and main specs exist for every task; more come from `from(…) { … }`.
+        (changes || specs > 2) ? [task: task.name, dir: canon(task.destinationDir)] : null
+    } catch (ignored) { null }
+}

 def describe = { Project p ->
    def m = [path: identity(p), dir: canon(p.projectDir), buildDir: canon(p.layout.buildDirectory.get().asFile)]
@@ -26,7 +44,10 @@ def describe = { Project p ->
            def test = s.name == 'test' || s.name == 'testFixtures' || suites.contains(s.name) ||
                s.name.endsWith('Test') || s.name.endsWith('Tests') ||
                (!src.isEmpty() && src.every { ideaTests.contains(it) })
-            [name: s.name, test: test, src: src, res: res]
+            def set = [name: s.name, test: test, src: src, res: res]
+            def processed = processedResources(p, s)
+            if (processed != null) set.processed = processed
+            set
        }
    }
    try {
@@ -335,6 +335,19 @@ pub struct ReportedSourceSet {
    /// Absolute resource folders.
    #[serde(default)]
    pub res: Vec<String>,
+    /// Set when Gradle changes the resources while copying them (filters, `expand`, renames,
+    /// includes / excludes, extra sources), which the Java language server's copy does not.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub processed: Option<ProcessedResources>,
+}
+
+/// The task that copies a source set's resources, and where it writes them.
+#[derive(Clone, Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
+pub struct ProcessedResources {
+    /// Task name in its project (`processResources`).
+    pub task: String,
+    /// Absolute output folder.
+    pub dir: String,
 }

 const MODEL_LINE: &str = "@openidea-model\t";
@@ -73,6 +73,9 @@ struct GradleCache {
    /// The projects as Gradle evaluated them; empty in caches from before they were read.
    #[serde(default)]
    projects: Vec<ReportedProject>,
+    /// The projects report how their resources are processed; older caches are stale.
+    #[serde(default)]
+    resources: bool,
 }

 /// Init script printing every project, its source sets and Java version once configured.
@@ -244,11 +247,12 @@ pub fn apply_gradle_cache(root: &Path, cache_dir: &Path, model: &mut GradleModel
            }
        }
    }
-    // A successful sync from before projects were cached has not seen computed includes.
-    let without_projects = cache.error.is_none() && cache.projects.is_empty();
+    // A successful sync from before projects were cached has not seen computed includes,
+    // nor, before resources were read, which launches need Gradle to process them.
+    let outdated = cache.error.is_none() && (cache.projects.is_empty() || !cache.resources);
    model.synced_at = Some(cache.at_ms);
    model.sync_error = cache.error;
-    model.stale = without_projects || cache.hash != hash_gradle_files(&gradle::build_files(root, model));
+    model.stale = outdated || cache.hash != hash_gradle_files(&gradle::build_files(root, model));
 }

 // ---------------------------------------------------------------- processes
@@ -462,6 +466,7 @@ pub fn sync_gradle_model(
        error: (!out.success).then(|| error_summary(&out.text)),
        tasks: if out.success { gradle::parse_tasks(&out.text) } else { Vec::new() },
        projects,
+        resources: true,
    };
    write_cache(&cache_dir, GRADLE_CACHE, &cache)
 }
@@ -713,6 +718,7 @@ mod tests {
            test,
            src: src.iter().map(|p| core_src(p)).collect(),
            res: res.iter().map(|p| core_src(p)).collect(),
+            processed: None,
        };
        // A settings script's own println comes first; task listing lines follow.
        let out = format!(
@@ -737,6 +743,15 @@ mod tests {
        );
        let projects = parse_reported_projects(&out);
        assert_eq!(projects.len(), 3, "the malformed line is skipped");
+        // As the init script prints a set whose resources Gradle filters or renames.
+        let processed = parse_reported_projects(&format!(
+            "@openidea-model\t{}",
+            r#"{"path":":","dir":"/p","sourceSets":[{"name":"main","test":false,"src":[],"res":[],"processed":{"task":"processResources","dir":"/p/build/resources/main"}}]}"#
+        ));
+        assert_eq!(
+            processed[0].source_sets[0].processed,
+            Some(gradle::ProcessedResources { task: "processResources".into(), dir: "/p/build/resources/main".into() })
+        );
        let src = |p: &str| PathBuf::from(core_src(p));
        assert_eq!(
            gradle::reported_source_roots(&projects),
@@ -802,12 +817,18 @@ mod tests {
                    ..Default::default()
                })
                .collect(),
+            resources: true,
        };
        write_cache(&cache_dir, GRADLE_CACHE, &cache).unwrap();
        apply_gradle_cache(&root, &cache_dir, &mut model);
        assert_eq!(model.projects[1].tasks.len(), 1);
        assert!(model.projects[0].tasks.iter().any(|t| t.name == "wrapper"), "root keeps its defaults");
        assert!(!model.stale);
+        write_cache(&cache_dir, GRADLE_CACHE, &GradleCache { resources: false, ..cache.clone() }).unwrap();
+        let mut m = gradle::load(&root).unwrap();
+        apply_gradle_cache(&root, &cache_dir, &mut m);
+        assert!(m.stale, "synced before resource processing was read");
+        write_cache(&cache_dir, GRADLE_CACHE, &cache).unwrap();

        let stale_after = |script: &str| {
            std::fs::write(root.join("app/build.gradle"), script).unwrap();
@@ -841,6 +862,7 @@ mod tests {
                dir: root.to_string_lossy().into_owned(),
                ..Default::default()
            }],
+            resources: true,
            ..Default::default()
        };
        write_cache(&cache_dir, GRADLE_CACHE, &cache).unwrap();
@@ -333,6 +333,23 @@ pub struct InlayHintsSettings {
    pub code_vision_usages: bool,
    /// Code vision: last author of a declaration (VCS blame) beside its usages.
    pub code_vision_author: bool,
+    /// Code vision: "no usages" on declarations nothing uses (IDEA shows it; clicking searches everywhere).
+    pub code_vision_no_usages: bool,
+    /// Code vision: where the usages go, above the declaration or at the end of its line.
+    pub code_vision_usages_position: CodeVisionPosition,
+    /// Code vision: where the author goes, above the declaration or at the end of its line.
+    pub code_vision_author_position: CodeVisionPosition,
+}
+
+/// Editor › Inlay Hints › Code Vision position (IDEA's "Default position for metrics").
+#[derive(Clone, Copy, Debug, Default, PartialEq, Serialize, Deserialize, Type)]
+#[serde(rename_all = "camelCase")]
+pub enum CodeVisionPosition {
+    /// A row of its own above the declaration.
+    #[default]
+    Top,
+    /// After the end of the declaration's first line.
+    Right,
 }

 impl Default for InlayHintsSettings {
@@ -344,6 +361,9 @@ impl Default for InlayHintsSettings {
            disabled_languages: Vec::new(),
            code_vision_usages: true,
            code_vision_author: true,
+            code_vision_no_usages: true,
+            code_vision_usages_position: CodeVisionPosition::Top,
+            code_vision_author_position: CodeVisionPosition::Top,
        }
    }
 }
@@ -273,6 +273,27 @@ pub enum DebugEvent {
        change_type: String,
        message: String,
    },
+    /// The session's capabilities changed after it started: fetch them again.
+    Capabilities,
+    JvmBackend {
+        state: JvmBackendState,
+    },
+}
+
+/// The backend that reads a debugged JVM's objects (instances, fields, referrers).
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, Type)]
+#[serde(tag = "kind", rename_all = "camelCase", rename_all_fields = "camelCase")]
+pub enum JvmBackendState {
+    Ready,
+    /// Turned off for the rest of the session: the debug adapter lacks what it needs.
+    Unsupported {
+        detail: String,
+    },
+    /// The Kotlin JDWP link only forwards now; inspecting stops.
+    Degraded {
+        detail: String,
+    },
+    Closed,
 }

 /// Where backend crates publish [`IdeEvent`]s. Cheap to clone via `Arc`.
@@ -294,3 +315,19 @@ pub struct NullSink;
 impl EventSink for NullSink {
    fn emit(&self, _: IdeEvent) {}
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    #[test]
+    fn jvm_backend_events_json_shape() {
+        let degraded = DebugEvent::JvmBackend { state: JvmBackendState::Degraded { detail: "x".into() } };
+        assert_eq!(
+            serde_json::to_value(&degraded).unwrap(),
+            json!({"kind": "jvmBackend", "state": {"kind": "degraded", "detail": "x"}})
+        );
+        assert_eq!(serde_json::to_value(DebugEvent::Capabilities).unwrap(), json!({"kind": "capabilities"}));
+    }
+}
@@ -21,6 +21,32 @@

 use std::collections::BTreeMap;

+/// Removes ANSI escape sequences (colour, cursor movement).
+pub fn strip_ansi(s: &str) -> String {
+    if !s.contains('\u{1b}') {
+        return s.to_owned();
+    }
+    let mut out = String::with_capacity(s.len());
+    let mut chars = s.chars().peekable();
+    while let Some(c) = chars.next() {
+        if c != '\u{1b}' {
+            out.push(c);
+            continue;
+        }
+        if chars.peek() == Some(&'[') {
+            chars.next();
+            for c in chars.by_ref() {
+                if c.is_ascii_alphabetic() || c == '~' {
+                    break;
+                }
+            }
+        } else {
+            chars.next();
+        }
+    }
+    out
+}
+
 /// Decodes complete lines or collected output from IDE-managed tools, preferring UTF-8.
 /// Windows launcher/native diagnostics fall back to the system ANSI code page.
 /// Callers must buffer incomplete lines so multibyte characters are never split.
@@ -72,6 +72,75 @@ pub fn kill_tree_of(pid: u32) {
    let _ = pid;
 }

+/// Whether a process with this id exists. PID reuse makes this conservative: a reused id keeps
+/// whatever was tied to the old process alive, a live process is never reported dead.
+pub fn process_alive(pid: u32) -> bool {
+    // `kill(0, 0)` would signal the caller's own process group.
+    if pid == 0 {
+        return false;
+    }
+    #[cfg(unix)]
+    {
+        // Above `i32::MAX` the id would wrap to a negative value, which `kill` reads as a group.
+        let Ok(pid) = libc::pid_t::try_from(pid) else { return false };
+        // SAFETY: signal 0 delivers nothing; it only checks that the process exists and is reachable.
+        unsafe { libc::kill(pid, 0) == 0 || std::io::Error::last_os_error().raw_os_error() == Some(libc::EPERM) }
+    }
+    #[cfg(windows)]
+    {
+        use std::os::windows::io::{AsRawHandle, FromRawHandle, OwnedHandle};
+        use windows_sys::Win32::Foundation::{ERROR_ACCESS_DENIED, GetLastError, STILL_ACTIVE};
+        use windows_sys::Win32::System::Threading::{
+            GetExitCodeProcess, OpenProcess, PROCESS_QUERY_LIMITED_INFORMATION,
+        };
+
+        // SAFETY: plain call; a non-null result is a process handle owned by this function.
+        let handle = unsafe { OpenProcess(PROCESS_QUERY_LIMITED_INFORMATION, 0, pid) };
+        if handle.is_null() {
+            // A process we may not open still exists.
+            // SAFETY: reads the thread's last error, set by the call above.
+            return unsafe { GetLastError() } == ERROR_ACCESS_DENIED;
+        }
+        // SAFETY: `handle` is valid and nobody else owns it.
+        let handle = unsafe { OwnedHandle::from_raw_handle(handle) };
+        let mut code = 0u32;
+        // SAFETY: a live process handle opened with query access; one out-parameter.
+        let ok = unsafe { GetExitCodeProcess(handle.as_raw_handle(), &mut code) };
+        // A handle that cannot be queried is not evidence that the process is gone.
+        ok == 0 || code == STILL_ACTIVE as u32
+    }
+    #[cfg(not(any(unix, windows)))]
+    {
+        true
+    }
+}
+
+/// The program and everything it started: an adapter may report the launcher shell, whose
+/// child is the debugged program, so stopping only the reported pid leaves the program alive.
+#[cfg(unix)]
+pub fn process_tree(pid: i32) -> Vec<i32> {
+    use std::collections::HashMap;
+    let Ok(out) = std::process::Command::new("ps").args(["-A", "-o", "pid=,ppid="]).output() else {
+        return vec![pid];
+    };
+    let mut children: HashMap<i32, Vec<i32>> = HashMap::new();
+    for line in String::from_utf8_lossy(&out.stdout).lines() {
+        let mut it = line.split_whitespace();
+        let (Some(p), Some(parent)) = (it.next(), it.next()) else { continue };
+        let (Ok(p), Ok(parent)) = (p.parse::<i32>(), parent.parse::<i32>()) else { continue };
+        children.entry(parent).or_default().push(p);
+    }
+    let mut tree = vec![pid];
+    let mut at = 0;
+    while at < tree.len() {
+        if let Some(kids) = children.get(&tree[at]) {
+            tree.extend(kids.iter().copied());
+        }
+        at += 1;
+    }
+    tree
+}
+
 /// [`kill_tree`] that first asks the group to exit: `SIGTERM`, then `grace`, then `SIGKILL`.
 /// Git removes its `index.lock` on `SIGTERM`, where `SIGKILL` leaves it behind and the next
 /// git command fails. Windows has no such signal to send from outside the child's console,
@@ -730,6 +799,30 @@ fn append_path_entries(path: &OsStr, extra: &OsStr) -> OsString {
 mod tests {
    use super::*;

+    #[test]
+    fn process_alive_tells_live_from_dead_pids() {
+        assert!(process_alive(std::process::id()));
+        assert!(!process_alive(0));
+        let mut cmd = if cfg!(windows) {
+            let mut cmd = Command::new("cmd");
+            cmd.args(["/C", "exit 0"]);
+            cmd
+        } else {
+            Command::new("true")
+        };
+        let mut child = cmd.spawn().unwrap();
+        let pid = child.id();
+        child.wait().unwrap();
+        drop(child);
+        assert!(!process_alive(pid));
+    }
+
+    #[cfg(unix)]
+    #[test]
+    fn process_alive_rejects_ids_that_are_not_a_single_process() {
+        assert!(!process_alive(u32::MAX));
+    }
+
    #[cfg(unix)]
    #[test]
    fn kill_tree_of_reaches_what_the_child_started() {
@@ -1050,6 +1143,25 @@ mod tests {
        }
        assert_eq!(find_executable("nope", &search), None);
    }
+
+    #[cfg(unix)]
+    #[test]
+    fn process_tree_lists_the_child_of_a_shell() {
+        let mut cmd = Command::new("sh");
+        own_group(&mut cmd).args(["-c", "sleep 5 & wait"]);
+        let mut child = cmd.spawn().unwrap();
+        let sh_pid = child.id() as i32;
+        // The shell forks `sleep` after it starts, so the tree grows a moment later.
+        let deadline = std::time::Instant::now() + Duration::from_secs(2);
+        let mut tree = process_tree(sh_pid);
+        while tree.len() < 2 && std::time::Instant::now() < deadline {
+            std::thread::sleep(Duration::from_millis(20));
+            tree = process_tree(sh_pid);
+        }
+        kill_tree(&mut child);
+        assert_eq!(tree[0], sh_pid, "the pid itself comes first");
+        assert!(tree.len() >= 2 && tree[1..].iter().all(|p| *p != sh_pid), "the shell's child follows: {tree:?}");
+    }
 }

 #[cfg(all(test, unix))]
@@ -261,6 +261,9 @@ pub enum WhileRunning {
 pub struct McpServerSettings {
    /// Serve the tools on 127.0.0.1 while the IDE runs.
    pub enabled: bool,
+    /// Run the tools that change the project or start processes without asking first
+    /// (IDEA's brave mode). Read from the IDE's own settings only, never a project's.
+    pub run_without_confirmation: bool,
 }

 /// Absent profiles retain automatic discovery for existing projects.
@@ -453,6 +456,10 @@ pub struct GeneralSettings {
    /// Where updates come from: `auto` (the CNB mirror for the Chinese UI, else GitHub) | `github` | `cnb`.
    pub update_source: String,
    pub auto_check_updates: bool,
+    /// System Settings › Updates: open the What's New page after the app starts a new version.
+    pub show_whats_new_after_update: bool,
+    /// The version the app last started with, so the first start after an update is recognisable.
+    pub last_run_version: String,
 }

 impl Default for GeneralSettings {
@@ -471,6 +478,8 @@ impl Default for GeneralSettings {
            project_name: String::new(),
            update_source: "auto".into(),
            auto_check_updates: true,
+            show_whats_new_after_update: true,
+            last_run_version: String::new(),
        }
    }
 }
@@ -574,7 +583,7 @@ pub struct AppearanceSettings {
    pub project_colors_in_toolbar: bool,
    pub dnd_with_alt_only: bool,
    pub menu_icons: bool,
-    /// `hamburger` | `merged`: main menu presentation in the frameless title bar.
+    /// `hamburger` | `merged` | `toolbar` (a separate row): main menu presentation in the frameless title bar.
    pub main_menu_style: String,
    /// A lone Alt press focuses or opens the main menu (IDEA `ide.windows.main.menu.focus.on.alt`).
    pub main_menu_alt: bool,
@@ -11,12 +11,17 @@ tempfile.workspace = true
 ide-core.workspace = true
 ide-language.workspace = true
 ide-jsonrpc.workspace = true
+ide-jvm-attach.workspace = true
+ide-jvm-inspect.workspace = true
+ide-jdwp.workspace = true
+ide-heap.workspace = true
 serde.workspace = true
 serde_json.workspace = true
 specta.workspace = true
 tokio = { workspace = true, features = ["net"] }
 parking_lot.workspace = true
 memchr = "2"
+percent-encoding = "2"
 portable-pty.workspace = true
 tracing.workspace = true

@@ -342,4 +342,16 @@ mod tests {
        assert_eq!(error_message(&msg), "cannot find a.py");
        assert_eq!(error_message(&json!({"message": "nope"})), "nope");
    }
+
+    #[test]
+    fn a_failed_launch_response_keeps_delves_too_old_error_verbatim() {
+        // The UI's repair offer keys off this exact wording (`missingAdapter`), so the
+        // conversion to the IPC error must not fold it into a generic message.
+        let message = "Failed to launch /p/api: Version of Delve is too old for Go version go1.27.0 \
+                       (maximum supported version 1.26, suppress this error with --check-go-version=false)";
+        let e = ide_core::IdeError::from(DapError::Failed { command: "launch".into(), message: message.into() });
+        assert_eq!(e.kind, ide_core::error::ErrorKind::InvalidInput);
+        assert_eq!(e.message, format!("launch failed: {message}"));
+        assert!(e.message.contains("Version of Delve is too old for Go version go1.27.0"));
+    }
 }
@@ -12,6 +12,35 @@ use tokio::net::{TcpListener, TcpStream};
 /// How long the relay waits for java-debug to attach once the JVM is connected.
 const ATTACH_TIMEOUT: Duration = Duration::from_secs(60);

+/// Loopback ports relaying a connected JVM, while their relay runs.
+static RELAY_PORTS: std::sync::Mutex<Vec<u16>> = std::sync::Mutex::new(Vec::new());
+
+fn relay_ports() -> std::sync::MutexGuard<'static, Vec<u16>> {
+    RELAY_PORTS.lock().unwrap_or_else(std::sync::PoisonError::into_inner)
+}
+
+/// Whether `port` is a relay of [`listen_for_jvm`]: a session attaching to it debugs a JVM that
+/// connected to the IDE, so the JVM has no JDWP listener the memory view could find it by.
+pub(crate) fn is_listen_relay(port: u16) -> bool {
+    relay_ports().contains(&port)
+}
+
+/// Registers a relay port for as long as it lives.
+struct RelayPort(u16);
+
+impl RelayPort {
+    fn new(port: u16) -> Self {
+        relay_ports().push(port);
+        Self(port)
+    }
+}
+
+impl Drop for RelayPort {
+    fn drop(&mut self) {
+        relay_ports().retain(|&p| p != self.0);
+    }
+}
+
 /// The address to listen on for `host`: the interface it names, or every interface when
 /// it is not one of this machine's (a DNS name or a NAT address the JVM connects to).
 async fn bind(host: &str, port: u16) -> IdeResult<TcpListener> {
@@ -40,11 +69,11 @@ pub async fn listen_for_jvm(host: &str, port: u16, cancel: CancellationToken) ->
    let _ = jvm.set_nodelay(true);
    let relay = TcpListener::bind(("127.0.0.1", 0)).await?;
    let local = relay.local_addr()?.port();
-    tokio::spawn(relay_one(relay, jvm));
+    tokio::spawn(relay_one(relay, jvm, RelayPort::new(local)));
    Ok(local)
 }

-async fn relay_one(relay: TcpListener, mut jvm: TcpStream) {
+async fn relay_one(relay: TcpListener, mut jvm: TcpStream, _registered: RelayPort) {
    let debugger = match tokio::time::timeout(ATTACH_TIMEOUT, relay.accept()).await {
        Ok(Ok((s, _))) => s,
        _ => return,
@@ -58,22 +87,64 @@ async fn relay_one(relay: TcpListener, mut jvm: TcpStream) {
 }

 #[cfg(test)]
-mod tests {
+pub(crate) mod tests {
    use super::*;
    use tokio::io::{AsyncReadExt, AsyncWriteExt};

-    #[tokio::test]
-    async fn relays_the_jdwp_handshake_both_ways() {
-        let port = TcpListener::bind(("127.0.0.1", 0)).await.unwrap().local_addr().unwrap().port();
-        let waiting = tokio::spawn(listen_for_jvm("127.0.0.1", port, CancellationToken::new()));
-        // The remote JVM (`server=n`) connects to the IDE.
-        let mut jvm = loop {
-            if let Ok(s) = TcpStream::connect(("127.0.0.1", port)).await {
-                break s;
+    /// A remote JVM (`server=n`) connected to [`listen_for_jvm`], and the relay port. A port
+    /// picked by binding to 0 can be taken by a parallel test before `listen_for_jvm` binds it,
+    /// so a failed listen is retried on a fresh one.
+    pub(crate) async fn connected_jvm() -> (TcpStream, u16) {
+        for _ in 0..20 {
+            let port = TcpListener::bind(("127.0.0.1", 0)).await.unwrap().local_addr().unwrap().port();
+            if let Some(connected) = jvm_on(port).await {
+                return connected;
+            }
+        }
+        panic!("listen_for_jvm never accepted a JVM");
+    }
+
+    /// One attempt of [`connected_jvm`]; `None` when the port turned out to be taken.
+    async fn jvm_on(port: u16) -> Option<(TcpStream, u16)> {
+        let mut waiting = tokio::spawn(listen_for_jvm("127.0.0.1", port, CancellationToken::new()));
+        let mut jvm = None;
+        for _ in 0..100 {
+            // A listen that failed to bind leaves nothing to connect to.
+            if waiting.is_finished() {
+                break;
+            }
+            // macOS does not refuse a connect to a port a socket holds without listening: the
+            // SYN goes unanswered until the connect times out.
+            let connect = TcpStream::connect(("127.0.0.1", port));
+            if let Ok(Ok(s)) = tokio::time::timeout(Duration::from_secs(1), connect).await {
+                jvm = Some(s);
+                break;
            }
            tokio::time::sleep(Duration::from_millis(20)).await;
-        };
-        let local = waiting.await.unwrap().unwrap();
+        }
+        // A connection to another test's listener never reaches `listen_for_jvm`.
+        let relay = tokio::time::timeout(Duration::from_secs(2), &mut waiting).await;
+        if let (Some(jvm), Ok(Ok(Ok(relay)))) = (jvm, relay) {
+            return Some((jvm, relay));
+        }
+        waiting.abort();
+        None
+    }
+
+    #[tokio::test]
+    async fn a_port_another_socket_holds_is_given_up() {
+        // Bound but not listening, as a parallel test's outgoing connection holds its local port.
+        let held = tokio::net::TcpSocket::new_v4().unwrap();
+        held.bind(([127, 0, 0, 1], 0).into()).unwrap();
+        let port = held.local_addr().unwrap().port();
+        let attempt = tokio::time::timeout(Duration::from_secs(5), jvm_on(port)).await;
+        assert!(matches!(attempt, Ok(None)), "an attempt on a taken port gives up instead of waiting forever");
+    }
+
+    #[tokio::test]
+    async fn relays_the_jdwp_handshake_both_ways() {
+        // The remote JVM (`server=n`) connects to the IDE.
+        let (mut jvm, local) = connected_jvm().await;
        // java-debug attaches to the relay and starts the handshake.
        let mut debugger = TcpStream::connect(("127.0.0.1", local)).await.unwrap();
        debugger.write_all(b"JDWP-Handshake").await.unwrap();
@@ -85,6 +156,22 @@ mod tests {
        assert_eq!(&buf, b"JDWP-Handshake");
    }

+    #[tokio::test]
+    async fn relay_port_is_known_while_the_relay_runs() {
+        let (jvm, local) = connected_jvm().await;
+        assert!(is_listen_relay(local));
+
+        let debugger = TcpStream::connect(("127.0.0.1", local)).await.unwrap();
+        drop((jvm, debugger));
+        for _ in 0..200 {
+            if !is_listen_relay(local) {
+                return;
+            }
+            tokio::time::sleep(Duration::from_millis(10)).await;
+        }
+        panic!("relay port {local} still registered after the relay ended");
+    }
+
    #[tokio::test]
    async fn waiting_is_cancellable() {
        let cancel = CancellationToken::new();
@@ -0,0 +1,92 @@
+//! The class table counted by the debugger (JDWP `InstanceCounts`) where no histogram can be
+//! taken: JDK 8-16 while suspended, Windows, a remote JVM. Only while suspended, and slow on a
+//! large heap (seconds on JDK 17+).
+
+use super::inspect::Call;
+use super::snapshot::diff_rows;
+use super::{ClassCounts, JvmMemoryResult, JvmMemorySnapshot, JvmMemorySource, JvmMemoryState, JvmMemoryUnavailable};
+use crate::session::Session;
+use ide_core::IdeResult;
+use ide_jvm_attach::HistogramRow;
+use ide_jvm_inspect::ClassCount;
+
+/// Reasons the debugger may be able to work around; the others (the program exited, the JVM
+/// did not answer) would fail for it too.
+pub(super) fn falls_back(reason: &JvmMemoryUnavailable) -> bool {
+    use JvmMemoryUnavailable as U;
+    matches!(
+        reason,
+        U::UnsafeWhileSuspended { .. }
+            | U::UnsupportedPlatform
+            | U::AttachedJvmNotFound
+            | U::ListenerNotRunning
+            | U::ListenModeNotSupported
+            | U::NoJvmProcess
+    )
+}
+
+/// `histogram` unless it is unavailable for a reason `session`'s inspector can work around.
+pub(super) async fn or_debugger_counts(
+    state: &mut JvmMemoryState,
+    histogram: Result<JvmMemorySnapshot, JvmMemoryUnavailable>,
+    session: Option<&Session>,
+) -> IdeResult<JvmMemoryResult> {
+    let reason = match histogram {
+        Ok(snapshot) => return Ok(JvmMemoryResult::Ready { snapshot }),
+        Err(reason) => reason,
+    };
+    let unavailable = |can_count_when_paused| Ok(JvmMemoryResult::Unavailable { reason, can_count_when_paused });
+    let Some(session) = session.filter(|_| falls_back(&reason)) else {
+        return unavailable(false);
+    };
+    let call = match Call::begin(session, None).await {
+        Ok(call) => call,
+        Err(super::JvmInspectUnavailable::NotSuspended) => return unavailable(true),
+        Err(_) => return unavailable(false),
+    };
+    let started = std::time::Instant::now();
+    let result = call.run(call.inspector.class_counts(&call.target)).await;
+    match call.finish(result) {
+        Ok(counts) => {
+            let elapsed_ms = u32::try_from(started.elapsed().as_millis()).unwrap_or(u32::MAX);
+            let previous = state.baseline(JvmMemorySource::Debugger);
+            let (snapshot, counts) = debugger_snapshot(counts, previous.as_deref(), elapsed_ms);
+            state.remember(JvmMemorySource::Debugger, counts);
+            Ok(JvmMemoryResult::Ready { snapshot })
+        }
+        Err(e) => {
+            tracing::debug!("debugger class counts: {e}");
+            let running = !session.suspended.get();
+            unavailable(running && session.jvm_inspector.lock().is_some())
+        }
+    }
+}
+
+/// Classes without instances are left out; most instances first.
+pub(super) fn debugger_snapshot(
+    counts: Vec<ClassCount>,
+    previous: Option<&ClassCounts>,
+    elapsed_ms: u32,
+) -> (JvmMemorySnapshot, ClassCounts) {
+    let rows = counts
+        .into_iter()
+        .filter(|c| c.instances > 0)
+        .map(|c| HistogramRow { name: c.name, module: None, instances: c.instances, bytes: 0 })
+        .collect();
+    let (mut rows, counts) = diff_rows(rows, previous);
+    for row in &mut rows {
+        row.delta_bytes = None;
+    }
+    rows.sort_by(|a, b| b.instances.cmp(&a.instances).then_with(|| a.name.cmp(&b.name)));
+    let snapshot = JvmMemorySnapshot {
+        total_instances: rows.iter().map(|r| r.instances).sum(),
+        total_bytes: 0,
+        rows,
+        java_version: String::new(),
+        taken_while_suspended: true,
+        has_baseline: previous.is_some(),
+        elapsed_ms,
+        source: JvmMemorySource::Debugger,
+    };
+    (snapshot, counts)
+}
@@ -0,0 +1,293 @@
+//! A heap dump of a debugged JVM, written over its attach listener into a snapshot directory.
+
+use super::snapshot::{AttachProbe, JdkProbe, attach_failure, gone, pick_jvm_pid};
+use super::{
+    JvmMemoryState, JvmMemoryUnavailable, JvmTarget, histogram_pid, keep_found, lookup, resolve_jvm,
+    safe_while_suspended, session_target,
+};
+use crate::session::Session;
+use ide_core::{CancellationToken, IdeError, IdeResult};
+use ide_jvm_attach::{AttachError, HeapDumpReply};
+use serde::{Deserialize, Serialize};
+use specta::Type;
+use std::future::Future;
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+
+const DUMP_FILE: &str = "dump.hprof";
+/// For a refusal whose reply holds no reason: empty, or only the line announcing the dump.
+const NO_REASON: &str = "the JVM wrote no heap dump and gave no reason";
+
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, Type)]
+#[serde(tag = "kind", rename_all = "camelCase", rename_all_fields = "camelCase")]
+pub enum JvmHeapUnavailable {
+    /// The reason a class table could not be read applies to the dump too.
+    Memory {
+        reason: JvmMemoryUnavailable,
+    },
+    /// Bytes.
+    NotEnoughDiskSpace {
+        needed: u64,
+        available: u64,
+    },
+    DumpFailed {
+        message: String,
+    },
+    DumpTimedOut {
+        seconds: u32,
+    },
+}
+
+#[derive(Clone, Debug, PartialEq)]
+pub struct JvmHeapDump {
+    pub path: PathBuf,
+    /// The class histogram taken just before the dump.
+    pub histogram: Vec<ide_heap::HistogramEntry>,
+    /// Raw `VM.flags`; empty when the JVM did not answer.
+    pub vm_flags: String,
+    /// 0 when unknown.
+    pub java_major: u32,
+    /// Empty when unknown.
+    pub java_version: String,
+    pub pid: u32,
+    pub taken_while_suspended: bool,
+    pub dump_ms: u32,
+}
+
+#[derive(Clone, Debug, PartialEq)]
+pub enum JvmHeapOutcome {
+    Ready(JvmHeapDump),
+    /// The snapshot directory has been deleted.
+    Unavailable(JvmHeapUnavailable),
+    /// The snapshot directory has been deleted.
+    Cancelled,
+}
+
+pub(crate) async fn session_heap_dump(
+    session: Arc<Session>,
+    dir: PathBuf,
+    cancel: CancellationToken,
+) -> IdeResult<JvmHeapOutcome> {
+    let launched = histogram_pid(&session);
+    let capture = {
+        let (dir, cancel) = (dir.clone(), cancel.clone());
+        async move {
+            let target = session_target(&session, launched);
+            let probe = |suspended| -> Arc<dyn AttachProbe> { Arc::new(JdkProbe { suspended }) };
+            let stopped = session.clone();
+            let suspended = move || capture_suspended(&stopped);
+            heap_dump_with(&session.jvm_memory, suspended, probe, target, dir, cancel).await
+        }
+    };
+    detached(capture, dir, cancel).await
+}
+
+/// The JVM can be suspended before the adapter reports the stop, so a session with a JDWP link
+/// also counts the link's view.
+pub(super) fn capture_suspended(session: &Session) -> bool {
+    session.suspended.get()
+        || session
+            .jdwp_link
+            .lock()
+            .as_ref()
+            .is_some_and(|link| matches!(link.state(), ide_jdwp::LinkState::Suspended { .. }))
+}
+
+/// Runs `capture` in a task of its own, so a caller that stops waiting cannot release the
+/// session's lock while the JVM is still writing the dump and let a class-table request run
+/// next to it. On `cancel` returns at once and deletes `dir`.
+pub(super) async fn detached(
+    capture: impl Future<Output = IdeResult<JvmHeapOutcome>> + Send + 'static,
+    dir: PathBuf,
+    cancel: CancellationToken,
+) -> IdeResult<JvmHeapOutcome> {
+    let task = tokio::spawn(capture);
+    tokio::select! {
+        joined = task => joined.map_err(|e| IdeError::internal(format!("JVM heap dump failed: {e}")))?,
+        () = cancel.cancelled() => {
+            discard_later(dir).await;
+            Ok(JvmHeapOutcome::Cancelled)
+        }
+    }
+}
+
+/// Holds `state`'s lock throughout, as a class-table request does; `suspended` and then `probe`
+/// are read once it is held, and `suspended` again by [`take_heap_dump`]. Anything but `Ready`
+/// deletes `dir`.
+pub(super) async fn heap_dump_with(
+    state: &tokio::sync::Mutex<JvmMemoryState>,
+    suspended: impl Fn() -> bool + Send + Sync + 'static,
+    probe: impl FnOnce(bool) -> Arc<dyn AttachProbe>,
+    target: JvmTarget<'_>,
+    dir: PathBuf,
+    cancel: CancellationToken,
+) -> IdeResult<JvmHeapOutcome> {
+    // The capture's histogram never becomes the Memory view's baseline, so the state is unused.
+    let _turn = state.lock().await;
+    let outcome = capture(suspended, probe, target, dir.clone(), cancel).await;
+    if !matches!(outcome, Ok(JvmHeapOutcome::Ready(_))) {
+        discard_later(dir).await;
+    }
+    outcome
+}
+
+async fn capture(
+    suspended_now: impl Fn() -> bool + Send + Sync + 'static,
+    probe: impl FnOnce(bool) -> Arc<dyn AttachProbe>,
+    target: JvmTarget<'_>,
+    dir: PathBuf,
+    cancel: CancellationToken,
+) -> IdeResult<JvmHeapOutcome> {
+    if cancel.is_cancelled() {
+        return Ok(JvmHeapOutcome::Cancelled);
+    }
+    let suspended = suspended_now();
+    let lookup = match lookup(target) {
+        Ok(lookup) => lookup,
+        Err(reason) => return Ok(memory(reason)),
+    };
+    let probe = probe(suspended);
+    let (found, outcome) = tokio::task::spawn_blocking(move || match resolve_jvm(lookup, &*probe, suspended) {
+        Ok((tree, found)) => (found, take_heap_dump(&*probe, &tree, suspended, &suspended_now, &dir, &cancel)),
+        Err(reason) => (None, Ok(memory(reason))),
+    })
+    .await
+    .map_err(|e| IdeError::internal(format!("JVM heap dump failed: {e}")))?;
+    keep_found(target, found);
+    outcome
+}
+
+/// `tree` starts with the session's reported pid, followed by its descendants. `suspended` is
+/// the stop read when the capture began; `suspended_now` reads it again. Anything but `Ready`
+/// deletes `dir`.
+pub(crate) fn take_heap_dump(
+    probe: &dyn AttachProbe,
+    tree: &[i32],
+    suspended: bool,
+    suspended_now: &dyn Fn() -> bool,
+    dir: &Path,
+    cancel: &CancellationToken,
+) -> IdeResult<JvmHeapOutcome> {
+    let outcome = dump(probe, tree, suspended, suspended_now, dir, cancel);
+    if !matches!(outcome, Ok(JvmHeapOutcome::Ready(_))) {
+        discard(dir);
+    }
+    outcome
+}
+
+fn dump(
+    probe: &dyn AttachProbe,
+    tree: &[i32],
+    suspended: bool,
+    suspended_now: &dyn Fn() -> bool,
+    dir: &Path,
+    cancel: &CancellationToken,
+) -> IdeResult<JvmHeapOutcome> {
+    use JvmHeapUnavailable as H;
+    use JvmMemoryUnavailable as U;
+    let Some(root) = tree.first().and_then(|&p| u32::try_from(p).ok()) else {
+        return Ok(memory(U::NoJvmProcess));
+    };
+    let Some(jvm) = pick_jvm_pid(tree, |p| probe.has_listener(p)) else {
+        return Ok(memory(gone(probe, root)));
+    };
+
+    let version = probe.version(jvm);
+    let java_major = version.as_ref().map_or(0, |v| v.major);
+    let safe = safe_while_suspended(version.as_ref());
+    let unsafe_stop = || Ok(memory(U::UnsafeWhileSuspended { java_major }));
+    if suspended && !safe {
+        return unsafe_stop();
+    }
+    // The program may stop during the histogram, and a stopped JDK below 17 hangs on the next
+    // request until it resumes.
+    let stopped_since = || !safe && suspended_now();
+
+    let text = match probe.histogram(jvm) {
+        Ok(text) => text,
+        Err(e) => return attach_failure(probe, jvm, e, "class histogram").map(memory),
+    };
+    let histogram: Vec<ide_heap::HistogramEntry> = ide_jvm_attach::parse_class_histogram(&text)
+        .rows
+        .into_iter()
+        .map(|r| ide_heap::HistogramEntry { name: r.name, instances: r.instances, bytes: r.bytes })
+        .collect();
+    let (instances, bytes) =
+        histogram.iter().fold((0u64, 0u64), |(i, b), e| (i.saturating_add(e.instances), b.saturating_add(e.bytes)));
+    let needed = ide_heap::capture_need(bytes, instances);
+    if let Some(available) = probe.free_bytes(dir)
+        && available < needed
+    {
+        return Ok(JvmHeapOutcome::Unavailable(H::NotEnoughDiskSpace { needed, available }));
+    }
+
+    if stopped_since() {
+        return unsafe_stop();
+    }
+    // Without flags the index assumes the default object layout.
+    let vm_flags = probe.vm_flags(jvm).unwrap_or_else(|e| {
+        tracing::debug!("VM.flags of JVM {jvm}: {e}");
+        String::new()
+    });
+    if cancel.is_cancelled() {
+        return Ok(JvmHeapOutcome::Cancelled);
+    }
+    if stopped_since() {
+        return unsafe_stop();
+    }
+
+    let path = dir.join(DUMP_FILE);
+    let timeout = ide_jvm_attach::dump_timeout(bytes);
+    let started = std::time::Instant::now();
+    let reply = probe.heap_dump(jvm, &path, timeout);
+    let dump_ms = u32::try_from(started.elapsed().as_millis()).unwrap_or(u32::MAX);
+    if cancel.is_cancelled() {
+        return Ok(JvmHeapOutcome::Cancelled);
+    }
+    let failed = |message: String| Ok(JvmHeapOutcome::Unavailable(H::DumpFailed { message }));
+    match reply {
+        Ok(HeapDumpReply::Created { .. }) => {
+            if !std::fs::metadata(&path).is_ok_and(|m| m.is_file() && m.len() > 0) {
+                return failed(format!("the JVM reported a heap dump, but {} is missing or empty", path.display()));
+            }
+            Ok(JvmHeapOutcome::Ready(JvmHeapDump {
+                path,
+                histogram,
+                vm_flags,
+                java_major,
+                java_version: version.map(|v| v.text).unwrap_or_default(),
+                pid: jvm,
+                taken_while_suspended: suspended,
+                dump_ms,
+            }))
+        }
+        Ok(HeapDumpReply::Refused { message }) if message.is_empty() || message.starts_with("Dumping heap to") => {
+            failed(NO_REASON.into())
+        }
+        Ok(HeapDumpReply::Refused { message }) => failed(message),
+        Err(AttachError::Timeout) => {
+            let seconds = u32::try_from(timeout.as_secs()).unwrap_or(u32::MAX);
+            Ok(JvmHeapOutcome::Unavailable(H::DumpTimedOut { seconds }))
+        }
+        // A running JDK rejects an argument with a non-zero result code.
+        Err(AttachError::Command { message, .. }) => failed(message),
+        Err(e) => attach_failure(probe, jvm, e, "heap dump").map(memory),
+    }
+}
+
+fn memory(reason: JvmMemoryUnavailable) -> JvmHeapOutcome {
+    JvmHeapOutcome::Unavailable(JvmHeapUnavailable::Memory { reason })
+}
+
+fn discard(dir: &Path) {
+    if let Err(e) = std::fs::remove_dir_all(dir)
+        && e.kind() != std::io::ErrorKind::NotFound
+    {
+        tracing::warn!("could not delete heap snapshot directory {}: {e}", dir.display());
+    }
+}
+
+/// A dump may be large, so it is deleted off the async threads.
+async fn discard_later(dir: PathBuf) {
+    let _ = tokio::task::spawn_blocking(move || discard(&dir)).await;
+}
@@ -0,0 +1,251 @@
+//! A suspended JVM's instances, fields and referrers, read through the session's
+//! [`JvmInspector`]. An object is only valid within the stop it was read in.
+
+use super::launch_pid;
+use crate::session::Session;
+use ide_core::event::{DebugEvent, JvmBackendState};
+use ide_core::{IdeError, IdeResult};
+use ide_jvm_inspect::{
+    DEFAULT_REFERRERS, InspectError, InspectFuture, JvmField, JvmInspector, JvmReferrer, JvmValue, MAX_INSTANCES,
+    MAX_REFERRERS, ObjectId, VmTarget,
+};
+use serde::{Deserialize, Serialize};
+use specta::Type;
+use std::sync::Arc;
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize, Type)]
+#[serde(rename_all = "camelCase")]
+pub struct JvmObjectRef {
+    /// The `generation` of the [`JvmInstancesResult`] the object came from.
+    pub generation: u64,
+    pub id: ObjectId,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, Type)]
+#[serde(tag = "kind", rename_all = "camelCase", rename_all_fields = "camelCase")]
+pub enum JvmInspectUnavailable {
+    /// The session has no inspector: not a JVM session, not ready yet, or turned off.
+    NoBackend,
+    /// The debug adapter lacks what the inspector needs; turned off for the session.
+    Unsupported {
+        detail: String,
+    },
+    NotSuspended,
+    /// The object is from an earlier stop, or the program stopped or resumed meanwhile.
+    Stale,
+    Collected,
+    ClassNotLoaded,
+    VmNotFound,
+    /// Transient: ask again.
+    TimedOut,
+    BackendClosed,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, Type)]
+#[serde(tag = "status", rename_all = "camelCase", rename_all_fields = "camelCase")]
+pub enum JvmInstancesResult {
+    Ready {
+        /// Goes into the [`JvmObjectRef`] of each instance.
+        generation: u64,
+        instances: Vec<JvmValue>,
+        /// The limit was reached; there may be more.
+        truncated: bool,
+    },
+    Unavailable {
+        reason: JvmInspectUnavailable,
+    },
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, Type)]
+#[serde(tag = "status", rename_all = "camelCase", rename_all_fields = "camelCase")]
+pub enum JvmFieldsResult {
+    Ready { fields: Vec<JvmField> },
+    Unavailable { reason: JvmInspectUnavailable },
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, Type)]
+#[serde(tag = "status", rename_all = "camelCase", rename_all_fields = "camelCase")]
+pub enum JvmReferrersResult {
+    Ready { referrers: Vec<JvmReferrer>, truncated: bool },
+    Unavailable { reason: JvmInspectUnavailable },
+}
+
+/// The session's JVM for the inspector: java-debug's VM is found by the launched program's pid
+/// or by the port of an attach (the thread lists of two JVMs can be identical).
+pub(crate) fn vm_target(s: &Session) -> VmTarget {
+    let (pid, attach) = if s.request == "launch" {
+        (u32::try_from(launch_pid(s)).ok().filter(|&p| p > 0), None)
+    } else {
+        (None, s.attach_target.get())
+    };
+    VmTarget {
+        pid,
+        attach_host: attach.map(|(host, _)| host.clone()).filter(|h| !h.is_empty()),
+        attach_port: attach.map(|&(_, port)| port),
+        generation: s.suspended.ticket(),
+    }
+}
+
+pub(crate) fn set_inspector(s: &Session, inspector: Arc<dyn JvmInspector>) {
+    *s.jvm_inspector.lock() = Some(inspector);
+    s.emit(DebugEvent::Capabilities);
+}
+
+pub(crate) async fn instances(s: &Session, class_name: &str, limit: u32) -> IdeResult<JvmInstancesResult> {
+    let limit = limit.clamp(1, MAX_INSTANCES);
+    let call = match Call::begin(s, None).await {
+        Ok(call) => call,
+        Err(reason) => return Ok(JvmInstancesResult::Unavailable { reason }),
+    };
+    let generation = call.target.generation;
+    let result = call.run(call.inspector.instances(&call.target, class_name, limit)).await;
+    Ok(match settle(call.finish(result))? {
+        Ok(instances) => {
+            JvmInstancesResult::Ready { generation, truncated: instances.len() >= limit as usize, instances }
+        }
+        Err(reason) => JvmInstancesResult::Unavailable { reason },
+    })
+}
+
+pub(crate) async fn fields(s: &Session, object: JvmObjectRef) -> IdeResult<JvmFieldsResult> {
+    let call = match Call::begin(s, Some(object.generation)).await {
+        Ok(call) => call,
+        Err(reason) => return Ok(JvmFieldsResult::Unavailable { reason }),
+    };
+    let result = call.run(call.inspector.fields(&call.target, object.id)).await;
+    Ok(match settle(call.finish(result))? {
+        Ok(fields) => JvmFieldsResult::Ready { fields },
+        Err(reason) => JvmFieldsResult::Unavailable { reason },
+    })
+}
+
+pub(crate) async fn referrers(s: &Session, object: JvmObjectRef, limit: u32) -> IdeResult<JvmReferrersResult> {
+    let limit = if limit == 0 { DEFAULT_REFERRERS } else { limit.min(MAX_REFERRERS) };
+    let call = match Call::begin(s, Some(object.generation)).await {
+        Ok(call) => call,
+        Err(reason) => return Ok(JvmReferrersResult::Unavailable { reason }),
+    };
+    let result = call.run(call.inspector.referrers(&call.target, object.id, limit)).await;
+    Ok(match settle(call.finish(result))? {
+        Ok(referrers) => JvmReferrersResult::Ready { truncated: referrers.len() >= limit as usize, referrers },
+        Err(reason) => JvmReferrersResult::Unavailable { reason },
+    })
+}
+
+/// One inspector request. Holds the session's inspect lock and a clone of its inspector, which
+/// [`Call::run`] lets go of when the session ends.
+pub(super) struct Call<'s> {
+    session: &'s Session,
+    pub(super) inspector: Arc<dyn JvmInspector>,
+    pub(super) target: VmTarget,
+    _turn: tokio::sync::MutexGuard<'s, ()>,
+}
+
+impl<'s> Call<'s> {
+    /// `object_generation`: the stop the request's object was read in.
+    pub(super) async fn begin(
+        s: &'s Session,
+        object_generation: Option<u64>,
+    ) -> Result<Call<'s>, JvmInspectUnavailable> {
+        // Checked under the lock: a request queued behind another must see the state it runs in.
+        let turn = s.inspect_lock.lock().await;
+        let Some(inspector) = s.jvm_inspector.lock().clone() else {
+            return Err(JvmInspectUnavailable::NoBackend);
+        };
+        if !s.suspended.get() {
+            return Err(JvmInspectUnavailable::NotSuspended);
+        }
+        let target = vm_target(s);
+        if object_generation.is_some_and(|g| g != target.generation) {
+            return Err(JvmInspectUnavailable::Stale);
+        }
+        Ok(Call { session: s, inspector, target, _turn: turn })
+    }
+
+    pub(super) async fn run<T>(&self, request: InspectFuture<'_, T>) -> Result<T, InspectError> {
+        tokio::select! {
+            result = request => result,
+            () = self.session.ended.cancelled() => Err(InspectError::Closed),
+        }
+    }
+
+    /// The result unless the program stopped or resumed while it was read: a request of several
+    /// commands can straddle a stop. `Unsupported` turns the backend off.
+    pub(super) fn finish<T>(self, result: Result<T, InspectError>) -> Result<T, InspectError> {
+        let s = self.session;
+        match result {
+            Err(InspectError::Unsupported(detail)) => {
+                turn_off(s, &self.inspector, &detail);
+                Err(InspectError::Unsupported(detail))
+            }
+            _ if !s.suspended.get() || s.suspended.ticket() != self.target.generation => Err(InspectError::Stale),
+            result => result,
+        }
+    }
+}
+
+/// An adapter that lacks what the inspector needs will not gain it during the session.
+fn turn_off(s: &Session, inspector: &Arc<dyn JvmInspector>, detail: &str) {
+    let removed = {
+        let mut slot = s.jvm_inspector.lock();
+        // Not one installed since this request began.
+        let current = slot.as_ref().is_some_and(|i| std::ptr::addr_eq(Arc::as_ptr(i), Arc::as_ptr(inspector)));
+        current && slot.take().is_some()
+    };
+    if removed {
+        s.emit(DebugEvent::JvmBackend { state: JvmBackendState::Unsupported { detail: detail.to_owned() } });
+        s.emit(DebugEvent::Capabilities);
+    }
+}
+
+/// A failure the inspector could not classify is an error, not a reason shown to the user.
+fn settle<T>(result: Result<T, InspectError>) -> IdeResult<Result<T, JvmInspectUnavailable>> {
+    use JvmInspectUnavailable as U;
+    let error = match result {
+        Ok(value) => return Ok(Ok(value)),
+        Err(error) => error,
+    };
+    Ok(Err(match error {
+        InspectError::NotSuspended => U::NotSuspended,
+        InspectError::Stale => U::Stale,
+        InspectError::Collected => U::Collected,
+        InspectError::ClassNotLoaded => U::ClassNotLoaded,
+        InspectError::VmNotFound => U::VmNotFound,
+        InspectError::Unsupported(detail) => U::Unsupported { detail },
+        // Also the Kotlin link's `Busy`: the backend stays.
+        InspectError::TimedOut => U::TimedOut,
+        InspectError::Closed => U::BackendClosed,
+        InspectError::Failed(message) => return Err(IdeError::internal(format!("JVM inspector: {message}"))),
+    }))
+}
+
+/// Reports a Kotlin link that stops inspecting or closes. Holds neither the link nor the
+/// session: it ends once the session drops the link, and never drops it itself (after
+/// `VirtualMachine.Dispose` the link is closed but still forwarding the reply).
+pub(crate) fn watch_link(s: &Arc<Session>, link: &ide_jdwp::JdwpLink) {
+    let mut states = link.subscribe();
+    let session = Arc::downgrade(s);
+    tokio::spawn(async move {
+        let mut reported = None;
+        loop {
+            let state = match &*states.borrow_and_update() {
+                ide_jdwp::LinkState::Degraded(detail) => Some(JvmBackendState::Degraded { detail: detail.clone() }),
+                ide_jdwp::LinkState::Closed => Some(JvmBackendState::Closed),
+                _ => None,
+            };
+            if let Some(state) = state
+                && reported.as_ref() != Some(&state)
+            {
+                let Some(s) = session.upgrade() else { return };
+                if s.ended.is_cancelled() || s.terminated.load(std::sync::atomic::Ordering::Acquire) {
+                    return;
+                }
+                s.emit(DebugEvent::JvmBackend { state: state.clone() });
+                reported = Some(state);
+            }
+            if states.changed().await.is_err() {
+                return;
+            }
+        }
+    });
+}
@@ -0,0 +1,338 @@
+//! Class-table snapshots of a debugged JVM, read over the HotSpot attach listener or, where
+//! that is unavailable, counted by the debugger; and the JVM's objects through the debugger.
+
+mod fallback;
+mod heap;
+mod inspect;
+mod scan;
+mod snapshot;
+#[cfg(test)]
+pub(crate) mod test_support;
+#[cfg(test)]
+mod tests;
+
+use crate::session::Session;
+use crate::session::Suspension;
+use ide_core::IdeResult;
+use serde::{Deserialize, Serialize};
+use specta::Type;
+use std::collections::HashMap;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicI64, Ordering};
+use std::time::Duration;
+
+pub use heap::*;
+pub use inspect::{JvmFieldsResult, JvmInspectUnavailable, JvmInstancesResult, JvmObjectRef, JvmReferrersResult};
+pub(crate) use inspect::{fields, instances, referrers, set_inspector, watch_link};
+#[cfg(all(test, unix))]
+pub(crate) use scan::watch_for_attached_jvm;
+pub(crate) use scan::{JvmFinder, attach_target, find_attached_jvm, find_local_jvm, local_jdwp_port};
+use snapshot::{AttachProbe, JdkProbe, take_snapshot};
+
+pub const JVM_MEMORY_ADAPTERS: &[&str] = &["java-debug", crate::kotlin::ADAPTER];
+pub const HISTOGRAM_TIMEOUT: Duration = Duration::from_secs(10);
+
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, Type)]
+#[serde(rename_all = "camelCase")]
+pub struct JvmClassRow {
+    pub name: String,
+    pub module: Option<String>,
+    pub instances: u64,
+    pub bytes: u64,
+    /// Change since the previous snapshot; `None` on the first one and for ambiguous rows.
+    pub delta_instances: Option<i64>,
+    pub delta_bytes: Option<i64>,
+    /// Several classes share this name (different class loaders), so it has no delta.
+    pub ambiguous: bool,
+}
+
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, Type)]
+#[serde(rename_all = "camelCase")]
+pub struct JvmMemorySnapshot {
+    /// Largest `bytes` first.
+    pub rows: Vec<JvmClassRow>,
+    pub total_instances: u64,
+    pub total_bytes: u64,
+    /// Empty when unknown.
+    pub java_version: String,
+    pub taken_while_suspended: bool,
+    /// A previous snapshot from the same source exists; deltas are against it.
+    pub has_baseline: bool,
+    /// Wall time of the histogram or count request.
+    pub elapsed_ms: u32,
+    pub source: JvmMemorySource,
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize, Type)]
+#[serde(rename_all = "camelCase")]
+pub enum JvmMemorySource {
+    /// `GC.class_histogram -all`: sizes, and objects not yet collected.
+    Histogram,
+    /// JDWP `InstanceCounts`: reachable objects only, no sizes (`bytes` is 0).
+    Debugger,
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Serialize, Deserialize, Type)]
+#[serde(tag = "kind", rename_all = "camelCase", rename_all_fields = "camelCase")]
+pub enum JvmMemoryUnavailable {
+    /// An attach session whose JVM is remote, or local but not found by its JDWP port.
+    AttachedJvmNotFound,
+    /// An attach through the IDE's listen-mode relay: the JVM connected to the IDE, so it has
+    /// no JDWP listener to be found by.
+    ListenModeNotSupported,
+    UnsupportedPlatform,
+    NoJvmProcess,
+    ListenerNotRunning,
+    /// `java_major` is 0 when the version is unknown.
+    UnsafeWhileSuspended {
+        java_major: u32,
+    },
+    ProgramExited,
+    TimedOut,
+}
+
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, Type)]
+#[serde(tag = "status", rename_all = "camelCase", rename_all_fields = "camelCase")]
+pub enum JvmMemoryResult {
+    Ready {
+        snapshot: JvmMemorySnapshot,
+    },
+    Unavailable {
+        reason: JvmMemoryUnavailable,
+        /// The debugger could count the classes if the program were paused.
+        can_count_when_paused: bool,
+    },
+}
+
+/// Instances and bytes per class name, for the next snapshot's deltas; `None` for a name
+/// several classes shared, so it gets no delta next time either.
+type ClassCounts = HashMap<String, Option<(u64, u64)>>;
+
+/// Per-session baseline; its mutex also keeps one class-table request per session at a time.
+#[derive(Default)]
+pub struct JvmMemoryState {
+    /// Shared with the blocking task, so a task that panics leaves it in place.
+    previous: Option<(JvmMemorySource, Arc<ClassCounts>)>,
+}
+
+impl JvmMemoryState {
+    /// The previous snapshot's counts if it came from `source`: counts from the other source
+    /// are not comparable (only reachable objects versus all of them).
+    pub(crate) fn baseline(&self, source: JvmMemorySource) -> Option<Arc<ClassCounts>> {
+        self.previous.as_ref().filter(|(s, _)| *s == source).map(|(_, counts)| counts.clone())
+    }
+
+    pub(crate) fn remember(&mut self, source: JvmMemorySource, counts: ClassCounts) {
+        self.previous = Some((source, Arc::new(counts)));
+    }
+}
+
+pub(crate) async fn session_classes(session: Arc<Session>) -> IdeResult<JvmMemoryResult> {
+    until_located(&session).await;
+    let launched = histogram_pid(&session);
+    // Detached, so a caller that gives up cannot release the lock while the JVM is still
+    // answering and let a second histogram request run next to it.
+    tokio::spawn(async move {
+        let target = session_target(&session, launched);
+        let probe = |suspended| -> Arc<dyn AttachProbe> { Arc::new(JdkProbe { suspended }) };
+        snapshot(&session.jvm_memory, &session.suspended, probe, target, Some(&session)).await
+    })
+    .await
+    .map_err(|e| ide_core::IdeError::internal(format!("JVM memory snapshot failed: {e}")))?
+}
+
+/// Bounds the wait of [`until_located`]; `configure` has its own timeouts, this only guards them.
+const LOCATE_WAIT: Duration = Duration::from_secs(30);
+
+/// Until `configure` has recorded where `session`'s JVM is. The Memory view asks as soon as the
+/// session appears; looked up earlier, an attached JVM is reported as missing.
+async fn until_located(session: &Session) {
+    let mut located = session.jvm_located.subscribe();
+    tokio::select! {
+        _ = located.wait_for(|v| *v) => {}
+        () = session.ended.cancelled() => {}
+        () = tokio::time::sleep(LOCATE_WAIT) => {}
+    }
+}
+
+/// `launched`: [`histogram_pid`] of `session`.
+fn session_target(session: &Session, launched: Result<Option<i32>, JvmMemoryUnavailable>) -> JvmTarget<'_> {
+    match launched {
+        Ok(Some(pid)) => JvmTarget::Launched(pid),
+        Ok(None) => JvmTarget::Attached { pid: &session.jvm_pid, jdwp_port: session.jdwp_port.get().copied() },
+        Err(reason) => JvmTarget::Unavailable(reason),
+    }
+}
+
+/// The pid of a launch session's program (`None` for an attach session), or why no histogram
+/// can be taken.
+fn histogram_pid(session: &Session) -> Result<Option<i32>, JvmMemoryUnavailable> {
+    if !JVM_MEMORY_ADAPTERS.contains(&session.adapter.as_str()) {
+        return Err(JvmMemoryUnavailable::NoJvmProcess);
+    }
+    if !cfg!(unix) {
+        return Err(JvmMemoryUnavailable::UnsupportedPlatform);
+    }
+    if session.request == "launch" {
+        return match i32::try_from(launch_pid(session)).ok().filter(|&p| p > 0) {
+            Some(pid) => Ok(Some(pid)),
+            None => Err(JvmMemoryUnavailable::NoJvmProcess),
+        };
+    }
+    if session.jvm_pid.load(Ordering::Acquire) <= 0 {
+        if session.terminated.load(Ordering::Acquire) {
+            return Err(JvmMemoryUnavailable::ProgramExited);
+        }
+        if session.listen_relay.load(Ordering::Acquire) {
+            return Err(JvmMemoryUnavailable::ListenModeNotSupported);
+        }
+    }
+    Ok(None)
+}
+
+/// The JVM a launch session started: the one the IDE started for an app-launched (Kotlin)
+/// session, else the program the adapter reported; 0 when neither is known.
+pub(crate) fn launch_pid(session: &Session) -> i64 {
+    match session.jvm_pid.load(Ordering::Acquire) {
+        p if p > 0 => p,
+        _ => session.debuggee_pid.load(Ordering::Acquire),
+    }
+}
+
+#[derive(Clone, Copy)]
+pub(crate) enum JvmTarget<'a> {
+    /// The program a launch session started: its pid and everything it started.
+    Launched(i32),
+    /// The JVM of an attach session. While `pid` is 0 it is looked up by `jdwp_port`, which is
+    /// `None` for a remote JVM.
+    Attached { pid: &'a AtomicI64, jdwp_port: Option<u16> },
+    /// Known without asking any JVM: no histogram can be taken, for this reason.
+    Unavailable(JvmMemoryUnavailable),
+}
+
+enum Lookup {
+    Tree(i32),
+    Known(i32),
+    Scan(u16),
+}
+
+/// `probe` is built from the suspension read once the lock is held: a request queued behind
+/// another must not decide on a state from before the program hit a breakpoint. Where no
+/// histogram can be taken, `debugger` (the session) may count the classes instead.
+pub(crate) async fn snapshot(
+    state: &tokio::sync::Mutex<JvmMemoryState>,
+    suspended: &Suspension,
+    probe: impl FnOnce(bool) -> Arc<dyn AttachProbe>,
+    target: JvmTarget<'_>,
+    debugger: Option<&Session>,
+) -> IdeResult<JvmMemoryResult> {
+    let mut state = state.lock().await;
+    let outcome = histogram(&mut state, suspended, probe, target).await?;
+    fallback::or_debugger_counts(&mut state, outcome, debugger).await
+}
+
+async fn histogram(
+    state: &mut JvmMemoryState,
+    suspended: &Suspension,
+    probe: impl FnOnce(bool) -> Arc<dyn AttachProbe>,
+    target: JvmTarget<'_>,
+) -> IdeResult<Result<JvmMemorySnapshot, JvmMemoryUnavailable>> {
+    let suspended = suspended.get();
+    let lookup = match lookup(target) {
+        Ok(lookup) => lookup,
+        Err(reason) => return Ok(Err(reason)),
+    };
+    let probe = probe(suspended);
+    let previous = state.baseline(JvmMemorySource::Histogram);
+    let (found, outcome) = tokio::task::spawn_blocking(move || match resolve_jvm(lookup, &*probe, suspended) {
+        Ok((tree, found)) => (found, take_snapshot(&*probe, &tree, suspended, previous.as_deref())),
+        Err(reason) => (None, Ok(Err(reason))),
+    })
+    .await
+    .map_err(|e| ide_core::IdeError::internal(format!("JVM memory snapshot failed: {e}")))?;
+    keep_found(target, found);
+    Ok(outcome?.map(|(snapshot, counts)| {
+        state.remember(JvmMemorySource::Histogram, counts);
+        snapshot
+    }))
+}
+
+/// Where to look for `target`'s JVM, or why there is none to ask.
+fn lookup(target: JvmTarget<'_>) -> Result<Lookup, JvmMemoryUnavailable> {
+    match target {
+        JvmTarget::Launched(pid) => Ok(Lookup::Tree(pid)),
+        JvmTarget::Attached { pid, jdwp_port } => match i32::try_from(pid.load(Ordering::Acquire)) {
+            Ok(p) if p > 0 => Ok(Lookup::Known(p)),
+            _ => jdwp_port.map(Lookup::Scan).ok_or(JvmMemoryUnavailable::AttachedJvmNotFound),
+        },
+        JvmTarget::Unavailable(reason) => Err(reason),
+    }
+}
+
+/// Blocking. The pids to pick the JVM from, and the pid a scan found (for [`keep_found`]).
+fn resolve_jvm(
+    lookup: Lookup,
+    probe: &dyn AttachProbe,
+    suspended: bool,
+) -> Result<(Vec<i32>, Option<i32>), JvmMemoryUnavailable> {
+    match lookup {
+        Lookup::Tree(pid) => Ok((pid_tree(pid), None)),
+        Lookup::Known(pid) => Ok((vec![pid], None)),
+        Lookup::Scan(port) => scan_for_jvm(probe, port, suspended).map(|pid| (vec![pid], Some(pid))),
+    }
+}
+
+fn keep_found(target: JvmTarget<'_>, found: Option<i32>) {
+    if let (Some(found), JvmTarget::Attached { pid, .. }) = (found, target) {
+        // The attach-time scan may have found it meanwhile; either answer is the same JVM.
+        let _ = pid.compare_exchange(0, i64::from(found), Ordering::AcqRel, Ordering::Acquire);
+    }
+}
+
+/// A suspended JDK below 17 hangs on an attach request, so while suspended only a JVM known to
+/// be 17 or later is asked anything. Pass a version read without attaching (the `release` file).
+fn safe_while_suspended(version: Option<&ide_jvm_attach::JavaVersion>) -> bool {
+    version.is_some_and(|v| ide_jvm_attach::histogram_safe_while_suspended(v.major))
+}
+
+/// The JVM listening on JDWP `port`. While suspended the scan reaches the session's own JVM
+/// too, so JVMs unsafe to ask are skipped, and when one of them may be the session's the
+/// answer is a refusal rather than "not found".
+fn scan_for_jvm(probe: &dyn AttachProbe, port: u16, suspended: bool) -> Result<i32, JvmMemoryUnavailable> {
+    // Majors of the JVMs left out; 0 for an unknown version.
+    let skipped = std::cell::RefCell::new(std::collections::BTreeSet::new());
+    let may_query = |pid: u32| {
+        if !suspended {
+            return true;
+        }
+        let version = probe.version(pid);
+        let safe = safe_while_suspended(version.as_ref());
+        if !safe {
+            skipped.borrow_mut().insert(version.map_or(0, |v| v.major));
+        }
+        safe
+    };
+    match probe.find_by_jdwp_port(port, &may_query).and_then(|p| i32::try_from(p).ok()) {
+        Some(pid) if pid > 0 => Ok(pid),
+        _ => {
+            let skipped = skipped.into_inner();
+            Err(match skipped.first() {
+                None => JvmMemoryUnavailable::AttachedJvmNotFound,
+                Some(&major) => JvmMemoryUnavailable::UnsafeWhileSuspended {
+                    java_major: if skipped.len() == 1 { major } else { 0 },
+                },
+            })
+        }
+    }
+}
+
+fn pid_tree(pid: i32) -> Vec<i32> {
+    #[cfg(unix)]
+    {
+        ide_core::process::process_tree(pid)
+    }
+    #[cfg(not(unix))]
+    {
+        vec![pid]
+    }
+}
@@ -0,0 +1,146 @@
+//! Finding the JVM behind an attach session by its JDWP port.
+
+use super::safe_while_suspended;
+use crate::session::Session;
+use serde_json::Value;
+use std::net::IpAddr;
+use std::sync::Arc;
+use std::sync::atomic::Ordering;
+use std::time::Duration;
+
+/// One try of the scan at attach time, and the limit per socket within it: a silent socket's
+/// query thread lives that long, so a longer per-socket limit piles threads up across tries.
+const ATTACH_SCAN_TRY: Duration = Duration::from_millis(500);
+const ATTACH_SCAN_INTERVAL: Duration = Duration::from_millis(150);
+/// The JVM's attach socket can appear a little after the JDWP "Listening" banner, which is
+/// when the IDE attaches.
+const ATTACH_SCAN_WINDOW: Duration = Duration::from_secs(3);
+/// The single scan a refresh makes while the attach-time one found nothing.
+pub(super) const RESCAN_PER_SOCKET: Duration = Duration::from_secs(1);
+pub(super) const RESCAN_BUDGET: Duration = Duration::from_secs(2);
+
+/// JDWP port of an attach to this machine (`hostName` loopback or absent): only such a JVM has
+/// an attach socket the IDE can find.
+pub(crate) fn local_jdwp_port(arguments: &Value) -> Option<u16> {
+    let local = match arguments.get("hostName") {
+        None | Some(Value::Null) => true,
+        Some(Value::String(host)) => {
+            let host = host.trim();
+            let ip = host.strip_prefix('[').and_then(|h| h.strip_suffix(']')).unwrap_or(host);
+            host.eq_ignore_ascii_case("localhost") || ip.parse::<IpAddr>().is_ok_and(|ip| ip.is_loopback())
+        }
+        Some(_) => false,
+    };
+    jdwp_port(arguments).filter(|_| local)
+}
+
+/// `port` of an attach request: a number or a numeric string, never 0.
+fn jdwp_port(arguments: &Value) -> Option<u16> {
+    let port = match arguments.get("port")? {
+        Value::Number(n) => n.as_u64().and_then(|p| u16::try_from(p).ok()),
+        Value::String(s) => s.trim().parse().ok(),
+        _ => None,
+    };
+    port.filter(|&p| p != 0)
+}
+
+/// `hostName` (empty when absent) and `port` of an attach request, local or not.
+pub(crate) fn attach_target(arguments: &Value) -> Option<(String, u16)> {
+    let host = arguments.get("hostName").and_then(Value::as_str).unwrap_or_default();
+    Some((host.to_owned(), jdwp_port(arguments)?))
+}
+
+/// Pid of the local JVM listening on a JDWP port (first argument), given whether the session is
+/// suspended (second). Replaceable so tests never send attach requests to real JVMs.
+pub(crate) type JvmFinder = Arc<dyn Fn(u16, bool) -> Option<u32> + Send + Sync>;
+
+/// The [`JvmFinder`] of the attach-time scan.
+pub(crate) fn find_local_jvm(port: u16, suspended: bool) -> Option<u32> {
+    let may_query = |pid| !suspended || safe_while_suspended(ide_jvm_attach::java_version_of(pid).as_ref());
+    ide_jvm_attach::find_jvm_by_jdwp_port(port, ATTACH_SCAN_TRY, ATTACH_SCAN_TRY, may_query)
+}
+
+/// Looks for the JVM behind a local attach: one try now, the rest in the background, so a JVM
+/// that never opens an attach socket holds the attach up by one try at most. Call it before the
+/// debugger connects: the JVM clears its JDWP listener address then, and a scan can no longer
+/// tell it by its port.
+pub(crate) async fn find_attached_jvm(session: &Arc<Session>, port: u16, finder: JvmFinder) {
+    let _ = session.jdwp_port.set(port);
+    let find = move |suspended: bool| finder(port, suspended);
+    let first = find.clone();
+    if let Ok(Some(pid)) = tokio::task::spawn_blocking(move || first(false)).await {
+        let _ = session.jvm_pid.compare_exchange(0, i64::from(pid), Ordering::AcqRel, Ordering::Acquire);
+        return;
+    }
+    drop(watch_for_attached_jvm(session, port, find, ATTACH_SCAN_INTERVAL, ATTACH_SCAN_WINDOW));
+}
+
+/// Retries `find` for the session's `jvm_pid`, holding the session weakly so an ended session
+/// is not kept alive. `find` is told whether the session is suspended.
+pub(crate) fn watch_for_attached_jvm<F>(
+    session: &Arc<Session>,
+    port: u16,
+    find: F,
+    interval: Duration,
+    window: Duration,
+) -> tokio::task::JoinHandle<()>
+where
+    F: Fn(bool) -> Option<u32> + Clone + Send + 'static,
+{
+    let weak = Arc::downgrade(session);
+    tokio::spawn(async move {
+        let gate = || match weak.upgrade() {
+            Some(s) if !s.terminated.load(Ordering::Acquire) && !s.conn.is_closed() => {
+                if s.jvm_pid.load(Ordering::Acquire) > 0 {
+                    ScanGate::Ended
+                } else {
+                    ScanGate::Scan { suspended: s.suspended.get() }
+                }
+            }
+            _ => ScanGate::Ended,
+        };
+        let Some(pid) = retry_find(find, gate, interval, window).await else {
+            tracing::debug!("no local JVM found for JDWP port {port} yet");
+            return;
+        };
+        if let Some(s) = weak.upgrade() {
+            let _ = s.jvm_pid.compare_exchange(0, i64::from(pid), Ordering::AcqRel, Ordering::Acquire);
+        }
+    })
+}
+
+#[derive(Clone, Copy, Debug, PartialEq)]
+pub(crate) enum ScanGate {
+    Scan { suspended: bool },
+    Ended,
+}
+
+/// Calls `find` on a blocking thread, on each try `gate` allows, until it returns a pid, `gate`
+/// reports the end or `window` has passed.
+pub(crate) async fn retry_find<F>(
+    find: F,
+    gate: impl Fn() -> ScanGate,
+    interval: Duration,
+    window: Duration,
+) -> Option<u32>
+where
+    F: Fn(bool) -> Option<u32> + Clone + Send + 'static,
+{
+    let deadline = tokio::time::Instant::now() + window;
+    loop {
+        match gate() {
+            ScanGate::Ended => break,
+            ScanGate::Scan { suspended } => {
+                let find = find.clone();
+                if let Ok(Some(pid)) = tokio::task::spawn_blocking(move || find(suspended)).await {
+                    return Some(pid);
+                }
+            }
+        }
+        if tokio::time::Instant::now() + interval >= deadline {
+            break;
+        }
+        tokio::time::sleep(interval).await;
+    }
+    None
+}
@@ -0,0 +1,214 @@
+//! One class histogram of a JVM and its deltas against the previous one.
+
+use super::scan::{RESCAN_BUDGET, RESCAN_PER_SOCKET};
+use super::{
+    ClassCounts, HISTOGRAM_TIMEOUT, JvmClassRow, JvmMemorySnapshot, JvmMemorySource, JvmMemoryUnavailable,
+    safe_while_suspended,
+};
+use ide_core::IdeResult;
+use ide_jvm_attach::{AttachError, HeapDumpReply, HistogramRow, JavaVersion};
+use std::collections::HashMap;
+use std::path::Path;
+use std::time::Duration;
+
+/// Version probe for a JVM whose `release` file could not be read.
+const VERSION_TIMEOUT: Duration = Duration::from_secs(2);
+const FLAGS_TIMEOUT: Duration = Duration::from_secs(2);
+
+/// The attach calls a snapshot or a heap dump makes, so tests can stand in for a JVM.
+pub(crate) trait AttachProbe: Send + Sync {
+    fn has_listener(&self, pid: u32) -> bool;
+    /// Asks only the JVMs whose pid `may_query` accepts.
+    fn find_by_jdwp_port(&self, port: u16, may_query: &dyn Fn(u32) -> bool) -> Option<u32>;
+    fn version(&self, pid: u32) -> Option<JavaVersion>;
+    fn histogram(&self, pid: u32) -> Result<String, AttachError>;
+    fn alive(&self, pid: u32) -> bool;
+    fn vm_flags(&self, pid: u32) -> Result<String, AttachError>;
+    fn heap_dump(&self, pid: u32, path: &Path, timeout: Duration) -> Result<HeapDumpReply, AttachError>;
+    /// `None` where the free space is not known.
+    fn free_bytes(&self, dir: &Path) -> Option<u64>;
+}
+
+pub(super) struct JdkProbe {
+    pub(super) suspended: bool,
+}
+
+impl AttachProbe for JdkProbe {
+    fn has_listener(&self, pid: u32) -> bool {
+        ide_jvm_attach::has_listener(pid)
+    }
+
+    fn find_by_jdwp_port(&self, port: u16, may_query: &dyn Fn(u32) -> bool) -> Option<u32> {
+        ide_jvm_attach::find_jvm_by_jdwp_port(port, RESCAN_PER_SOCKET, RESCAN_BUDGET, may_query)
+    }
+
+    fn version(&self, pid: u32) -> Option<JavaVersion> {
+        ide_jvm_attach::java_version_of(pid).or_else(|| {
+            // The version decides whether attaching to a suspended VM hangs it, so never
+            // attach to learn it while suspended.
+            if self.suspended {
+                return None;
+            }
+            let out = ide_jvm_attach::jcmd(pid, "VM.version", VERSION_TIMEOUT).ok()?;
+            ide_jvm_attach::parse_vm_version(&out)
+        })
+    }
+
+    fn histogram(&self, pid: u32) -> Result<String, AttachError> {
+        ide_jvm_attach::jcmd(pid, "GC.class_histogram -all", HISTOGRAM_TIMEOUT)
+    }
+
+    fn alive(&self, pid: u32) -> bool {
+        alive(pid)
+    }
+
+    fn vm_flags(&self, pid: u32) -> Result<String, AttachError> {
+        ide_jvm_attach::vm_flags(pid, FLAGS_TIMEOUT)
+    }
+
+    fn heap_dump(&self, pid: u32, path: &Path, timeout: Duration) -> Result<HeapDumpReply, AttachError> {
+        ide_jvm_attach::heap_dump(pid, path, timeout)
+    }
+
+    fn free_bytes(&self, dir: &Path) -> Option<u64> {
+        ide_heap::free_bytes(dir)
+    }
+}
+
+fn alive(pid: u32) -> bool {
+    #[cfg(unix)]
+    {
+        // 0 and negative pids address process groups.
+        let Ok(pid) = i32::try_from(pid) else { return false };
+        // SAFETY: signal 0 only checks that the process exists and may be signalled.
+        pid > 0
+            && (unsafe { libc::kill(pid, 0) } == 0
+                || std::io::Error::last_os_error().raw_os_error() == Some(libc::EPERM))
+    }
+    #[cfg(not(unix))]
+    {
+        let _ = pid;
+        false
+    }
+}
+
+/// Why `pid`'s attach listener cannot answer: it stopped listening or the program ended.
+pub(super) fn gone(probe: &dyn AttachProbe, pid: u32) -> JvmMemoryUnavailable {
+    if probe.alive(pid) { JvmMemoryUnavailable::ListenerNotRunning } else { JvmMemoryUnavailable::ProgramExited }
+}
+
+/// The reason for an attach request (`what`) to `jvm` that failed, or an error where no reason
+/// fits.
+pub(super) fn attach_failure(
+    probe: &dyn AttachProbe,
+    jvm: u32,
+    error: AttachError,
+    what: &str,
+) -> IdeResult<JvmMemoryUnavailable> {
+    use JvmMemoryUnavailable as U;
+    match error {
+        AttachError::Timeout => Ok(U::TimedOut),
+        AttachError::Refused | AttachError::NoListener => Ok(gone(probe, jvm)),
+        AttachError::ForeignListener => {
+            tracing::warn!("attach socket of JVM {jvm} is no longer private to this user; not using it");
+            Ok(U::ListenerNotRunning)
+        }
+        AttachError::Unsupported => Ok(U::UnsupportedPlatform),
+        e => Err(ide_core::IdeError::internal(format!("{what} of JVM {jvm}: {e}"))),
+    }
+}
+
+pub(crate) fn pick_jvm_pid(tree: &[i32], has_listener: impl Fn(u32) -> bool) -> Option<u32> {
+    tree.iter().filter_map(|&p| u32::try_from(p).ok()).filter(|&p| p > 0).find(|&p| has_listener(p))
+}
+
+pub(crate) fn diff_rows(rows: Vec<HistogramRow>, previous: Option<&ClassCounts>) -> (Vec<JvmClassRow>, ClassCounts) {
+    let mut seen: HashMap<&str, u32> = HashMap::with_capacity(rows.len());
+    for r in &rows {
+        *seen.entry(r.name.as_str()).or_default() += 1;
+    }
+    let ambiguous: Vec<bool> = rows.iter().map(|r| seen[r.name.as_str()] > 1).collect();
+    drop(seen);
+
+    let mut counts = ClassCounts::with_capacity(rows.len());
+    let mut out: Vec<JvmClassRow> = rows
+        .into_iter()
+        .zip(ambiguous)
+        .map(|(r, ambiguous)| {
+            counts.insert(r.name.clone(), (!ambiguous).then_some((r.instances, r.bytes)));
+            let delta = match previous {
+                _ if ambiguous => None,
+                None => None,
+                Some(p) => match p.get(&r.name) {
+                    Some(None) => None,
+                    // Absent: new since the previous snapshot, so it grew from nothing.
+                    before => {
+                        let (instances, bytes) = before.copied().flatten().unwrap_or_default();
+                        Some((delta(r.instances, instances), delta(r.bytes, bytes)))
+                    }
+                },
+            };
+            JvmClassRow {
+                name: r.name,
+                module: r.module,
+                instances: r.instances,
+                bytes: r.bytes,
+                delta_instances: delta.map(|d| d.0),
+                delta_bytes: delta.map(|d| d.1),
+                ambiguous,
+            }
+        })
+        .collect();
+    out.sort_by(|a, b| b.bytes.cmp(&a.bytes).then_with(|| a.name.cmp(&b.name)));
+    (out, counts)
+}
+
+/// Exact while the difference fits in an i64 (two's complement).
+fn delta(now: u64, before: u64) -> i64 {
+    now.wrapping_sub(before) as i64
+}
+
+/// `pid_tree` starts with the session's reported pid, followed by its descendants.
+pub(crate) fn take_snapshot(
+    probe: &dyn AttachProbe,
+    pid_tree: &[i32],
+    suspended: bool,
+    previous: Option<&ClassCounts>,
+) -> IdeResult<Result<(JvmMemorySnapshot, ClassCounts), JvmMemoryUnavailable>> {
+    use JvmMemoryUnavailable as U;
+    let Some(root) = pid_tree.first().and_then(|&p| u32::try_from(p).ok()) else {
+        return Ok(Err(U::NoJvmProcess));
+    };
+    let Some(jvm) = pick_jvm_pid(pid_tree, |p| probe.has_listener(p)) else {
+        return Ok(Err(gone(probe, root)));
+    };
+
+    let version = probe.version(jvm);
+    if suspended && !safe_while_suspended(version.as_ref()) {
+        return Ok(Err(U::UnsafeWhileSuspended { java_major: version.map_or(0, |v| v.major) }));
+    }
+
+    let started = std::time::Instant::now();
+    let text = match probe.histogram(jvm) {
+        Ok(text) => text,
+        Err(e) => return attach_failure(probe, jvm, e, "class histogram").map(Err),
+    };
+    let elapsed_ms = u32::try_from(started.elapsed().as_millis()).unwrap_or(u32::MAX);
+
+    let histogram = ide_jvm_attach::parse_class_histogram(&text);
+    if histogram.skipped_lines > 0 {
+        tracing::debug!("class histogram of JVM {jvm}: {} unreadable lines", histogram.skipped_lines);
+    }
+    let (rows, counts) = diff_rows(histogram.rows, previous);
+    let snapshot = JvmMemorySnapshot {
+        total_instances: rows.iter().map(|r| r.instances).sum(),
+        total_bytes: rows.iter().map(|r| r.bytes).sum(),
+        rows,
+        java_version: version.map(|v| v.text).unwrap_or_default(),
+        taken_while_suspended: suspended,
+        has_baseline: previous.is_some(),
+        elapsed_ms,
+        source: JvmMemorySource::Histogram,
+    };
+    Ok(Ok((snapshot, counts)))
+}
@@ -0,0 +1,119 @@
+//! A scripted [`JvmInspector`] and a session without an adapter behind it.
+
+use crate::SessionOwner;
+use crate::adapter::{AdapterSpec, Transport};
+use crate::connection::Connection;
+use crate::session::Session;
+use ide_core::{DebugSessionId, IdeEvent, SharedSink};
+use ide_jvm_inspect::{
+    ClassCount, InspectError, InspectFuture, JvmField, JvmInspector, JvmReferrer, JvmValue, ObjectId, VmTarget,
+};
+use parking_lot::Mutex;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicUsize, Ordering};
+use tokio::sync::Notify;
+
+#[derive(Default)]
+pub(crate) struct FakeInspector {
+    pub targets: Mutex<Vec<VmTarget>>,
+    /// The `limit` of each `instances` / `referrers` call.
+    pub limits: Mutex<Vec<u32>>,
+    pub calls: AtomicUsize,
+    /// Returned by every call when set.
+    pub error: Mutex<Option<InspectError>>,
+    pub counts: Mutex<Vec<ClassCount>>,
+    pub instances: Mutex<Vec<JvmValue>>,
+    /// Every call signals `entered`, then waits for `release`.
+    pub hold: bool,
+    pub entered: Notify,
+    pub release: Notify,
+}
+
+impl FakeInspector {
+    pub fn failing(error: InspectError) -> Self {
+        Self { error: Mutex::new(Some(error)), ..Self::default() }
+    }
+
+    pub fn calls(&self) -> usize {
+        self.calls.load(Ordering::SeqCst)
+    }
+
+    async fn answer<T: Send>(&self, target: &VmTarget, limit: Option<u32>, ok: T) -> Result<T, InspectError> {
+        self.calls.fetch_add(1, Ordering::SeqCst);
+        self.targets.lock().push(target.clone());
+        self.limits.lock().extend(limit);
+        if self.hold {
+            self.entered.notify_one();
+            self.release.notified().await;
+        }
+        match self.error.lock().clone() {
+            Some(e) => Err(e),
+            None => Ok(ok),
+        }
+    }
+}
+
+impl JvmInspector for FakeInspector {
+    fn class_counts<'a>(&'a self, target: &'a VmTarget) -> InspectFuture<'a, Vec<ClassCount>> {
+        Box::pin(self.answer(target, None, self.counts.lock().clone()))
+    }
+
+    fn instances<'a>(&'a self, target: &'a VmTarget, _: &'a str, limit: u32) -> InspectFuture<'a, Vec<JvmValue>> {
+        Box::pin(self.answer(target, Some(limit), self.instances.lock().clone()))
+    }
+
+    fn fields<'a>(&'a self, target: &'a VmTarget, _: ObjectId) -> InspectFuture<'a, Vec<JvmField>> {
+        Box::pin(self.answer(target, None, Vec::new()))
+    }
+
+    fn referrers<'a>(&'a self, target: &'a VmTarget, _: ObjectId, limit: u32) -> InspectFuture<'a, Vec<JvmReferrer>> {
+        Box::pin(self.answer(target, Some(limit), Vec::new()))
+    }
+}
+
+pub(crate) struct BareSession {
+    pub session: Arc<Session>,
+    pub events: Arc<Mutex<Vec<IdeEvent>>>,
+    /// The adapter's end of the connection, kept open.
+    _adapter: tokio::io::DuplexStream,
+}
+
+/// A java-debug session that no adapter answers; must be called inside a Tokio runtime.
+pub(crate) fn bare_session(request: &str) -> BareSession {
+    bare_session_of("java-debug", request)
+}
+
+/// A session of adapter `name` that no adapter answers; must be called inside a Tokio runtime.
+pub(crate) fn bare_session_of(name: &str, request: &str) -> BareSession {
+    let events: Arc<Mutex<Vec<IdeEvent>>> = Arc::default();
+    let sink_events = events.clone();
+    let sink: SharedSink = Arc::new(move |e: IdeEvent| sink_events.lock().push(e));
+    let (client, adapter) = tokio::io::duplex(1 << 10);
+    let (reader, writer) = tokio::io::split(client);
+    let (conn, _) = Connection::start(reader, writer);
+    let spec = AdapterSpec {
+        adapter: name.into(),
+        command: String::new(),
+        args: Vec::new(),
+        transport: Transport::Stdio,
+        install_hint: None,
+        cwd: std::env::temp_dir(),
+        env: Vec::new(),
+        server_host: None,
+        connect: None,
+        pinned: None,
+    };
+    let session = Session::new(
+        DebugSessionId(1),
+        SessionOwner::new("test", sink),
+        "test".into(),
+        name.into(),
+        None,
+        request.into(),
+        conn,
+        spec,
+        None,
+        None,
+    );
+    BareSession { session: Arc::new(session), events, _adapter: adapter }
+}
@@ -9,6 +9,13 @@ use std::{

 pub const ADAPTER: &str = "kotlin-debug-adapter";

+/// The adapter attaches to a JVM the IDE starts, through the JDWP link. It resumes the VM while
+/// constructing its debuggee, before its breakpoint hooks exist, so the link holds resumes back
+/// until the attach is done.
+pub(crate) fn link_options() -> ide_jdwp::LinkOptions {
+    ide_jdwp::LinkOptions { hold_resumes: true }
+}
+
 pub struct SourceMirror {
    dir: tempfile::TempDir,
    paths: HashMap<PathBuf, PathBuf>,
@@ -1,166 +0,0 @@
-//! Keep a suspended JVM stopped until Kotlin's adapter has installed its listeners.
-//! The official 0.4.4 adapter resumes VMStart while constructing its debuggee, before
-//! installing breakpoint hooks. DAP attach completes only after those hooks exist.
-use ide_core::IdeResult;
-use std::{io, time::Duration};
-use tokio::{
-    io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt},
-    net::{TcpListener, TcpStream},
-    sync::oneshot,
-    task::JoinHandle,
-};
-
-pub(crate) struct StartupRelay {
-    release: Option<oneshot::Sender<()>>,
-    task: JoinHandle<()>,
-}
-
-impl StartupRelay {
-    pub async fn start(host: String, port: u16) -> IdeResult<(Self, u16)> {
-        let listener = TcpListener::bind(("127.0.0.1", 0)).await?;
-        let local = listener.local_addr()?.port();
-        let (release, ready) = oneshot::channel();
-        let task = tokio::spawn(async move {
-            let connect = async {
-                let (adapter, _) = listener.accept().await?;
-                let jvm = TcpStream::connect((host.as_str(), port)).await?;
-                adapter.set_nodelay(true)?;
-                jvm.set_nodelay(true)?;
-                Ok::<_, io::Error>((adapter, jvm))
-            };
-            if let Ok(Ok((adapter, jvm))) = tokio::time::timeout(Duration::from_secs(60), connect).await {
-                let (mut ar, mut aw) = adapter.into_split();
-                let (mut jr, mut jw) = jvm.into_split();
-                let upstream = forward(&mut ar, &mut jw, ready);
-                let downstream = tokio::io::copy(&mut jr, &mut aw);
-                // Close both sockets as soon as either side disconnects.
-                tokio::select! {
-                    r = upstream => if let Err(e) = r { tracing::debug!("Kotlin JDWP relay: {e}"); },
-                    _ = downstream => {},
-                }
-            }
-        });
-        Ok((Self { release: Some(release), task }, local))
-    }
-
-    pub fn release(&mut self) {
-        if let Some(ready) = self.release.take() {
-            let _ = ready.send(());
-        }
-    }
-}
-
-impl Drop for StartupRelay {
-    fn drop(&mut self) {
-        self.task.abort();
-    }
-}
-
-async fn packet(reader: &mut (impl AsyncRead + Unpin)) -> io::Result<Vec<u8>> {
-    let mut header = [0; 11];
-    reader.read_exact(&mut header).await?;
-    let len = u32::from_be_bytes(header[..4].try_into().unwrap()) as usize;
-    if !(11..=16 * 1024 * 1024).contains(&len) {
-        return Err(io::Error::new(io::ErrorKind::InvalidData, "Invalid JDWP packet length"));
-    }
-    let mut packet = vec![0; len];
-    packet[..11].copy_from_slice(&header);
-    reader.read_exact(&mut packet[11..]).await?;
-    Ok(packet)
-}
-
-async fn forward(
-    reader: &mut (impl AsyncRead + Unpin),
-    writer: &mut (impl AsyncWrite + Unpin),
-    mut ready: oneshot::Receiver<()>,
-) -> io::Result<()> {
-    let mut handshake = [0; 14];
-    reader.read_exact(&mut handshake).await?;
-    if &handshake != b"JDWP-Handshake" {
-        return Err(io::Error::new(io::ErrorKind::InvalidData, "Invalid JDWP handshake"));
-    }
-    writer.write_all(&handshake).await?;
-    let mut released = false;
-    let mut held: Vec<Vec<u8>> = Vec::new();
-    loop {
-        // Retain the in-flight read when attach completes, including partial packets.
-        let read = packet(reader);
-        tokio::pin!(read);
-        let packet = tokio::select! {
-            result = &mut read => result?,
-            result = &mut ready, if !released => {
-                result.map_err(|_| io::Error::new(io::ErrorKind::ConnectionAborted, "Kotlin attach cancelled"))?;
-                released = true;
-                for packet in held.drain(..) {
-                    writer.write_all(&packet).await?;
-                }
-                read.await?
-            }
-        };
-        // JDWP VirtualMachine.Resume (1/9), ThreadReference.Resume (11/3).
-        // Other commands must proceed: attach itself queries threads and installs requests.
-        if !released && packet[8] == 0 && matches!((packet[9], packet[10]), (1, 9) | (11, 3)) {
-            if held.len() >= 1024 {
-                return Err(io::Error::new(io::ErrorKind::InvalidData, "Too many pending JDWP resumes"));
-            }
-            held.push(packet);
-        } else {
-            writer.write_all(&packet).await?;
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    fn command(id: u8, set: u8, cmd: u8) -> Vec<u8> {
-        vec![0, 0, 0, 11, 0, 0, 0, id, 0, set, cmd]
-    }
-
-    #[tokio::test]
-    async fn holds_resumes_but_allows_setup_and_preserves_partial_reads() {
-        let (mut adapter, mut reader) = tokio::io::duplex(4096);
-        let (mut writer, mut jvm) = tokio::io::duplex(4096);
-        let (release, ready) = oneshot::channel();
-        let task = tokio::spawn(async move { forward(&mut reader, &mut writer, ready).await });
-        adapter.write_all(b"JDWP-Handshake").await.unwrap();
-        let mut handshake = [0; 14];
-        jvm.read_exact(&mut handshake).await.unwrap();
-        assert_eq!(&handshake, b"JDWP-Handshake");
-        let resume = command(1, 1, 9);
-        let thread_resume = command(2, 11, 3);
-        let setup = command(3, 1, 4);
-        adapter.write_all(&resume).await.unwrap();
-        adapter.write_all(&thread_resume).await.unwrap();
-        adapter.write_all(&setup).await.unwrap();
-        assert_eq!(packet(&mut jvm).await.unwrap(), setup);
-        let next = command(4, 1, 9);
-        adapter.write_all(&next[..6]).await.unwrap();
-        release.send(()).unwrap();
-        assert_eq!(packet(&mut jvm).await.unwrap(), resume);
-        assert_eq!(packet(&mut jvm).await.unwrap(), thread_resume);
-        adapter.write_all(&next[6..]).await.unwrap();
-        assert_eq!(packet(&mut jvm).await.unwrap(), next);
-        drop(adapter);
-        assert!(task.await.unwrap().is_err());
-    }
-
-    #[tokio::test]
-    async fn dropping_session_closes_relay() {
-        let (relay, port) = StartupRelay::start("127.0.0.1".into(), 1).await.unwrap();
-        let mut adapter = TcpStream::connect(("127.0.0.1", port)).await.unwrap();
-        drop(relay);
-        let result = tokio::time::timeout(Duration::from_secs(1), adapter.read_u8()).await.unwrap();
-        assert!(result.is_err());
-    }
-
-    #[tokio::test]
-    async fn rejects_invalid_packet_lengths() {
-        for len in [0_u32, 10, 16 * 1024 * 1024 + 1] {
-            let mut bytes = command(1, 1, 9);
-            bytes[..4].copy_from_slice(&len.to_be_bytes());
-            assert_eq!(packet(&mut bytes.as_slice()).await.unwrap_err().kind(), io::ErrorKind::InvalidData);
-        }
-    }
-}
@@ -13,8 +13,8 @@ pub mod attach;
 mod cargo;
 pub mod connection;
 pub mod jdwp;
+mod jvm_memory;
 pub mod kotlin;
-mod kotlin_jdwp;
 mod session;

 use adapter::{AdapterLauncher, AdapterSpec, ProcessLauncher};
@@ -22,6 +22,7 @@ use connection::{Connection, DapError};
 use futures::future::join_all;
 use ide_core::{DebugSessionId, IdeError, IdeResult, SharedSink};
 use ide_language::LanguageRegistry;
+pub use jvm_memory::*;
 use parking_lot::Mutex;
 use serde::{Deserialize, Serialize};
 use serde_json::{Value, json};
@@ -126,6 +127,14 @@ pub struct DebugCapabilities {
    pub log_points: bool,
    /// HotSwap ("Reload Changed Classes"): java-debug's `redefineClasses` request.
    pub hot_code_replace: bool,
+    /// The session debugs a JVM, so it has a Memory view; [`DebugManager::jvm_memory_classes`]
+    /// says why when it has no data (platform, remote JVM, ...).
+    pub jvm_memory: bool,
+    /// The session has a JVM inspector: instances, fields and referrers while suspended. Can
+    /// change during the session ([`ide_core::event::DebugEvent::Capabilities`]).
+    pub jvm_instances: bool,
+    /// The session can be asked for a heap dump ([`DebugManager::jvm_heap_dump`]).
+    pub jvm_heap: bool,
 }

 /// Result of [`DebugManager::redefine_classes`].
@@ -302,6 +311,7 @@ pub struct DebugManager {
    data_breakpoints: Mutex<Vec<DataBreakpoint>>,
    /// In-flight stops, by session: a second Stop sets `stop_force`, the waiting one kills.
    stopping: Mutex<HashMap<DebugSessionId, Arc<Session>>>,
+    jvm_finder: Mutex<jvm_memory::JvmFinder>,
 }

 /// rdbg (Ruby's debug gem) starts the debuggee itself and ignores `launch`'s program,
@@ -566,9 +576,21 @@ pub(crate) fn capabilities_from(caps: &Value) -> DebugCapabilities {
        hit_conditional_breakpoints: has("supportsHitConditionalBreakpoints"),
        log_points: has("supportsLogPoints"),
        hot_code_replace: false,
+        jvm_memory: false,
+        jvm_instances: false,
+        jvm_heap: false,
    }
 }

+pub(crate) fn session_capabilities(session: &Session) -> DebugCapabilities {
+    let mut caps = capabilities_from(&session.caps.lock());
+    caps.hot_code_replace = HOT_CODE_REPLACE_ADAPTERS.contains(&session.adapter.as_str());
+    caps.jvm_memory = JVM_MEMORY_ADAPTERS.contains(&session.adapter.as_str());
+    caps.jvm_instances = session.jvm_inspector.lock().is_some();
+    caps.jvm_heap = JVM_MEMORY_ADAPTERS.contains(&session.adapter.as_str());
+    caps
+}
+
 /// States of a file's line breakpoints that no session has been sent.
 fn unsent(breakpoints: &[Breakpoint]) -> Vec<BreakpointState> {
    breakpoints
@@ -657,9 +679,15 @@ impl DebugManager {
            function_breakpoints: Mutex::default(),
            data_breakpoints: Mutex::default(),
            stopping: Mutex::default(),
+            jvm_finder: Mutex::new(Arc::new(jvm_memory::find_local_jvm)),
        })
    }

+    #[cfg(test)]
+    pub(crate) fn set_jvm_finder(&self, finder: jvm_memory::JvmFinder) {
+        *self.jvm_finder.lock() = finder;
+    }
+
    fn session(&self, s: DebugSessionId) -> IdeResult<Arc<Session>> {
        self.sessions.lock().get(&s).cloned().ok_or_else(|| no_session(s))
    }
@@ -979,13 +1007,31 @@ impl DebugManager {
            .await?;
        *s.caps.lock() = if caps.is_object() { caps } else { json!({}) };

-        // The Kotlin adapter ignores the resolved classpath and project runtime on launch.
+        // The Kotlin adapters ignore the resolved classpath and project runtime on launch.
        let request = if s.adapter == kotlin::ADAPTER && s.request == "launch" {
            session::start_kotlin_debuggee(s, &mut arguments).await?;
            "attach"
        } else {
            &s.request
        };
+        // Both blocks below run before the Kotlin link replaces the JVM's address with its own.
+        if request == "attach"
+            && let Some(target) = jvm_memory::attach_target(&arguments)
+        {
+            let _ = s.attach_target.set(target);
+        }
+        if cfg!(unix)
+            && s.request == "attach"
+            && JVM_MEMORY_ADAPTERS.contains(&s.adapter.as_str())
+            && let Some(port) = jvm_memory::local_jdwp_port(&arguments)
+        {
+            if jdwp::is_listen_relay(port) {
+                s.listen_relay.store(true, Ordering::Release);
+            } else {
+                let finder = self.jvm_finder.lock().clone();
+                jvm_memory::find_attached_jvm(s, port, finder).await;
+            }
+        }
        if s.adapter == kotlin::ADAPTER {
            let host = arguments
                .get("hostName")
@@ -998,11 +1044,14 @@ impl DebugManager {
                .and_then(|p| u16::try_from(p).ok())
                .filter(|p| *p != 0)
                .ok_or_else(|| IdeError::invalid("Invalid Kotlin JDWP port"))?;
-            let (relay, port) = kotlin_jdwp::StartupRelay::start(host, port).await?;
-            *s.kotlin_relay.lock() = Some(relay);
+            let (link, port) = ide_jdwp::JdwpLink::start(host, port, kotlin::link_options()).await?;
+            *s.jvm_inspector.lock() = Some(Arc::new(ide_jdwp::LinkInspector::new(link.clone())));
+            jvm_memory::watch_link(s, &link);
+            *s.jdwp_link.lock() = Some(link);
            arguments["hostName"] = json!("127.0.0.1");
            arguments["port"] = json!(port);
        }
+        s.jvm_located.send_replace(true);
        // Launch/attach may wait for a slow runtime or a remote target. Let the adapter report
        // failure; Stop cancels configuration and a closed connection fails the pending reply.
        // Many adapters answer only after `configurationDone`, so don't wait for the reply yet.
@@ -1097,8 +1146,8 @@ impl DebugManager {
                reply.await.unwrap_or(Err(DapError::Closed))?;
            }
        }
-        if let Some(relay) = s.kotlin_relay.lock().as_mut() {
-            relay.release();
+        if let Some(link) = s.jdwp_link.lock().as_ref() {
+            link.release();
        }
        Ok(())
    }
@@ -1331,8 +1380,9 @@ impl DebugManager {
    }

    pub async fn threads(&self, s: DebugSessionId) -> IdeResult<Vec<ThreadInfo>> {
-        let body = self.session(s)?.conn.request_timeout("threads", Value::Null, REQUEST_TIMEOUT).await?;
-        Ok(body
+        let session = self.session(s)?;
+        let body = session.conn.request_timeout("threads", Value::Null, REQUEST_TIMEOUT).await?;
+        let threads: Vec<ThreadInfo> = body
            .get("threads")
            .and_then(Value::as_array)
            .map(|t| {
@@ -1343,7 +1393,8 @@ impl DebugManager {
                    })
                    .collect()
            })
-            .unwrap_or_default())
+            .unwrap_or_default();
+        Ok(threads)
    }

    pub async fn stack_trace(&self, s: DebugSessionId, thread_id: i64) -> IdeResult<Vec<StackFrame>> {
@@ -1472,8 +1523,11 @@ impl DebugManager {

    pub async fn resume(&self, s: DebugSessionId, thread_id: i64) -> IdeResult<()> {
        let session = self.session(s)?;
+        let ticket = session.suspended.ticket();
        let body = session.conn.request_timeout("continue", json!({ "threadId": thread_id }), REQUEST_TIMEOUT).await?;
+        // DAP: a response without the flag continued every thread.
        let all = body.get("allThreadsContinued").and_then(Value::as_bool).unwrap_or(true);
+        session.suspended.resumed(ticket, thread_id, all);
        session.emit(ide_core::event::DebugEvent::Continued { thread_id, all_threads_continued: all });
        Ok(())
    }
@@ -1513,7 +1567,10 @@ impl DebugManager {
            StepKind::Out => "stepOut",
        };
        let session = self.session(s)?;
+        let ticket = session.suspended.ticket();
        session.conn.request_timeout(command, json!({ "threadId": thread_id }), REQUEST_TIMEOUT).await?;
+        // A step runs its thread; other threads stopped by their own events stay stopped.
+        session.suspended.resumed(ticket, thread_id, false);
        session.emit(ide_core::event::DebugEvent::Continued { thread_id, all_threads_continued: true });
        Ok(())
    }
@@ -1611,10 +1668,12 @@ impl DebugManager {
    /// Step into a specific call of the current line (Smart Step Into).
    pub async fn step_in_target(&self, s: DebugSessionId, thread_id: i64, target_id: i64) -> IdeResult<()> {
        let session = self.session(s)?;
+        let ticket = session.suspended.ticket();
        session
            .conn
            .request_timeout("stepIn", json!({ "threadId": thread_id, "targetId": target_id }), REQUEST_TIMEOUT)
            .await?;
+        session.suspended.resumed(ticket, thread_id, false);
        session.emit(ide_core::event::DebugEvent::Continued { thread_id, all_threads_continued: true });
        Ok(())
    }
@@ -1648,10 +1707,68 @@ impl DebugManager {
    }

    pub fn capabilities(&self, s: DebugSessionId) -> IdeResult<DebugCapabilities> {
+        Ok(session_capabilities(&*self.session(s)?))
+    }
+
+    /// Class table of the session's JVM (`GC.class_histogram -all`, or counted by the debugger
+    /// while suspended where no histogram can be taken), with deltas against the session's
+    /// previous snapshot from the same source.
+    pub async fn jvm_memory_classes(&self, s: DebugSessionId) -> IdeResult<JvmMemoryResult> {
        let session = self.session(s)?;
-        let mut caps = capabilities_from(&session.caps.lock());
-        caps.hot_code_replace = HOT_CODE_REPLACE_ADAPTERS.contains(&session.adapter.as_str());
-        Ok(caps)
+        jvm_memory::session_classes(session).await
+    }
+
+    /// Writes a heap dump of the session's JVM to `dir/dump.hprof`, one class-table or dump
+    /// request per session at a time. Anything but `Ready` deletes `dir`, unless `s` is not a
+    /// session. On `cancel` the result is `Cancelled` at once, while a dump the JVM is writing
+    /// still holds the session's turn until the JVM answers.
+    pub async fn jvm_heap_dump(
+        &self,
+        s: DebugSessionId,
+        dir: PathBuf,
+        cancel: ide_core::CancellationToken,
+    ) -> IdeResult<JvmHeapOutcome> {
+        let session = self.session(s)?;
+        jvm_memory::session_heap_dump(session, dir, cancel).await
+    }
+
+    /// Gives the session the backend for [`Self::jvm_memory_instances`] and the debugger's class
+    /// counts; the UI is told to reload the session's capabilities.
+    pub fn set_jvm_inspector(
+        &self,
+        s: DebugSessionId,
+        inspector: Arc<dyn ide_jvm_inspect::JvmInspector>,
+    ) -> IdeResult<()> {
+        let session = self.session(s)?;
+        if session.ended.is_cancelled() {
+            return Err(no_session(s));
+        }
+        jvm_memory::set_inspector(&session, inspector);
+        Ok(())
+    }
+
+    /// Up to `limit` (clamped to `1..=MAX_INSTANCES`) instances of `class_name` while suspended.
+    pub async fn jvm_memory_instances(
+        &self,
+        s: DebugSessionId,
+        class_name: String,
+        limit: u32,
+    ) -> IdeResult<JvmInstancesResult> {
+        jvm_memory::instances(&*self.session(s)?, &class_name, limit).await
+    }
+
+    pub async fn jvm_memory_fields(&self, s: DebugSessionId, object: JvmObjectRef) -> IdeResult<JvmFieldsResult> {
+        jvm_memory::fields(&*self.session(s)?, object).await
+    }
+
+    /// Objects that point at `object`; `limit` 0 means `DEFAULT_REFERRERS`, at most `MAX_REFERRERS`.
+    pub async fn jvm_memory_referrers(
+        &self,
+        s: DebugSessionId,
+        object: JvmObjectRef,
+        limit: u32,
+    ) -> IdeResult<JvmReferrersResult> {
+        jvm_memory::referrers(&*self.session(s)?, object, limit).await
    }

    /// HotSwap: reloads the classes compiled since the session started (or since the
@@ -5,6 +5,8 @@ use crate::adapter::AdapterSpec;
 use crate::connection::{Connection, DapError, Incoming};
 use crate::{Breakpoint, BreakpointState, DataBreakpoint, DebugManager, FunctionBreakpoint, SessionOwner};
 use ide_core::event::DebugEvent;
+#[cfg(unix)]
+use ide_core::process::process_tree;
 use ide_core::{DebugSessionId, IdeEvent, TaskId};
 use parking_lot::Mutex;
 use serde_json::{Value, json};
@@ -62,6 +64,15 @@ pub(crate) struct Session {
    pub stop_force: AtomicBool,
    /// `systemProcessId` from the adapter's `process` event: the program being debugged.
    pub debuggee_pid: AtomicI64,
+    /// The JVM an app-launched (Kotlin) session started, or the local JVM an attach session
+    /// found by its JDWP port. Separate from `debuggee_pid`, which decides how Stop signals the
+    /// program and is 0 for Kotlin.
+    pub jvm_pid: AtomicI64,
+    /// JDWP port of an attach to a local JVM, to look the JVM up again; unset for a remote one.
+    pub jdwp_port: std::sync::OnceLock<u16>,
+    /// The attach goes through the listen-mode relay of [`crate::jdwp::listen_for_jvm`].
+    pub listen_relay: AtomicBool,
+    pub suspended: Suspension,
    /// Pids signalled for the current stop: the reported pid and everything it started (an
    /// adapter may report the launcher shell, whose child is the program).
    #[cfg_attr(not(unix), allow(dead_code))]
@@ -75,8 +86,105 @@ pub(crate) struct Session {
    /// Resolved path → the path as the IDE knows it, for adapters given resolved paths.
    resolved_paths: Mutex<HashMap<PathBuf, PathBuf>>,
    progress: Mutex<HashMap<String, TaskId>>,
-    pub kotlin_relay: Mutex<Option<crate::kotlin_jdwp::StartupRelay>>,
+    /// Relay between a Kotlin adapter and its JVM. Dropping it closes both sockets.
+    pub jdwp_link: parking_lot::Mutex<Option<ide_jdwp::JdwpLink>>,
+    pub jvm_inspector: parking_lot::Mutex<Option<Arc<dyn ide_jvm_inspect::JvmInspector>>>,
+    /// One inspector request per session at a time.
+    pub inspect_lock: tokio::sync::Mutex<()>,
+    /// Cancelled when the session ends, so a request still waiting on the inspector lets go of
+    /// it (and of the JDWP link it holds).
+    pub ended: ide_core::CancellationToken,
+    /// `hostName` (empty when absent) and `port` of the session's `attach` request, as the IDE
+    /// was given them: Kotlin then replaces them with the link's. Unlike `jdwp_port`, also set
+    /// for a remote JVM.
+    pub attach_target: std::sync::OnceLock<(String, u16)>,
+    /// Set once `configure` has recorded where the program's JVM is (the fields above, the
+    /// Kotlin link), just before the `launch`/`attach` request goes out.
+    pub jvm_located: watch::Sender<bool>,
    pub kotlin_mirror: Mutex<Option<crate::kotlin::SourceMirror>>,
+    pub jvm_memory: tokio::sync::Mutex<crate::JvmMemoryState>,
+}
+
+/// Which threads of the program are suspended: each from its `stopped` event until a
+/// `continued` event or a successful resume/step request for it. Never cleared elsewhere, since
+/// a stale "suspended" only makes a later check refuse. fwcd and java-debug stop a breakpoint's
+/// thread alone, so a thread can run while another stays stopped.
+///
+/// A resume reply ends only the stops taken before the request was sent: a thread's next
+/// `stopped` event can overtake the reply. The check and the removal happen under one lock, so
+/// no stop can land between them and be lost.
+#[derive(Default)]
+pub(crate) struct Suspension(Mutex<Stops>);
+
+#[derive(Default)]
+struct Stops {
+    /// Counts stops, and moves back to an earlier stopped thread: an object read under one
+    /// ticket is stale under another.
+    ticket: u64,
+    /// Threads stopped alone by their event, oldest first, with the ticket of that stop.
+    threads: Vec<(i64, u64)>,
+    /// The ticket of a stop of every thread (`allThreadsStopped`, or no thread id): any resume
+    /// ends it.
+    all: Option<u64>,
+}
+
+impl Stops {
+    fn latest(&self) -> Option<u64> {
+        self.threads.last().map(|&(_, ticket)| ticket).max(self.all)
+    }
+
+    /// Ends the stops taken by ticket `upto` of `thread`, or of every thread.
+    fn release(&mut self, thread: Option<i64>, every: bool, upto: u64) {
+        let latest = self.latest();
+        if every || thread.is_none() || self.all.is_some_and(|t| t <= upto) {
+            self.all = self.all.filter(|&t| t > upto);
+            self.threads.retain(|&(_, t)| t > upto);
+        } else {
+            self.threads.retain(|&(id, t)| Some(id) != thread || t > upto);
+        }
+        // Back on an earlier thread's stop, as the JDWP link moves to it with a new generation.
+        let now = self.latest();
+        if now.is_some() && now != latest {
+            self.ticket += 1;
+        }
+    }
+}
+
+impl Suspension {
+    /// Whether some thread is stopped.
+    pub fn get(&self) -> bool {
+        let stops = self.0.lock();
+        stops.all.is_some() || !stops.threads.is_empty()
+    }
+
+    /// A `stopped` event; `all`: its `allThreadsStopped`.
+    pub fn stopped(&self, thread: Option<i64>, all: bool) {
+        let mut stops = self.0.lock();
+        stops.ticket += 1;
+        let ticket = stops.ticket;
+        match thread.filter(|_| !all) {
+            Some(thread) => {
+                stops.threads.retain(|&(id, _)| id != thread);
+                stops.threads.push((thread, ticket));
+            }
+            None => stops.all = Some(ticket),
+        }
+    }
+
+    /// A `continued` event; `all`: its `allThreadsContinued`.
+    pub fn continued(&self, thread: Option<i64>, all: bool) {
+        self.0.lock().release(thread, all, u64::MAX);
+    }
+
+    /// Take before sending a resume-like request and pass to [`Self::resumed`].
+    pub fn ticket(&self) -> u64 {
+        self.0.lock().ticket
+    }
+
+    /// After a successful resume-like request for `thread`; `all`: every thread ran.
+    pub fn resumed(&self, ticket: u64, thread: i64, all: bool) {
+        self.0.lock().release(Some(thread), all, ticket);
+    }
 }

 /// Adapters that only match files by their resolved path: Godot maps a script to
@@ -125,6 +233,10 @@ impl Session {
            terminated: AtomicBool::new(false),
            stop_force: AtomicBool::new(false),
            debuggee_pid: AtomicI64::new(0),
+            jvm_pid: AtomicI64::new(0),
+            jdwp_port: std::sync::OnceLock::new(),
+            listen_relay: AtomicBool::new(false),
+            suspended: Suspension::default(),
            stop_pids: parking_lot::Mutex::new(Vec::new()),
            #[cfg(windows)]
            stop_tree: Mutex::new(None),
@@ -132,8 +244,14 @@ impl Session {
            breakpoint_ids: Mutex::new(HashMap::new()),
            resolved_paths: Mutex::new(HashMap::new()),
            progress: Mutex::new(HashMap::new()),
-            kotlin_relay: Mutex::new(None),
+            jdwp_link: parking_lot::Mutex::new(None),
+            jvm_inspector: parking_lot::Mutex::new(None),
+            inspect_lock: tokio::sync::Mutex::default(),
+            ended: ide_core::CancellationToken::new(),
+            attach_target: std::sync::OnceLock::new(),
+            jvm_located: watch::channel(false).0,
            kotlin_mirror: Mutex::new(None),
+            jvm_memory: tokio::sync::Mutex::default(),
        }
    }

@@ -294,7 +412,10 @@ impl Session {

    /// Kills the adapter and every debuggee we started.
    pub fn kill_processes(&self) {
-        self.kotlin_relay.lock().take();
+        self.ended.cancel();
+        // The inspector holds a clone of the link, so both go.
+        self.jvm_inspector.lock().take();
+        self.jdwp_link.lock().take();
        #[cfg(unix)]
        if self.request == "launch" {
            for group in launch_groups(self) {
@@ -351,32 +472,6 @@ pub(crate) async fn remember_debuggee(s: &Arc<Session>) {
    let _ = s;
 }

-/// The program and everything it started: an adapter may report the launcher shell, whose
-/// child is the debugged program, so stopping only the reported pid leaves the program alive.
-#[cfg(unix)]
-fn process_tree(pid: i32) -> Vec<i32> {
-    use std::collections::HashMap;
-    let Ok(out) = std::process::Command::new("ps").args(["-A", "-o", "pid=,ppid="]).output() else {
-        return vec![pid];
-    };
-    let mut children: HashMap<i32, Vec<i32>> = HashMap::new();
-    for line in String::from_utf8_lossy(&out.stdout).lines() {
-        let mut it = line.split_whitespace();
-        let (Some(p), Some(parent)) = (it.next(), it.next()) else { continue };
-        let (Ok(p), Ok(parent)) = (p.parse::<i32>(), parent.parse::<i32>()) else { continue };
-        children.entry(parent).or_default().push(p);
-    }
-    let mut tree = vec![pid];
-    let mut at = 0;
-    while at < tree.len() {
-        if let Some(kids) = children.get(&tree[at]) {
-            tree.extend(kids.iter().copied());
-        }
-        at += 1;
-    }
-    tree
-}
-
 /// Asks the program itself to exit (`force` = SIGKILL) without ending the adapter session:
 /// its output keeps flowing over DAP while it shuts down. The first call remembers the pid
 /// tree, because the reported pid may die before its program does.
@@ -600,18 +695,27 @@ async fn on_event(mgr: &Arc<DebugManager>, s: &Arc<Session>, event: &str, mut bo
            s.emit(DebugEvent::Initialized);
        }
        "stopped" => {
+            let all = body.get("allThreadsStopped").and_then(Value::as_bool).unwrap_or(false);
+            s.suspended.stopped(int("threadId"), all);
+            let reason = str_("reason").unwrap_or_else(|| "pause".into());
            mgr.restore_breakpoints(s).await;
            s.emit(DebugEvent::Stopped {
                thread_id: int("threadId").unwrap_or(0),
-                reason: str_("reason").unwrap_or_else(|| "pause".into()),
+                reason,
                description: str_("description").or_else(|| str_("text")),
-                all_threads_stopped: body.get("allThreadsStopped").and_then(Value::as_bool).unwrap_or(false),
+                all_threads_stopped: all,
+            });
+        }
+        "continued" => {
+            let all = body.get("allThreadsContinued").and_then(Value::as_bool);
+            // DAP: an event without the flag continues its thread alone. The UI keeps treating
+            // it as all threads; the suspension, which holds only threads that stopped, need not.
+            s.suspended.continued(int("threadId"), all.unwrap_or(false));
+            s.emit(DebugEvent::Continued {
+                thread_id: int("threadId").unwrap_or(0),
+                all_threads_continued: all.unwrap_or(true),
            });
        }
-        "continued" => s.emit(DebugEvent::Continued {
-            thread_id: int("threadId").unwrap_or(0),
-            all_threads_continued: body.get("allThreadsContinued").and_then(Value::as_bool).unwrap_or(true),
-        }),
        "output" => {
            let category = str_("category").unwrap_or_else(|| "console".into());
            if category == "telemetry" {
@@ -917,6 +1021,19 @@ fn forward_output<R: AsyncRead + Unpin + Send + 'static>(s: &Session, mut stream
    });
 }

+/// The attach listener must be on the command line before the user's arguments, which may
+/// override it; the memory view needs it because the JVM is started by the IDE, not attached to.
+/// Not on Windows (`windows`), which the memory view does not support.
+fn kotlin_jvm_args(user: Vec<String>, windows: bool) -> Vec<String> {
+    if windows || !ide_jvm_attach::wants_listener_flag(&user.join(" ")) {
+        return user;
+    }
+    let mut args = Vec::with_capacity(user.len() + 1);
+    args.push(ide_jvm_attach::START_LISTENER_FLAG.to_owned());
+    args.extend(user);
+    args
+}
+
 /// Launch with the IDE's resolved runtime and classpath; the adapter only attaches.
 pub(crate) async fn start_kotlin_debuggee(s: &Arc<Session>, args: &mut Value) -> ide_core::IdeResult<()> {
    use tokio::io::{AsyncBufReadExt, BufReader};
@@ -931,7 +1048,7 @@ pub(crate) async fn start_kotlin_debuggee(s: &Arc<Session>, args: &mut Value) ->
        std::env::join_paths(strings("classPaths")).map_err(|e| ide_core::IdeError::invalid(e.to_string()))?;
    let mut cmd = tokio::process::Command::new(java);
    ide_core::process::prepare(cmd.as_std_mut());
-    cmd.args(strings("vmArguments"))
+    cmd.args(kotlin_jvm_args(strings("vmArguments"), cfg!(windows)))
        // Let the JVM pick its own port (`:0`) and report it below: pre-picking a free
        // port here races with other processes that may take it before the JVM binds.
        .arg("-agentlib:jdwp=transport=dt_socket,server=y,suspend=y,address=127.0.0.1:0")
@@ -965,11 +1082,12 @@ pub(crate) async fn start_kotlin_debuggee(s: &Arc<Session>, args: &mut Value) ->
    // forced stop), the session's job on Windows, so a stop reaches what it forks.
    ide_core::process::own_group(cmd.as_std_mut());
    let mut child = cmd.spawn()?;
-    #[cfg(windows)]
-    if let Some(pid) = child.id()
-        && let Some(job) = s.job.lock().as_ref()
-    {
-        job.assign_pid(pid);
+    if let Some(pid) = child.id() {
+        s.jvm_pid.store(i64::from(pid), Ordering::Release);
+        #[cfg(windows)]
+        if let Some(job) = s.job.lock().as_ref() {
+            job.assign_pid(pid);
+        }
    }
    if let Some(err) = child.stderr.take() {
        forward_output(s, err, "stderr");
@@ -1010,7 +1128,162 @@ fn parse_jdwp_port(line: &str) -> Option<u16> {

 #[cfg(test)]
 mod tests {
-    use super::parse_jdwp_port;
+    use super::{Suspension, kotlin_jvm_args, parse_jdwp_port};
+
+    #[test]
+    fn a_resume_reply_clears_the_stop_it_was_sent_for() {
+        let s = Suspension::default();
+        assert!(!s.get());
+        s.stopped(Some(1), true);
+        let ticket = s.ticket();
+        s.resumed(ticket, 1, true);
+        assert!(!s.get());
+    }
+
+    #[test]
+    fn a_stop_after_the_resume_request_keeps_the_session_suspended() {
+        let s = Suspension::default();
+        s.stopped(Some(1), true);
+        let ticket = s.ticket();
+        // The next breakpoint's `stopped` event overtakes the reply to `continue`.
+        s.continued(Some(1), true);
+        s.stopped(Some(1), true);
+        s.resumed(ticket, 1, true);
+        assert!(s.get());
+        // A reply for the new stop still clears it.
+        s.resumed(s.ticket(), 1, true);
+        assert!(!s.get());
+    }
+
+    #[test]
+    fn a_reply_after_a_continued_event_changes_nothing() {
+        let s = Suspension::default();
+        s.stopped(Some(1), true);
+        let ticket = s.ticket();
+        s.continued(Some(1), true);
+        s.resumed(ticket, 1, true);
+        assert!(!s.get());
+        s.stopped(Some(1), true);
+        assert!(s.get());
+    }
+
+    #[test]
+    fn concurrent_stops_are_never_lost() {
+        let s = std::sync::Arc::new(Suspension::default());
+        s.stopped(Some(1), true);
+        let ticket = s.ticket();
+        let stopping = {
+            let s = s.clone();
+            std::thread::spawn(move || (0..10_000).for_each(|_| s.stopped(Some(1), true)))
+        };
+        for _ in 0..10_000 {
+            s.resumed(ticket, 1, true);
+        }
+        stopping.join().unwrap();
+        // However they interleave, a resume for the old ticket never clears a later stop.
+        assert!(s.get());
+        assert_eq!(s.ticket(), ticket + 10_000);
+    }
+
+    /// fwcd and java-debug stop a breakpoint's thread alone (`SUSPEND_EVENT_THREAD`): 7, then 8.
+    fn two_threads_stopped() -> Suspension {
+        let s = Suspension::default();
+        s.stopped(Some(7), false);
+        s.stopped(Some(8), false);
+        s
+    }
+
+    #[test]
+    fn continuing_the_latest_stopped_thread_keeps_the_other_as_a_new_stop() {
+        let s = two_threads_stopped();
+        let ticket = s.ticket();
+        s.continued(Some(8), false);
+        assert!(s.get());
+        // Objects read on thread 8's stop are stale.
+        assert_eq!(s.ticket(), ticket + 1);
+        s.continued(Some(7), false);
+        assert!(!s.get());
+    }
+
+    #[test]
+    fn continuing_an_earlier_stopped_thread_keeps_the_latest_stop() {
+        let s = two_threads_stopped();
+        let ticket = s.ticket();
+        s.continued(Some(7), false);
+        assert!(s.get());
+        assert_eq!(s.ticket(), ticket);
+    }
+
+    #[test]
+    fn continuing_all_threads_ends_every_stop() {
+        let s = two_threads_stopped();
+        s.continued(Some(7), true);
+        assert!(!s.get());
+        let s = two_threads_stopped();
+        s.continued(None, false);
+        assert!(!s.get());
+    }
+
+    #[test]
+    fn continuing_a_thread_that_did_not_stop_changes_nothing() {
+        let s = two_threads_stopped();
+        let ticket = s.ticket();
+        s.continued(Some(5), false);
+        assert!(s.get());
+        assert_eq!(s.ticket(), ticket);
+    }
+
+    #[test]
+    fn a_step_reply_resumes_its_thread_alone() {
+        let s = two_threads_stopped();
+        let ticket = s.ticket();
+        s.resumed(ticket, 8, false);
+        assert!(s.get());
+        assert_eq!(s.ticket(), ticket + 1);
+        s.resumed(s.ticket(), 7, true);
+        assert!(!s.get());
+    }
+
+    #[test]
+    fn a_thread_stopping_again_before_its_resume_reply_stays_stopped() {
+        let s = two_threads_stopped();
+        let ticket = s.ticket();
+        // Thread 7's step ends, and its `stopped` event overtakes the reply to `next`.
+        s.stopped(Some(7), false);
+        s.resumed(ticket, 7, false);
+        s.continued(Some(8), false);
+        assert!(s.get());
+    }
+
+    #[test]
+    fn a_stop_of_every_thread_ends_with_any_resume() {
+        let s = Suspension::default();
+        s.stopped(Some(7), false);
+        s.stopped(Some(8), true);
+        s.continued(Some(8), false);
+        assert!(!s.get());
+        // A stop without a thread.
+        s.stopped(None, false);
+        s.resumed(s.ticket(), 5, false);
+        assert!(!s.get());
+    }
+
+    #[test]
+    fn kotlin_launch_starts_attach_listener_first() {
+        assert_eq!(
+            kotlin_jvm_args(vec!["-Xmx1g".into()], false),
+            vec!["-XX:+StartAttachListener".to_string(), "-Xmx1g".into()]
+        );
+        assert_eq!(
+            kotlin_jvm_args(vec!["-XX:+DisableAttachMechanism".into()], false),
+            vec!["-XX:+DisableAttachMechanism".to_string()]
+        );
+    }
+
+    #[test]
+    fn kotlin_launch_on_windows_gets_no_attach_listener() {
+        assert_eq!(kotlin_jvm_args(vec!["-Xmx1g".into()], true), vec!["-Xmx1g".to_string()]);
+    }

    #[test]
    fn reads_the_port_the_jvm_chose() {
@@ -1,5 +1,7 @@
 //! Protocol flow against an in-memory fake adapter.

+mod jvm_inspect;
+
 use super::*;
 use crate::adapter::{AdapterIo, LaunchFuture};
 use ide_core::IdeEvent;
@@ -120,6 +122,14 @@ struct Script {
    supports_terminate: bool,
    /// Answer `setBreakpoints` only once this many are waiting (a client that waits for each never gets one).
    hold_set_breakpoints: usize,
+    /// Send a `continued` event right after the first `stopped` one.
+    continued_after_stop: bool,
+    /// Close the connection when a `pause` request arrives (the adapter dies).
+    hang_up_on_pause: bool,
+    /// Answer `continue` with this body and keep the program running, instead of letting it exit.
+    continue_reply: Option<Value>,
+    /// Answer `initialize` only after this long (a slow adapter keeps the session connecting).
+    delay_initialize: Option<Duration>,
    initialized_delay: Duration,
    launch_delay: Duration,
    /// Leave startup pending before `initialized` (false) or after configuration (true).
@@ -144,6 +154,10 @@ impl Default for Script {
            output_after_terminate: false,
            supports_terminate: true,
            hold_set_breakpoints: 0,
+            continued_after_stop: false,
+            hang_up_on_pause: false,
+            continue_reply: None,
+            delay_initialize: None,
            initialized_delay: Duration::ZERO,
            launch_delay: Duration::ZERO,
            pending_startup: None,
@@ -210,6 +224,9 @@ async fn fake_adapter<S: AsyncRead + AsyncWrite + Unpin + Send + 'static>(stream
        }
        match msg["command"].as_str().unwrap_or_default() {
            "initialize" => {
+                if let Some(delay) = script.delay_initialize {
+                    tokio::time::sleep(delay).await;
+                }
                ok!(
                    msg,
                    json!({
@@ -341,6 +358,9 @@ async fn fake_adapter<S: AsyncRead + AsyncWrite + Unpin + Send + 'static>(stream
                send!(
                    json!({"type": "event", "event": "stopped", "body": {"reason": "breakpoint", "threadId": 1, "allThreadsStopped": true}})
                );
+                if script.continued_after_stop {
+                    send!(json!({"type": "event", "event": "continued", "body": {"threadId": 1}}));
+                }
                send!(
                    json!({"type": "event", "event": "breakpoint", "body": {"reason": "changed", "breakpoint": {"id": 1, "verified": false, "message": "moved", "line": 9, "source": {"path": script.source_path}}}})
                );
@@ -372,11 +392,13 @@ async fn fake_adapter<S: AsyncRead + AsyncWrite + Unpin + Send + 'static>(stream
                    {"label": ""},
                ]})
            ),
+            "continue" if script.continue_reply.is_some() => ok!(msg, script.continue_reply.clone().unwrap()),
            "continue" => {
                ok!(msg, json!({"allThreadsContinued": true}));
                send!(json!({"type": "event", "event": "exited", "body": {"exitCode": 3}}));
                send!(json!({"type": "event", "event": "terminated"}));
            }
+            "pause" if script.hang_up_on_pause => break,
            "stepIn" | "stepOut" | "pause" => ok!(msg, Value::Null),
            "redefineClasses" => {
                send!(
@@ -430,6 +452,8 @@ fn harness(script: Script) -> Harness {
        Arc::new(LanguageRegistry::from_configs(Vec::new())),
        Arc::new(FakeLauncher { log: log.clone(), script }),
    );
+    // A local attach must not send requests to the developer's own JVMs.
+    manager.set_jvm_finder(Arc::new(|_, _| None));
    Harness { manager, log, events, owner }
 }

@@ -1352,6 +1376,36 @@ async fn hot_swap_only_for_java_debug() {
    assert_eq!(failed.error_message.as_deref(), Some("boom"));
 }

+/// A loopback port nothing listens on until `listen` is called on the socket, and that no parallel
+/// test can bind meanwhile. A port picked by binding to 0 and closing the listener can be taken
+/// by another test before the code under test connects to it.
+fn reserved_port() -> (tokio::net::TcpSocket, u16) {
+    // Without SO_REUSEADDR, which every `TcpListener::bind` sets: no other socket may bind it.
+    let socket = tokio::net::TcpSocket::new_v4().unwrap();
+    socket.bind(([127, 0, 0, 1], 0).into()).unwrap();
+    let port = socket.local_addr().unwrap().port();
+    (socket, port)
+}
+
+/// An editor that starts listening on the reserved port after `delay` and serves one session.
+#[cfg(unix)]
+fn fake_editor(socket: tokio::net::TcpSocket, delay: Duration, log: Log) {
+    tokio::spawn(async move {
+        tokio::time::sleep(delay).await;
+        let listener = socket.listen(1).unwrap();
+        let (stream, _) = listener.accept().await.unwrap();
+        // Closed once its session is accepted, so a later start finds no editor open.
+        drop(listener);
+        fake_adapter(stream, log, Script::default()).await;
+    });
+}
+
+#[tokio::test]
+async fn a_reserved_port_cannot_be_taken_by_a_parallel_test() {
+    let (_socket, port) = reserved_port();
+    assert!(tokio::net::TcpListener::bind(("127.0.0.1", port)).await.is_err(), "port {port} is still free to bind");
+}
+
 fn godot_registry(port: u16, command: &str, args: &[&str]) -> Arc<LanguageRegistry> {
    let lang = ide_language::LanguageConfig {
        id: ide_core::LanguageId::new("gdscript"),
@@ -1419,7 +1473,7 @@ async fn godot_connects_to_the_editor_adapter() {

 #[tokio::test]
 async fn godot_without_editor_or_executable_says_what_to_do() {
-    let port = std::net::TcpListener::bind(("127.0.0.1", 0)).unwrap().local_addr().unwrap().port();
+    let (_no_editor, port) = reserved_port();
    let (owner, _events) = recording_owner("main");
    let manager = DebugManager::new(godot_registry(port, "openidea-no-such-godot", &[]));
    let err = manager.start(godot_launch(&std::env::temp_dir()), owner).await.unwrap_err();
@@ -1432,15 +1486,9 @@ async fn godot_without_editor_or_executable_says_what_to_do() {
 #[cfg(unix)]
 #[tokio::test]
 async fn godot_starts_a_headless_editor_when_none_is_open() {
-    let port = std::net::TcpListener::bind(("127.0.0.1", 0)).unwrap().local_addr().unwrap().port();
+    let (editor, port) = reserved_port();
    let log: Log = Arc::default();
-    let server_log = log.clone();
-    tokio::spawn(async move {
-        tokio::time::sleep(std::time::Duration::from_millis(300)).await;
-        let listener = tokio::net::TcpListener::bind(("127.0.0.1", port)).await.unwrap();
-        let (stream, _) = listener.accept().await.unwrap();
-        fake_adapter(stream, server_log, Script::default()).await;
-    });
+    fake_editor(editor, Duration::from_millis(300), log.clone());
    let (owner, _events) = recording_owner("main");
    // `sleep` stands in for `godot --headless --editor --path ${root} …`.
    let manager = DebugManager::new(godot_registry(port, "sleep", &["30"]));
@@ -1643,7 +1691,7 @@ async fn a_plugin_language_runs_its_debug_adapter_from_the_approved_program() {
    std::fs::write(&program, "#!/bin/sh\nexec sleep 30\n").unwrap();
    std::fs::set_permissions(&program, std::fs::Permissions::from_mode(0o755)).unwrap();

-    let port = std::net::TcpListener::bind(("127.0.0.1", 0)).unwrap().local_addr().unwrap().port();
+    let (editor, port) = reserved_port();
    let mut gdscript = parse_definition(include_str!("../../../languages/gdscript.toml")).unwrap();
    let debugger = gdscript.debugger.as_mut().unwrap();
    debugger.args.clear();
@@ -1660,13 +1708,7 @@ async fn a_plugin_language_runs_its_debug_adapter_from_the_approved_program() {
    let _language = store.add("acme.godot".into(), gdscript);

    let log: Log = Arc::default();
-    let server_log = log.clone();
-    tokio::spawn(async move {
-        tokio::time::sleep(std::time::Duration::from_millis(300)).await;
-        let listener = tokio::net::TcpListener::bind(("127.0.0.1", port)).await.unwrap();
-        let (stream, _) = listener.accept().await.unwrap();
-        fake_adapter(stream, server_log, Script::default()).await;
-    });
+    fake_editor(editor, Duration::from_millis(300), log.clone());
    let (owner, _events) = recording_owner("main");
    let manager = DebugManager::new(registry.clone());
    let id = manager.start(godot_launch(&dir), owner).await.unwrap();
@@ -2018,6 +2060,418 @@ fn delve_leftovers_are_removed_without_touching_other_files() {
    std::fs::remove_dir_all(&dir).unwrap();
 }

+fn is_suspended(h: &Harness, id: DebugSessionId) -> bool {
+    h.manager.session(id).unwrap().suspended.get()
+}
+
+#[tokio::test]
+async fn stopped_event_marks_session_suspended() {
+    let h = harness(Script::default());
+    let id = h.start(config()).await.unwrap();
+    h.wait_event(|e| matches!(e, DebugEvent::Stopped { .. })).await;
+    assert!(is_suspended(&h, id));
+}
+
+#[tokio::test]
+async fn resume_clears_suspended_without_continued_event() {
+    let h = harness(Script::default());
+    let id = h.start(config()).await.unwrap();
+    h.wait_event(|e| matches!(e, DebugEvent::Stopped { .. })).await;
+    // The fake adapter ends the session after `continue`, so keep the session alive here.
+    let session = h.manager.session(id).unwrap();
+    assert!(session.suspended.get());
+    h.manager.resume(id, 1).await.unwrap();
+    assert!(!session.suspended.get());
+}
+
+/// After the fake adapter's stop of every thread ends, threads 7 then 8 stop alone, as a
+/// breakpoint does on fwcd and java-debug (`SUSPEND_EVENT_THREAD`).
+async fn threads_stopped(script: Script) -> (Harness, Arc<Session>, DebugSessionId) {
+    let h = harness(script);
+    let id = h.start(config()).await.unwrap();
+    h.wait_event(|e| matches!(e, DebugEvent::Stopped { .. })).await;
+    let session = h.manager.session(id).unwrap();
+    session.suspended.continued(Some(1), true);
+    session.suspended.stopped(Some(7), false);
+    session.suspended.stopped(Some(8), false);
+    (h, session, id)
+}
+
+#[tokio::test]
+async fn resume_of_one_thread_keeps_the_others_suspended() {
+    let only_it = Script { continue_reply: Some(json!({"allThreadsContinued": false})), ..Script::default() };
+    let (h, session, id) = threads_stopped(only_it).await;
+    h.manager.resume(id, 7).await.unwrap();
+    assert!(session.suspended.get());
+    h.manager.resume(id, 8).await.unwrap();
+    assert!(!session.suspended.get());
+}
+
+#[tokio::test]
+async fn resume_reply_without_the_flag_continues_every_thread() {
+    let (h, session, id) = threads_stopped(Script { continue_reply: Some(json!({})), ..Script::default() }).await;
+    h.manager.resume(id, 7).await.unwrap();
+    assert!(!session.suspended.get());
+}
+
+#[tokio::test]
+async fn step_of_one_thread_keeps_the_others_suspended() {
+    let (h, session, id) = threads_stopped(Script::default()).await;
+    let ticket = session.suspended.ticket();
+    h.manager.step(id, 8, StepKind::Into).await.unwrap();
+    assert!(session.suspended.get());
+    // Thread 8 was the latest stop: what was read on it is stale.
+    assert_eq!(session.suspended.ticket(), ticket + 1);
+    h.manager.step_in_target(id, 7, 1).await.unwrap();
+    assert!(!session.suspended.get());
+}
+
+#[tokio::test]
+async fn continued_event_clears_suspended() {
+    let h = harness(Script { continued_after_stop: true, ..Script::default() });
+    let id = h.start(config()).await.unwrap();
+    h.wait_event(|e| matches!(e, DebugEvent::Continued { .. })).await;
+    assert!(!is_suspended(&h, id));
+}
+
+fn java_config(request: &str) -> LaunchConfig {
+    LaunchConfig { adapter: "java-debug".into(), request: request.into(), extra_json: None, ..config() }
+}
+
+fn remote_attach(adapter: &str, project: &std::path::Path) -> LaunchConfig {
+    LaunchConfig {
+        adapter: adapter.into(),
+        cwd: Some(project.to_owned().into()),
+        extra_json: Some(json!({"projectRoot": project, "hostName": "10.0.0.5", "port": 5005}).to_string()),
+        ..java_config("attach")
+    }
+}
+
+#[tokio::test]
+async fn attach_session_reports_capability() {
+    let h = harness(Script::default());
+    let project = tempfile::tempdir().unwrap();
+    for adapter in ["java-debug", kotlin::ADAPTER] {
+        let id = h.start(remote_attach(adapter, project.path())).await.unwrap();
+        let caps = h.manager.capabilities(id).unwrap();
+        assert!(caps.jvm_memory, "{adapter}");
+        assert_eq!(caps.jvm_heap, caps.jvm_memory, "{adapter}");
+    }
+
+    let id = h.start(config()).await.unwrap();
+    let caps = h.manager.capabilities(id).unwrap();
+    assert!(!caps.jvm_memory);
+    assert_eq!(caps.jvm_heap, caps.jvm_memory);
+    assert_eq!(
+        h.manager.jvm_memory_classes(id).await.unwrap(),
+        JvmMemoryResult::Unavailable { reason: JvmMemoryUnavailable::NoJvmProcess, can_count_when_paused: false }
+    );
+}
+
+#[cfg(unix)]
+#[tokio::test]
+async fn ended_attach_session_without_pid_reports_program_exited() {
+    let h = harness(Script::default());
+    let project = tempfile::tempdir().unwrap();
+    let id = h.start(remote_attach("java-debug", project.path())).await.unwrap();
+    h.manager.session(id).unwrap().terminated.store(true, Ordering::Release);
+    assert_eq!(
+        h.manager.jvm_memory_classes(id).await.unwrap(),
+        JvmMemoryResult::Unavailable { reason: JvmMemoryUnavailable::ProgramExited, can_count_when_paused: false }
+    );
+}
+
+#[cfg(unix)]
+#[tokio::test]
+async fn listen_mode_attach_is_not_scanned() {
+    let h = harness(Script::default());
+    let scans = Arc::new(Mutex::new(Vec::new()));
+    let counted = scans.clone();
+    h.manager.set_jvm_finder(Arc::new(move |port, _| {
+        counted.lock().push(port);
+        None
+    }));
+    let (_jvm, relay) = jdwp::tests::connected_jvm().await;
+    let mut cfg = java_config("attach");
+    cfg.extra_json = Some(json!({"hostName": "127.0.0.1", "port": relay}).to_string());
+    let id = h.start(cfg).await.unwrap();
+    assert_eq!(
+        h.manager.jvm_memory_classes(id).await.unwrap(),
+        JvmMemoryResult::Unavailable {
+            reason: JvmMemoryUnavailable::ListenModeNotSupported,
+            can_count_when_paused: false
+        }
+    );
+    tokio::time::sleep(Duration::from_millis(50)).await;
+    assert!(scans.lock().is_empty());
+}
+
+#[cfg(unix)]
+#[tokio::test]
+async fn remote_host_is_not_scanned() {
+    let h = harness(Script::default());
+    let project = tempfile::tempdir().unwrap();
+    let id = h.start(remote_attach("java-debug", project.path())).await.unwrap();
+    let session = h.manager.session(id).unwrap();
+    assert_eq!(session.jdwp_port.get(), None);
+    assert_eq!(
+        h.manager.jvm_memory_classes(id).await.unwrap(),
+        JvmMemoryResult::Unavailable {
+            reason: JvmMemoryUnavailable::AttachedJvmNotFound,
+            can_count_when_paused: false
+        }
+    );
+    assert_eq!(session.jvm_pid.load(Ordering::Acquire), 0);
+}
+
+/// The Memory view asks as soon as the session appears, while its JVM is still being looked up:
+/// answering then would claim there is no JVM to ask.
+#[cfg(unix)]
+#[tokio::test]
+async fn memory_asked_while_an_attach_connects_waits_for_its_jvm() {
+    let h = harness(Script { delay_initialize: Some(Duration::from_millis(300)), ..Script::default() });
+    let project = tempfile::tempdir().unwrap();
+    let (manager, owner) = (h.manager.clone(), h.owner.clone());
+    let config = remote_attach("java-debug", project.path());
+    let start = tokio::spawn(async move { manager.start(config, owner).await });
+    h.wait_event(|e| matches!(e, DebugEvent::Started { .. })).await;
+    let id = DebugSessionId(*event_sessions(&h.events).first().unwrap());
+    let session = h.manager.session(id).unwrap();
+    assert!(session.attach_target.get().is_none(), "asked too late to cover the connecting session");
+
+    h.manager.jvm_memory_classes(id).await.unwrap();
+    assert!(session.attach_target.get().is_some(), "answered before the attach was looked at");
+    start.await.unwrap().unwrap();
+}
+
+/// A JVM clears its JDWP listener address once a debugger connects, so only a scan made before
+/// the `attach` request can tell it by its port.
+#[cfg(unix)]
+#[tokio::test]
+async fn local_attach_finds_its_jvm_before_the_debugger_connects() {
+    let h = harness(Script::default());
+    let log = h.log.clone();
+    h.manager.set_jvm_finder(Arc::new(move |port, _| {
+        std::thread::sleep(Duration::from_millis(50));
+        let attached = log.lock().iter().any(|m| m["command"] == "attach");
+        (port == 5005 && !attached).then_some(4242)
+    }));
+    let mut cfg = java_config("attach");
+    cfg.extra_json = Some(json!({"hostName": "127.0.0.1", "port": 5005}).to_string());
+    let id = h.start(cfg).await.unwrap();
+    let session = h.manager.session(id).unwrap();
+    assert_eq!(session.jvm_pid.load(Ordering::Acquire), 4242);
+}
+
+#[cfg(unix)]
+#[tokio::test]
+async fn local_attach_remembers_its_jdwp_port() {
+    let h = harness(Script::default());
+    let ports = Arc::new(Mutex::new(Vec::new()));
+    let asked = ports.clone();
+    h.manager.set_jvm_finder(Arc::new(move |port, _| {
+        asked.lock().push(port);
+        Some(4242)
+    }));
+    let mut cfg = java_config("attach");
+    cfg.extra_json = Some(json!({"hostName": "localhost", "port": "5005"}).to_string());
+    let id = h.start(cfg).await.unwrap();
+    let session = h.manager.session(id).unwrap();
+    assert_eq!(session.jdwp_port.get(), Some(&5005));
+    assert_eq!(h.request("attach")["arguments"]["port"], "5005");
+    for _ in 0..200 {
+        if session.jvm_pid.load(Ordering::Acquire) != 0 {
+            break;
+        }
+        tokio::time::sleep(Duration::from_millis(10)).await;
+    }
+    assert_eq!(session.jvm_pid.load(Ordering::Acquire), 4242);
+    assert_eq!(ports.lock().first(), Some(&5005));
+}
+
+/// A fake `find_jvm_by_jdwp_port` for the JVM 4242, recording whether each scan ran while
+/// suspended. A JVM not `safe_while_suspended` (JDK below 17) is skipped by a suspended scan.
+#[cfg(unix)]
+fn counting_find(
+    safe_while_suspended: bool,
+) -> (Arc<Mutex<Vec<bool>>>, impl Fn(bool) -> Option<u32> + Clone + Send + 'static) {
+    let scans = Arc::new(Mutex::new(Vec::new()));
+    let counted = scans.clone();
+    (scans, move |suspended| {
+        counted.lock().push(suspended);
+        (!suspended || safe_while_suspended).then_some(4242)
+    })
+}
+
+#[cfg(unix)]
+async fn suspended_remote_attach(h: &Harness) -> Arc<Session> {
+    let project = tempfile::tempdir().unwrap();
+    let id = h.start(remote_attach("java-debug", project.path())).await.unwrap();
+    h.wait_event(|e| matches!(e, DebugEvent::Stopped { .. })).await;
+    let session = h.manager.session(id).unwrap();
+    assert!(session.suspended.get());
+    session
+}
+
+#[cfg(unix)]
+#[tokio::test]
+async fn background_scan_skips_an_old_jdk_while_suspended() {
+    let h = harness(Script::default());
+    let session = suspended_remote_attach(&h).await;
+    let (scans, find) = counting_find(false);
+    let task =
+        jvm_memory::watch_for_attached_jvm(&session, 5005, find, Duration::from_millis(5), Duration::from_secs(5));
+    tokio::time::sleep(Duration::from_millis(100)).await;
+    assert!(!scans.lock().is_empty() && scans.lock().iter().all(|&suspended| suspended));
+    assert_eq!(session.jvm_pid.load(Ordering::Acquire), 0);
+
+    session.suspended.continued(None, true);
+    tokio::time::timeout(Duration::from_secs(2), task).await.unwrap().unwrap();
+    assert_eq!(scans.lock().last(), Some(&false));
+    assert_eq!(session.jvm_pid.load(Ordering::Acquire), 4242);
+}
+
+#[cfg(unix)]
+#[tokio::test]
+async fn background_scan_finds_a_jdk_17_or_later_while_suspended() {
+    let h = harness(Script::default());
+    let session = suspended_remote_attach(&h).await;
+    let (scans, find) = counting_find(true);
+    let task =
+        jvm_memory::watch_for_attached_jvm(&session, 5005, find, Duration::from_millis(5), Duration::from_secs(5));
+    tokio::time::timeout(Duration::from_secs(2), task).await.unwrap().unwrap();
+    assert_eq!(*scans.lock(), [true]);
+    assert_eq!(session.jvm_pid.load(Ordering::Acquire), 4242);
+}
+
+#[cfg(unix)]
+#[tokio::test]
+async fn background_scan_keeps_a_known_pid_and_stops_with_the_session() {
+    let h = harness(Script::default());
+    let project = tempfile::tempdir().unwrap();
+    let id = h.start(remote_attach("java-debug", project.path())).await.unwrap();
+    h.wait_event(|e| matches!(e, DebugEvent::Stopped { .. })).await;
+    let session = h.manager.session(id).unwrap();
+    session.suspended.continued(None, true);
+    session.jvm_pid.store(7, Ordering::Release);
+    let watch = |find| {
+        jvm_memory::watch_for_attached_jvm(&session, 5005, find, Duration::from_millis(5), Duration::from_secs(5))
+    };
+    let (scans, find) = counting_find(true);
+    tokio::time::timeout(Duration::from_secs(2), watch(find)).await.unwrap().unwrap();
+    assert!(scans.lock().is_empty(), "a known pid needs no scan");
+    assert_eq!(session.jvm_pid.load(Ordering::Acquire), 7);
+
+    // A refresh stores the pid while this scan is still running.
+    session.jvm_pid.store(0, Ordering::Release);
+    let racing = session.clone();
+    let find = move |_| {
+        racing.jvm_pid.store(7, Ordering::Release);
+        Some(4242)
+    };
+    let task =
+        jvm_memory::watch_for_attached_jvm(&session, 5005, find, Duration::from_millis(5), Duration::from_secs(5));
+    tokio::time::timeout(Duration::from_secs(2), task).await.unwrap().unwrap();
+    assert_eq!(session.jvm_pid.load(Ordering::Acquire), 7, "a pid stored first is kept");
+
+    h.manager.stop(id).await.unwrap();
+    let (scans, find) = counting_find(true);
+    tokio::time::timeout(Duration::from_secs(2), watch(find)).await.unwrap().unwrap();
+    assert!(scans.lock().is_empty());
+}
+
+#[cfg(unix)]
+#[tokio::test]
+async fn unknown_pid_reports_no_jvm_process() {
+    let h = harness(Script::default());
+    let id = h.start(java_config("launch")).await.unwrap();
+    assert!(h.manager.capabilities(id).unwrap().jvm_memory);
+    assert_eq!(
+        h.manager.jvm_memory_classes(id).await.unwrap(),
+        JvmMemoryResult::Unavailable { reason: JvmMemoryUnavailable::NoJvmProcess, can_count_when_paused: false }
+    );
+}
+
+fn kotlin_attach(project: &std::path::Path, host: &str, port: u16) -> LaunchConfig {
+    LaunchConfig {
+        adapter: kotlin::ADAPTER.into(),
+        cwd: Some(project.to_owned().into()),
+        extra_json: Some(json!({"projectRoot": project, "hostName": host, "port": port}).to_string()),
+        ..java_config("attach")
+    }
+}
+
+#[tokio::test]
+async fn kotlin_session_installs_link_inspector() {
+    let h = harness(Script::default());
+    let project = tempfile::tempdir().unwrap();
+    let id = h.start(kotlin_attach(project.path(), "127.0.0.1", 5005)).await.unwrap();
+    let session = h.manager.session(id).unwrap();
+    assert!(session.jdwp_link.lock().is_some());
+    assert!(session.jvm_inspector.lock().is_some());
+}
+
+#[tokio::test]
+async fn attach_target_recorded_before_kotlin_rewrite() {
+    let h = harness(Script::default());
+    let project = tempfile::tempdir().unwrap();
+    let id = h.start(kotlin_attach(project.path(), "10.0.0.5", 5005)).await.unwrap();
+    let session = h.manager.session(id).unwrap();
+    assert_eq!(session.attach_target.get(), Some(&("10.0.0.5".to_owned(), 5005)));
+    let sent = &h.request("attach")["arguments"];
+    assert_eq!(sent["hostName"], "127.0.0.1");
+    let link_port = sent["port"].as_u64().unwrap();
+    assert!(link_port != 0 && link_port != 5005);
+}
+
+#[tokio::test]
+async fn java_attach_records_target() {
+    let h = harness(Script::default());
+    let mut cfg = java_config("attach");
+    cfg.extra_json = Some(json!({"hostName": "localhost", "port": 5005}).to_string());
+    let id = h.start(cfg).await.unwrap();
+    let session = h.manager.session(id).unwrap();
+    assert_eq!(session.attach_target.get(), Some(&("localhost".to_owned(), 5005)));
+    assert!(session.jdwp_link.lock().is_none());
+    assert!(session.jvm_inspector.lock().is_none());
+}
+
+#[tokio::test]
+async fn attach_target_recorded_for_any_adapter() {
+    let h = harness(Script::default());
+    let mut cfg = java_config("attach");
+    // "x" is not a registered adapter; debugpy has no JVM link or memory support.
+    cfg.adapter = "debugpy".into();
+    cfg.extra_json = Some(json!({"hostName": "h", "port": 7}).to_string());
+    let id = h.start(cfg).await.unwrap();
+    assert_eq!(h.manager.session(id).unwrap().attach_target.get(), Some(&("h".to_owned(), 7)));
+}
+
+#[tokio::test]
+async fn launch_sessions_have_no_attach_target() {
+    let h = harness(Script::default());
+    let id = h.start(java_config("launch")).await.unwrap();
+    assert_eq!(h.manager.session(id).unwrap().attach_target.get(), None);
+}
+
+#[tokio::test]
+async fn closing_the_session_drops_the_link() {
+    let h = harness(Script { hang_up_on_pause: true, ..Script::default() });
+    let project = tempfile::tempdir().unwrap();
+    let id = h.start(kotlin_attach(project.path(), "127.0.0.1", 5005)).await.unwrap();
+    let session = h.manager.session(id).unwrap();
+    assert!(session.jdwp_link.lock().is_some() && session.jvm_inspector.lock().is_some());
+    let _ = session.conn.request_timeout("pause", json!({"threadId": 1}), Duration::from_secs(2)).await;
+    for _ in 0..200 {
+        if h.manager.session(id).is_err() {
+            break;
+        }
+        tokio::time::sleep(Duration::from_millis(10)).await;
+    }
+    assert!(session.jdwp_link.lock().is_none());
+    assert!(session.jvm_inspector.lock().is_none());
+}
+
 #[cfg(unix)]
 #[tokio::test]
 async fn rdbg_drains_test_output_after_dap_closes_before_killing_the_process() {
@@ -0,0 +1,401 @@
+//! Instances, fields and referrers through a session's JVM inspector.
+
+use super::*;
+use crate::jvm_memory::test_support::FakeInspector;
+use ide_core::event::JvmBackendState;
+use ide_jvm_inspect::{InspectError, JvmValue, VmTarget};
+use tokio::io::AsyncReadExt;
+
+fn debug_events(h: &Harness) -> Vec<DebugEvent> {
+    h.events
+        .lock()
+        .iter()
+        .filter_map(|e| match e {
+            IdeEvent::Debug { event, .. } => Some(event.clone()),
+            _ => None,
+        })
+        .collect()
+}
+
+fn backend_events(h: &Harness) -> Vec<JvmBackendState> {
+    debug_events(h)
+        .into_iter()
+        .filter_map(|e| match e {
+            DebugEvent::JvmBackend { state } => Some(state),
+            _ => None,
+        })
+        .collect()
+}
+
+/// A java-debug session stopped at its breakpoint, inspected by `fake`.
+async fn stopped_with(h: &Harness, config: LaunchConfig, fake: &Arc<FakeInspector>) -> DebugSessionId {
+    let id = h.start(config).await.unwrap();
+    h.wait_event(|e| matches!(e, DebugEvent::Stopped { .. })).await;
+    h.manager.set_jvm_inspector(id, fake.clone()).unwrap();
+    id
+}
+
+fn object(id: u64) -> JvmValue {
+    JvmValue::Object { id, type_name: "A".into() }
+}
+
+const NOT_SUSPENDED: JvmInstancesResult =
+    JvmInstancesResult::Unavailable { reason: JvmInspectUnavailable::NotSuspended };
+
+#[tokio::test]
+async fn instances_need_a_suspended_session() {
+    let h = harness(Script { continued_after_stop: true, ..Script::default() });
+    let id = h.start(java_config("launch")).await.unwrap();
+    h.wait_event(|e| matches!(e, DebugEvent::Continued { .. })).await;
+    let fake = Arc::new(FakeInspector::default());
+    h.manager.set_jvm_inspector(id, fake.clone()).unwrap();
+    assert_eq!(h.manager.jvm_memory_instances(id, "A".into(), 10).await.unwrap(), NOT_SUSPENDED);
+    assert_eq!(fake.calls(), 0);
+}
+
+#[tokio::test]
+async fn instances_without_an_inspector_have_no_backend() {
+    let h = harness(Script::default());
+    let id = h.start(java_config("launch")).await.unwrap();
+    h.wait_event(|e| matches!(e, DebugEvent::Stopped { .. })).await;
+    assert_eq!(
+        h.manager.jvm_memory_instances(id, "A".into(), 10).await.unwrap(),
+        JvmInstancesResult::Unavailable { reason: JvmInspectUnavailable::NoBackend }
+    );
+}
+
+#[tokio::test]
+async fn object_from_previous_stop_is_stale() {
+    let h = harness(Script::default());
+    let fake = Arc::new(FakeInspector::default());
+    fake.instances.lock().push(object(7));
+    let id = stopped_with(&h, java_config("launch"), &fake).await;
+    let JvmInstancesResult::Ready { generation, instances, truncated } =
+        h.manager.jvm_memory_instances(id, "A".into(), 10).await.unwrap()
+    else {
+        panic!("instances unavailable");
+    };
+    assert_eq!((generation, instances, truncated), (1, vec![object(7)], false));
+
+    h.manager.session(id).unwrap().suspended.stopped(None, true);
+    let stale = JvmObjectRef { generation, id: 7 };
+    assert_eq!(
+        h.manager.jvm_memory_fields(id, stale).await.unwrap(),
+        JvmFieldsResult::Unavailable { reason: JvmInspectUnavailable::Stale }
+    );
+    assert_eq!(
+        h.manager.jvm_memory_referrers(id, stale, 10).await.unwrap(),
+        JvmReferrersResult::Unavailable { reason: JvmInspectUnavailable::Stale }
+    );
+    assert_eq!(fake.calls(), 1, "a stale object never reaches the inspector");
+    let current = JvmObjectRef { generation: 2, id: 7 };
+    assert_eq!(h.manager.jvm_memory_fields(id, current).await.unwrap(), JvmFieldsResult::Ready { fields: Vec::new() });
+}
+
+#[tokio::test]
+async fn instances_stay_readable_while_another_thread_is_stopped() {
+    let h = harness(Script::default());
+    let fake = Arc::new(FakeInspector::default());
+    fake.instances.lock().push(object(7));
+    let id = stopped_with(&h, java_config("launch"), &fake).await;
+    let session = h.manager.session(id).unwrap();
+    // Threads 20, 21, then 22 stop alone (SUSPEND_EVENT_THREAD).
+    session.suspended.continued(Some(1), true);
+    for thread in [20, 21, 22] {
+        session.suspended.stopped(Some(thread), false);
+    }
+    let instances = || h.manager.jvm_memory_instances(id, "A".into(), 10);
+    let result = instances().await.unwrap();
+    let JvmInstancesResult::Ready { generation, .. } = result else { panic!("{result:?}") };
+    let read_on_22 = JvmObjectRef { generation, id: 7 };
+
+    // An earlier thread resumes: the stop on 22 goes on, and so does what was read on it.
+    session.suspended.continued(Some(20), false);
+    assert_eq!(
+        h.manager.jvm_memory_fields(id, read_on_22).await.unwrap(),
+        JvmFieldsResult::Ready { fields: Vec::new() }
+    );
+
+    // The latest one resumes: 21 is still stopped, but what was read on 22 is stale.
+    session.suspended.continued(Some(22), false);
+    assert_eq!(
+        h.manager.jvm_memory_fields(id, read_on_22).await.unwrap(),
+        JvmFieldsResult::Unavailable { reason: JvmInspectUnavailable::Stale }
+    );
+    let result = instances().await.unwrap();
+    assert!(matches!(result, JvmInstancesResult::Ready { .. }), "{result:?}");
+}
+
+/// Starts `instances` on a held fake inspector; returns once the inspector was entered.
+async fn held_request(
+    h: &Harness,
+    id: DebugSessionId,
+    fake: &Arc<FakeInspector>,
+) -> tokio::task::JoinHandle<IdeResult<JvmInstancesResult>> {
+    let manager = h.manager.clone();
+    let request = tokio::spawn(async move { manager.jvm_memory_instances(id, "A".into(), 10).await });
+    tokio::time::timeout(Duration::from_secs(2), fake.entered.notified()).await.unwrap();
+    request
+}
+
+#[tokio::test]
+async fn stop_during_request_is_stale() {
+    let h = harness(Script::default());
+    let fake = Arc::new(FakeInspector { hold: true, ..FakeInspector::default() });
+    fake.instances.lock().push(object(7));
+    let id = stopped_with(&h, java_config("launch"), &fake).await;
+    let request = held_request(&h, id, &fake).await;
+    h.manager.session(id).unwrap().suspended.stopped(None, true);
+    fake.release.notify_one();
+    assert_eq!(
+        request.await.unwrap().unwrap(),
+        JvmInstancesResult::Unavailable { reason: JvmInspectUnavailable::Stale }
+    );
+}
+
+#[tokio::test]
+async fn resume_during_request_is_stale() {
+    let h = harness(Script::default());
+    let fake = Arc::new(FakeInspector { hold: true, ..FakeInspector::default() });
+    let id = stopped_with(&h, java_config("launch"), &fake).await;
+    let request = held_request(&h, id, &fake).await;
+    h.manager.session(id).unwrap().suspended.continued(None, true);
+    fake.release.notify_one();
+    assert_eq!(
+        request.await.unwrap().unwrap(),
+        JvmInstancesResult::Unavailable { reason: JvmInspectUnavailable::Stale }
+    );
+}
+
+#[tokio::test]
+async fn one_request_at_a_time_per_session() {
+    let h = harness(Script::default());
+    let fake = Arc::new(FakeInspector { hold: true, ..FakeInspector::default() });
+    let id = stopped_with(&h, java_config("launch"), &fake).await;
+    let first = held_request(&h, id, &fake).await;
+    let manager = h.manager.clone();
+    let second = tokio::spawn(async move { manager.jvm_memory_instances(id, "B".into(), 10).await });
+    tokio::time::sleep(Duration::from_millis(50)).await;
+    assert_eq!(fake.calls(), 1, "the second request waits for the first");
+    fake.release.notify_one();
+    first.await.unwrap().unwrap();
+    tokio::time::timeout(Duration::from_secs(2), fake.entered.notified()).await.unwrap();
+    fake.release.notify_one();
+    second.await.unwrap().unwrap();
+    assert_eq!(fake.calls(), 2);
+}
+
+#[tokio::test]
+async fn session_end_lets_go_of_a_waiting_inspector() {
+    let h = harness(Script::default());
+    let fake = Arc::new(FakeInspector { hold: true, ..FakeInspector::default() });
+    let id = stopped_with(&h, java_config("launch"), &fake).await;
+    let request = held_request(&h, id, &fake).await;
+    h.manager.session(id).unwrap().kill_processes();
+    assert_eq!(
+        tokio::time::timeout(Duration::from_secs(2), request).await.unwrap().unwrap().unwrap(),
+        JvmInstancesResult::Unavailable { reason: JvmInspectUnavailable::BackendClosed }
+    );
+    assert_eq!(Arc::strong_count(&fake), 1, "no clone of the inspector outlives the session");
+}
+
+#[tokio::test]
+async fn target_carries_pid_for_launch() {
+    let h = harness(Script { debuggee_pid: Some(4242), java_process_id: true, ..Script::default() });
+    let fake = Arc::new(FakeInspector::default());
+    let id = stopped_with(&h, java_config("launch"), &fake).await;
+    h.manager.jvm_memory_instances(id, "A".into(), 10).await.unwrap();
+    assert_eq!(
+        *fake.targets.lock(),
+        [VmTarget { pid: Some(4242), attach_host: None, attach_port: None, generation: 1 }]
+    );
+}
+
+#[tokio::test]
+async fn target_carries_attach_port_and_host() {
+    let h = harness(Script::default());
+    let project = tempfile::tempdir().unwrap();
+    let fake = Arc::new(FakeInspector::default());
+    let id = stopped_with(&h, remote_attach("java-debug", project.path()), &fake).await;
+    h.manager.jvm_memory_instances(id, "A".into(), 10).await.unwrap();
+    assert_eq!(
+        *fake.targets.lock(),
+        [VmTarget { pid: None, attach_host: Some("10.0.0.5".into()), attach_port: Some(5005), generation: 1 }]
+    );
+}
+
+#[tokio::test]
+async fn unsupported_backend_turns_instances_off() {
+    let h = harness(Script::default());
+    let fake = Arc::new(FakeInspector::failing(InspectError::Unsupported("x".into())));
+    let id = stopped_with(&h, java_config("launch"), &fake).await;
+    let before = debug_events(&h).len();
+    assert_eq!(
+        h.manager.jvm_memory_instances(id, "A".into(), 10).await.unwrap(),
+        JvmInstancesResult::Unavailable { reason: JvmInspectUnavailable::Unsupported { detail: "x".into() } }
+    );
+    let after: Vec<_> = debug_events(&h).split_off(before);
+    assert!(
+        matches!(
+            after.as_slice(),
+            [DebugEvent::JvmBackend { state: JvmBackendState::Unsupported { detail } }, DebugEvent::Capabilities]
+                if detail == "x"
+        ),
+        "{after:?}"
+    );
+    assert!(!h.manager.capabilities(id).unwrap().jvm_instances);
+    assert_eq!(
+        h.manager.jvm_memory_instances(id, "A".into(), 10).await.unwrap(),
+        JvmInstancesResult::Unavailable { reason: JvmInspectUnavailable::NoBackend }
+    );
+    assert_eq!(fake.calls(), 1);
+}
+
+/// `error` is reported but leaves the backend on.
+async fn keeps_the_backend(error: InspectError, reason: JvmInspectUnavailable) {
+    let h = harness(Script::default());
+    let fake = Arc::new(FakeInspector::failing(error));
+    let id = stopped_with(&h, java_config("launch"), &fake).await;
+    let object = JvmObjectRef { generation: 1, id: 7 };
+    assert_eq!(h.manager.jvm_memory_fields(id, object).await.unwrap(), JvmFieldsResult::Unavailable { reason });
+    assert!(h.manager.capabilities(id).unwrap().jvm_instances);
+    assert!(backend_events(&h).is_empty());
+    assert!(h.manager.session(id).unwrap().jvm_inspector.lock().is_some());
+}
+
+#[tokio::test]
+async fn collected_keeps_the_backend() {
+    keeps_the_backend(InspectError::Collected, JvmInspectUnavailable::Collected).await;
+}
+
+#[tokio::test]
+async fn timed_out_keeps_the_backend() {
+    keeps_the_backend(InspectError::TimedOut, JvmInspectUnavailable::TimedOut).await;
+}
+
+#[tokio::test]
+async fn failed_inspection_is_an_error() {
+    let h = harness(Script::default());
+    let fake = Arc::new(FakeInspector::failing(InspectError::Failed("boom".into())));
+    let id = stopped_with(&h, java_config("launch"), &fake).await;
+    let err = h.manager.jvm_memory_instances(id, "A".into(), 10).await.unwrap_err();
+    assert!(err.to_string().contains("boom"), "{err}");
+    assert!(h.manager.capabilities(id).unwrap().jvm_instances);
+}
+
+#[tokio::test]
+async fn limits_are_clamped() {
+    let h = harness(Script::default());
+    let fake = Arc::new(FakeInspector::default());
+    let id = stopped_with(&h, java_config("launch"), &fake).await;
+    for limit in [5000, 0] {
+        h.manager.jvm_memory_instances(id, "A".into(), limit).await.unwrap();
+    }
+    let object = JvmObjectRef { generation: 1, id: 7 };
+    for limit in [0, 900, 20] {
+        h.manager.jvm_memory_referrers(id, object, limit).await.unwrap();
+    }
+    assert_eq!(*fake.limits.lock(), [1000, 1, 50, 200, 20]);
+}
+
+#[tokio::test]
+async fn a_full_page_is_truncated() {
+    let h = harness(Script::default());
+    let fake = Arc::new(FakeInspector::default());
+    fake.instances.lock().extend([object(1), object(2)]);
+    let id = stopped_with(&h, java_config("launch"), &fake).await;
+    let truncated = |r: JvmInstancesResult| match r {
+        JvmInstancesResult::Ready { truncated, .. } => truncated,
+        other => panic!("{other:?}"),
+    };
+    assert!(truncated(h.manager.jvm_memory_instances(id, "A".into(), 2).await.unwrap()));
+    assert!(!truncated(h.manager.jvm_memory_instances(id, "A".into(), 3).await.unwrap()));
+}
+
+#[tokio::test]
+async fn capability_follows_the_inspector() {
+    let h = harness(Script::default());
+    let id = h.start(java_config("launch")).await.unwrap();
+    assert!(!h.manager.capabilities(id).unwrap().jvm_instances);
+    let before = debug_events(&h).len();
+    h.manager.set_jvm_inspector(id, Arc::new(FakeInspector::default())).unwrap();
+    assert!(h.manager.capabilities(id).unwrap().jvm_instances);
+    assert!(matches!(debug_events(&h).split_off(before).as_slice(), [DebugEvent::Capabilities]));
+}
+
+#[tokio::test]
+async fn remote_attach_counts_classes_through_the_debugger() {
+    let h = harness(Script::default());
+    let project = tempfile::tempdir().unwrap();
+    let fake = Arc::new(FakeInspector::default());
+    fake.counts.lock().push(ide_jvm_inspect::ClassCount { name: "A".into(), instances: 3 });
+    let id = stopped_with(&h, remote_attach("java-debug", project.path()), &fake).await;
+    let result = h.manager.jvm_memory_classes(id).await.unwrap();
+    assert!(
+        matches!(&result, JvmMemoryResult::Ready { snapshot }
+            if snapshot.source == JvmMemorySource::Debugger && snapshot.rows.len() == 1),
+        "{result:?}"
+    );
+    assert_eq!(fake.targets.lock()[0].attach_port, Some(5005));
+}
+
+const HANDSHAKE: &[u8; 14] = b"JDWP-Handshake";
+
+/// A Kotlin attach whose link reaches a fake JVM: (adapter end, JVM end) after both handshakes.
+async fn linked_kotlin(h: &Harness) -> (DebugSessionId, tokio::net::TcpStream, tokio::net::TcpStream) {
+    let jvm = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let project = tempfile::tempdir().unwrap();
+    let id = h.start(kotlin_attach(project.path(), "127.0.0.1", jvm.local_addr().unwrap().port())).await.unwrap();
+    let link_port = h.request("attach")["arguments"]["port"].as_u64().unwrap() as u16;
+    let mut adapter = tokio::net::TcpStream::connect(("127.0.0.1", link_port)).await.unwrap();
+    adapter.write_all(HANDSHAKE).await.unwrap();
+    let (mut jvm, _) = tokio::time::timeout(Duration::from_secs(2), jvm.accept()).await.unwrap().unwrap();
+    let mut buf = [0; 14];
+    jvm.read_exact(&mut buf).await.unwrap();
+    jvm.write_all(HANDSHAKE).await.unwrap();
+    adapter.read_exact(&mut buf).await.unwrap();
+    (id, adapter, jvm)
+}
+
+/// A command the adapter sends with an id from the link's own range: the link then only forwards.
+async fn degrade(adapter: &mut tokio::net::TcpStream) {
+    let mut packet = 11u32.to_be_bytes().to_vec();
+    packet.extend(ide_jdwp::OWN_IDS.start.to_be_bytes());
+    packet.extend([0, 1, 1]);
+    adapter.write_all(&packet).await.unwrap();
+}
+
+async fn wait_backend(h: &Harness, want: impl Fn(&JvmBackendState) -> bool) {
+    h.wait_event(|e| matches!(e, DebugEvent::JvmBackend { state } if want(state))).await;
+}
+
+#[tokio::test]
+async fn kotlin_link_reports_degraded_then_closed() {
+    let h = harness(Script::default());
+    let (id, mut adapter, jvm) = linked_kotlin(&h).await;
+    degrade(&mut adapter).await;
+    wait_backend(&h, |s| matches!(s, JvmBackendState::Degraded { .. })).await;
+    drop(jvm);
+    wait_backend(&h, |s| *s == JvmBackendState::Closed).await;
+    assert_eq!(backend_events(&h).len(), 2, "{:?}", backend_events(&h));
+    let session = h.manager.session(id).unwrap();
+    assert!(session.jdwp_link.lock().is_some(), "a closed link is kept until the session ends");
+}
+
+#[tokio::test]
+async fn kotlin_link_is_quiet_once_the_program_ended() {
+    let h = harness(Script::default());
+    let (id, _adapter, jvm) = linked_kotlin(&h).await;
+    let session = h.manager.session(id).unwrap();
+    // The adapter reported the end; the session keeps its link until it shuts down.
+    session.terminated.store(true, Ordering::Release);
+    drop(jvm);
+    for _ in 0..100 {
+        if session.jdwp_link.lock().as_ref().is_some_and(|l| l.state() == ide_jdwp::LinkState::Closed) {
+            break;
+        }
+        tokio::time::sleep(Duration::from_millis(10)).await;
+    }
+    tokio::time::sleep(Duration::from_millis(50)).await;
+    assert!(backend_events(&h).is_empty(), "{:?}", backend_events(&h));
+}
@@ -136,24 +136,19 @@ pub struct DockerRegistry {
    pub username: String,
 }

-#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, Type)]
+#[derive(Clone, Debug, Default, PartialEq, Eq, Serialize, Deserialize, Type)]
 #[serde(rename_all = "camelCase")]
 pub struct DockerSettings {
-    #[serde(default = "default_connections")]
+    #[serde(default)]
    pub connections: Vec<DockerConnection>,
    #[serde(default)]
    pub registries: Vec<DockerRegistry>,
 }

-impl Default for DockerSettings {
-    fn default() -> Self {
-        Self { connections: default_connections(), registries: Vec::new() }
-    }
-}
-
-/// Before anything is configured the CLI's own daemon is shown, as before connections existed.
-fn default_connections() -> Vec<DockerConnection> {
-    vec![DockerConnection { id: "default".into(), name: "Docker".into(), endpoint: DockerEndpoint::Default }]
+/// Whether `c` is the connection older versions injected when nothing was configured:
+/// Services lists only connections the user added, so this one is dropped on load.
+fn is_injected_default(c: &DockerConnection) -> bool {
+    c.id == "default" && c.name == "Docker" && c.endpoint == DockerEndpoint::Default
 }

 const SETTINGS_FILE: &str = "docker.json";
@@ -161,10 +156,17 @@ const PASSWORDS_FILE: &str = "docker-passwords.json";

 pub fn load_settings(data_dir: &Path) -> DockerSettings {
    match std::fs::read(data_dir.join(SETTINGS_FILE)) {
-        Ok(bytes) => serde_json::from_slice(&bytes).unwrap_or_else(|e| {
-            tracing::warn!("{SETTINGS_FILE}: {e}");
-            DockerSettings::default()
-        }),
+        // A file written before the opt-in Services rework carries the injected default
+        // (saved along with a registry or another connection); drop it unless edited.
+        Ok(bytes) => serde_json::from_slice(&bytes)
+            .map(|mut s: DockerSettings| {
+                s.connections.retain(|c| !is_injected_default(c));
+                s
+            })
+            .unwrap_or_else(|e| {
+                tracing::warn!("{SETTINGS_FILE}: {e}");
+                DockerSettings::default()
+            }),
        Err(_) => DockerSettings::default(),
    }
 }
@@ -373,7 +375,8 @@ mod tests {
    fn settings_round_trip_and_passwords() {
        let tmp = tempfile::tempdir().unwrap();
        let dir = tmp.path();
-        assert_eq!(load_settings(dir).connections[0].endpoint, DockerEndpoint::Default);
+        // Nothing configured: no connection is invented.
+        assert!(load_settings(dir).connections.is_empty());
        let settings = DockerSettings {
            connections: vec![DockerConnection {
                id: "c1".into(),
@@ -395,8 +398,32 @@ mod tests {
        assert!(auth_for(dir, "postgres").is_none());
        save_settings(dir, &DockerSettings { registries: Vec::new(), ..settings }).unwrap();
        assert_eq!(registry_password(dir, "r1"), None);
-        // An empty list stays empty (the default connection is only for a missing file).
+        // An empty list stays empty.
        save_settings(dir, &DockerSettings { connections: Vec::new(), registries: Vec::new() }).unwrap();
        assert!(load_settings(dir).connections.is_empty());
    }
+
+    #[test]
+    fn load_drops_the_injected_default_connection() {
+        let tmp = tempfile::tempdir().unwrap();
+        let dir = tmp.path();
+        let injected =
+            || DockerConnection { id: "default".into(), name: "Docker".into(), endpoint: DockerEndpoint::Default };
+        let registry =
+            DockerRegistry { id: "r1".into(), name: "GHCR".into(), address: "ghcr.io".into(), username: "me".into() };
+        // Saved untouched by an older version (e.g. along with a registry): dropped on load.
+        save_settings(dir, &DockerSettings { connections: vec![injected()], registries: vec![registry.clone()] })
+            .unwrap();
+        let loaded = load_settings(dir);
+        assert!(loaded.connections.is_empty());
+        assert_eq!(loaded.registries, vec![registry]);
+        // A connection the user renamed or re-pointed stays theirs.
+        let renamed = DockerConnection { name: "Local engine".into(), ..injected() };
+        save_settings(dir, &DockerSettings { connections: vec![renamed.clone()], registries: Vec::new() }).unwrap();
+        assert_eq!(load_settings(dir).connections, vec![renamed]);
+        let repointed =
+            DockerConnection { endpoint: DockerEndpoint::Socket { path: "/var/run/docker.sock".into() }, ..injected() };
+        save_settings(dir, &DockerSettings { connections: vec![repointed.clone()], registries: Vec::new() }).unwrap();
+        assert_eq!(load_settings(dir).connections, vec![repointed]);
+    }
 }
@@ -71,7 +71,6 @@ impl Capability {
        match self {
            Self::ProjectRead | Self::EditorRead | Self::Storage | Self::Events | Self::VcsRead => Risk::Low,
            Self::EditorWrite
-            | Self::ProjectWrite
            | Self::DatabaseSources
            | Self::VcsWrite
            | Self::Net(_)
@@ -80,7 +79,14 @@ impl Capability {
            | Self::Secrets
            | Self::InputEvents
            | Self::Clipboard => Risk::Medium,
-            Self::IdeCommands | Self::Ipc(_) | Self::Process(_) | Self::FsRead(_) | Self::FsWrite(_) => Risk::High,
+            // A trusted project's own files name the programs the IDE starts (its settings, its
+            // build scripts), so writing them is running code.
+            Self::ProjectWrite
+            | Self::IdeCommands
+            | Self::Ipc(_)
+            | Self::Process(_)
+            | Self::FsRead(_)
+            | Self::FsWrite(_) => Risk::High,
        }
    }

@@ -402,6 +408,15 @@ mod tests {
        assert!(Capability::Storage.auto_migrates() && !Capability::EditorWrite.auto_migrates());
    }

+    /// A trusted project's files name programs the IDE starts by itself: the git and language
+    /// server executables in `.openidea/settings.toml`, its build scripts. Writing them is
+    /// running code, and the install dialog says so.
+    #[test]
+    fn writing_project_files_is_a_high_risk_grant() {
+        assert_eq!(Capability::ProjectWrite.risk(), Risk::High);
+        assert!(!Capability::ProjectWrite.auto_migrates());
+    }
+
    #[test]
    fn host_patterns() {
        assert!(host_matches("crates.io", "crates.io"));
@@ -352,6 +352,22 @@ pub struct IndexEntry {
    /// the IDE before showing it.
    #[serde(default)]
    pub icon: Option<PluginIcon>,
+    /// Where to read more: the manifest's homepage, else the registry's source repository.
+    #[serde(default)]
+    pub homepage: Option<String>,
+    /// This version's release notes (the manifest's `change-notes`).
+    #[serde(default)]
+    pub notes: Option<String>,
+    /// Downloads of the plugin's packages, as counted by whoever serves them.
+    #[serde(default)]
+    pub downloads: Option<u64>,
+    /// Picked by the marketplace's maintainers.
+    #[serde(default)]
+    pub featured: bool,
+    /// Project-relative paths whose presence makes the plugin worth suggesting
+    /// (`pom.xml`, `config/routes.rb`).
+    #[serde(default)]
+    pub suggest: Vec<String>,
 }

 impl IndexEntry {
@@ -389,10 +405,31 @@ impl Index {
    /// revocations, publisher keys (revoked ones) and `moved` — and lists no
    /// plugins ([`Index::is_newer`]): the IDE needs updating to install from it.
    pub fn parse(text: &str) -> Result<Self, String> {
+        let (index, dropped) = Self::parse_dropping(text)?;
+        if let Some(first) = dropped.first() {
+            // One line per index, and a bounded one: whoever wrote the index chose the text.
+            let id: String = first.id.chars().take(80).collect();
+            tracing::warn!(
+                "index: {} entries are not offered, their id or version is not one a manifest may carry (first: {id:?})",
+                dropped.len()
+            );
+        }
+        Ok(index)
+    }
+
+    /// [`Index::parse`], with the entries it does not offer: those whose id or
+    /// version no manifest may carry, so no IDE could install them.
+    pub fn parse_dropping(text: &str) -> Result<(Self, Vec<IndexEntry>), String> {
        let value: serde_json::Value = serde_json::from_str(text).map_err(|e| format!("index: {e}"))?;
        let format = value.get("format").and_then(serde_json::Value::as_u64).ok_or("index: `format` is missing")?;
        if format == u64::from(INDEX_FORMAT) {
-            return serde_json::from_value(value).map_err(|e| format!("index: {e}"));
+            let mut index: Index = serde_json::from_value(value).map_err(|e| format!("index: {e}"))?;
+            // What no manifest may carry is not offered: the id and version also name files on disk.
+            let (offered, dropped) = std::mem::take(&mut index.plugins)
+                .into_iter()
+                .partition(|e| crate::validate_plugin_id(&e.id).is_ok() && e.semver().is_some());
+            index.plugins = offered;
+            return Ok((index, dropped));
        }
        if format < u64::from(INDEX_FORMAT) {
            return Err(format!("index format {format} is not supported (this IDE reads {INDEX_FORMAT})"));
@@ -401,13 +438,14 @@ impl Index {
            let items = value.get(field).and_then(serde_json::Value::as_array);
            items.into_iter().flatten().filter_map(|v| serde_json::from_value(v.clone()).ok()).collect()
        }
-        Ok(Index {
+        let index = Index {
            format: u32::try_from(format).unwrap_or(u32::MAX),
            publishers: each(&value, "publishers"),
            plugins: Vec::new(),
            revoked: each(&value, "revoked"),
            moved: value.get("moved").and_then(serde_json::Value::as_str).map(str::to_owned),
-        })
+        };
+        Ok((index, Vec::new()))
    }

    /// Written in a format newer than this IDE reads: it lists no plugins.
@@ -685,9 +723,33 @@ mod tests {
            engine: "^1.0".into(),
            categories: Vec::new(),
            icon: None,
+            homepage: None,
+            notes: None,
+            downloads: None,
+            featured: false,
+            suggest: Vec::new(),
        }
    }

+    #[test]
+    fn listing_details_are_optional() {
+        let mut json = serde_json::to_value(entry("acme.a", "acme")).unwrap();
+        for key in ["homepage", "notes", "downloads", "featured", "suggest"] {
+            json.as_object_mut().unwrap().remove(key);
+        }
+        assert_eq!(serde_json::from_value::<IndexEntry>(json).unwrap(), entry("acme.a", "acme"));
+        let full = IndexEntry {
+            homepage: Some("https://example.com".into()),
+            notes: Some("Fixes".into()),
+            downloads: Some(42),
+            featured: true,
+            suggest: vec!["pom.xml".into()],
+            ..entry("acme.a", "acme")
+        };
+        let back: IndexEntry = serde_json::from_str(&serde_json::to_string(&full).unwrap()).unwrap();
+        assert_eq!(back, full);
+    }
+
    fn version(id: &str, version: &str, engine: &str) -> IndexEntry {
        IndexEntry { version: version.into(), engine: engine.into(), ..entry(id, "acme") }
    }
@@ -868,6 +930,51 @@ mod tests {
        }
    }

+    #[test]
+    fn an_index_offers_only_the_ids_and_versions_a_manifest_may_carry() {
+        let invalid_ids = [
+            "../x",
+            "../../escape",
+            "acme/../x",
+            r"acme.\..\x",
+            "acme./../../../escape",
+            "/abs",
+            r"C:\x",
+            "ACME.x",
+            "acme.",
+            "acme..x",
+            "acme.x y",
+            "",
+        ];
+        let invalid_versions =
+            ["../1.0.0", "1.0.0/../../../escape", r"1.0.0\..\x", "1.0.0-..", " 1.0.0", "1.0", "latest", ""];
+        let valid = [("acme.x", "1.0.0"), ("acme.tools-2_b", "1.0.0-rc.1+b.2")];
+        let mut listed: Vec<(&str, &str)> = valid.to_vec();
+        listed.extend(invalid_ids.iter().map(|id| (*id, "1.0.0")));
+        listed.extend(invalid_versions.iter().map(|version| ("acme.x", *version)));
+        let plugins: Vec<serde_json::Value> = listed
+            .iter()
+            .map(|(id, version)| {
+                serde_json::json!({
+                    "id": id, "name": "X", "version": version, "publisher": "acme",
+                    "download": "x.oiplugin", "digest": "d", "engine": "^1.0",
+                })
+            })
+            .collect();
+        let text = serde_json::json!({ "format": INDEX_FORMAT, "plugins": plugins }).to_string();
+        let pairs = |entries: &[IndexEntry]| -> Vec<(String, String)> {
+            entries.iter().map(|e| (e.id.clone(), e.version.clone())).collect()
+        };
+        let owned = |pairs: &[(&str, &str)]| -> Vec<(String, String)> {
+            pairs.iter().map(|(id, version)| ((*id).to_owned(), (*version).to_owned())).collect()
+        };
+
+        assert_eq!(pairs(&Index::parse(&text).unwrap().plugins), owned(&valid));
+        let (index, dropped) = Index::parse_dropping(&text).unwrap();
+        assert_eq!(pairs(&index.plugins), owned(&valid));
+        assert_eq!(pairs(&dropped), owned(&listed[valid.len()..]));
+    }
+
    #[test]
    fn a_newer_index_still_revokes_and_moves() {
        let newer = format!(
@@ -102,6 +102,14 @@ fn parse<T>(r: serde_json::Result<T>) -> IdeResult<T> {
    r.map_err(|e| IdeError::new(ide_core::error::ErrorKind::Protocol, format!("unexpected gh output: {e}")))
 }

+/// The repository of `root`, free to execute: the host runs the `gh_*` commands only in a
+/// trusted project.
+fn trusted_repo(root: &Path) -> IdeResult<Repo> {
+    let repo = Repo::discover(root)?.ok_or_else(|| IdeError::invalid("not a git repository"))?;
+    repo.set_execution_allowed(true);
+    Ok(repo)
+}
+
 /// Whether gh is installed and signed in, and which GitHub repository `root` maps to.
 pub fn availability(root: &Path) -> GhAvailability {
    if gh_executable().is_none() {
@@ -233,7 +241,7 @@ pub fn create(
        return Err(IdeError::invalid("choose a base branch"));
    }
    if req.push {
-        let repo = Repo::discover(root)?.ok_or_else(|| IdeError::invalid("not a git repository"))?;
+        let repo = trusted_repo(root)?;
        repo.push(false, true, cancel, on_output)?;
    }
    let mut args = vec!["pr", "create", "--title", req.title.trim(), "--body", &req.body, "--base", req.base.trim()];
@@ -354,7 +362,7 @@ pub fn prepare_diff(
    cancel: &CancellationToken,
    on_output: &dyn Fn(String),
 ) -> IdeResult<PrDiffRevisions> {
-    let repo = Repo::discover(root)?.ok_or_else(|| IdeError::invalid("not a git repository"))?;
+    let repo = trusted_repo(root)?;
    if !repo.has_commit(head_oid) || !repo.has_commit(base_oid) {
        let remotes = repo.remotes()?;
        let remote = remote_for(&remotes, name_with_owner)
@@ -0,0 +1,28 @@
+[package]
+name = "ide-heap"
+description = "HotSpot .hprof parser and heap-snapshot analysis for the debugger's JVM Memory view"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+ide-core.workspace = true
+memmap2.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+specta.workspace = true
+thiserror.workspace = true
+tracing.workspace = true
+uuid.workspace = true
+
+[target.'cfg(unix)'.dependencies]
+libc = "0.2"
+
+[dev-dependencies]
+criterion = { version = "0.5", default-features = false }
+proptest.workspace = true
+tempfile.workspace = true
+
+[[bench]]
+name = "heap"
+harness = false
@@ -0,0 +1,100 @@
+//! `cargo bench -p ide-heap`: whole-dump analysis and the dominator tree at a million objects.
+
+use criterion::{Criterion, black_box, criterion_group, criterion_main};
+use ide_core::CancellationToken;
+use ide_heap::testing::{MiniHeap, XorShift};
+use ide_heap::{AnalyzeOptions, BasicType, HeapError, HeapPhase, HeapRootKind, Store, analyze, dominators};
+
+const CHAINS: u64 = 100_000;
+const CHAIN_LEN: u64 = 5;
+const NODES: u64 = CHAINS * CHAIN_LEN;
+/// Every object takes one slot: node `k` at slot `2k`, its payload at `2k + 1`.
+const BASE: u64 = 0x1_0000_0000;
+const SLOT: u64 = 0x20;
+
+fn node_addr(k: u64) -> u64 {
+    BASE + 2 * k * SLOT
+}
+
+fn payload_addr(k: u64) -> u64 {
+    node_addr(k) + SLOT
+}
+
+/// 100k rooted chains of five `Node { next, payload }`, each node with its own `byte[16]`. Every
+/// seventh node points at the payload of the same position in the next chain instead, so that
+/// payload has two owners and its own one is garbage: 1M objects in all.
+fn million_objects() -> MiniHeap {
+    let mut heap = MiniHeap::new();
+    heap.class("Node", Some("java.lang.Object"), &[("next", BasicType::Object), ("payload", BasicType::Object)]);
+    for k in 0..NODES {
+        let next = if (k + 1) % CHAIN_LEN == 0 { 0 } else { node_addr(k + 1) };
+        let payload = if k % 7 == 0 { payload_addr((k + CHAIN_LEN) % NODES) } else { payload_addr(k) };
+        heap.object(node_addr(k), "Node", &[("next", next), ("payload", payload)]);
+        heap.prim(payload_addr(k), BasicType::Byte, 16);
+        if k % CHAIN_LEN == 0 {
+            heap.root(HeapRootKind::JniGlobal, node_addr(k));
+        }
+    }
+    heap
+}
+
+/// CSR of a random tree on `nodes` nodes rooted at 0 plus `extra` random edges, targets of each
+/// node in scrambled order.
+fn random_graph(nodes: u32, extra: u32, seed: u64) -> (Vec<u32>, Vec<u32>) {
+    let mut rng = XorShift::new(seed);
+    let mut edges: Vec<(u32, u32)> = (1..nodes).map(|v| (rng.below(v), v)).collect();
+    edges.extend((0..extra).map(|_| (rng.below(nodes), rng.below(nodes))));
+    for i in (1..edges.len()).rev() {
+        edges.swap(i, rng.below(i as u32 + 1) as usize);
+    }
+    let mut start = vec![0u32; nodes as usize + 1];
+    for &(from, _) in &edges {
+        start[from as usize + 1] += 1;
+    }
+    for v in 0..nodes as usize {
+        start[v + 1] += start[v];
+    }
+    let mut cursor = start.clone();
+    let mut out = vec![0u32; edges.len()];
+    for &(from, to) in &edges {
+        out[cursor[from as usize] as usize] = to;
+        cursor[from as usize] += 1;
+    }
+    (start, out)
+}
+
+fn no_progress(_: HeapPhase, _: u64, _: u64) -> Result<(), HeapError> {
+    Ok(())
+}
+
+fn bench(c: &mut Criterion) {
+    let dir = tempfile::tempdir().expect("temp dir");
+    let hprof = dir.path().join("dump.hprof");
+    million_objects().write_to(&hprof);
+    let (start, out) = random_graph(1_000_000, 500_000, 0x5EED_D0D0);
+    let cancel = CancellationToken::new();
+    let opts = AnalyzeOptions { in_memory: true, ..AnalyzeOptions::default() };
+    let summary = analyze(&hprof, dir.path(), opts, &no_progress, &cancel).expect("analyze").summary;
+    assert_eq!(summary.objects, 2 * NODES, "{summary:?}");
+    assert_eq!(summary.unreachable_objects, NODES.div_ceil(7), "{summary:?}");
+
+    let mut g = c.benchmark_group("heap");
+    g.sample_size(10);
+    g.bench_function("index_1m_objects", |b| {
+        b.iter(|| {
+            let opts = AnalyzeOptions { in_memory: true, ..AnalyzeOptions::default() };
+            let index = analyze(&hprof, dir.path(), opts, &no_progress, &cancel).expect("analyze");
+            black_box(index.summary.objects)
+        })
+    });
+    g.bench_function("dominators_1m_random", |b| {
+        b.iter(|| {
+            let d = dominators(&start, &out, 0, &Store::Ram, &no_progress, &cancel).expect("dominators");
+            black_box(d.reachable)
+        })
+    });
+    g.finish();
+}
+
+criterion_group!(benches, bench);
+criterion_main!(benches);
@@ -0,0 +1,202 @@
+//! Ties the passes together: graph, dominators, shallow and retained sizes, the class table and
+//! paths to GC roots, into a [`HeapIndex`].
+
+mod classes;
+mod paths;
+mod retained;
+
+use std::collections::HashMap;
+use std::fs::File;
+use std::path::Path;
+use std::time::Instant;
+
+use ide_core::CancellationToken;
+use memmap2::Mmap;
+
+use crate::column::Store;
+use crate::dominators::{Dominators, dominators};
+use crate::error::HeapError;
+use crate::files::INDEX_FILES;
+use crate::graph::{HeapGraph, build_graph};
+use crate::index::{HeapIndex, HeapSummary, write_summary};
+use crate::progress::{CHECK_EVERY, HeapPhase, Progress};
+use crate::shallow::{HistogramEntry, ShallowModel, VmLayout};
+
+#[derive(Clone, Debug, Default)]
+pub struct AnalyzeOptions {
+    /// The JVM's class histogram from the same stop as the dump.
+    pub histogram: Option<Vec<HistogramEntry>>,
+    /// `VM.flags` output of the JVM that wrote the dump.
+    pub vm_flags: Option<String>,
+    pub java_major: Option<u32>,
+    /// Keep every array in RAM and write nothing to `index_dir`.
+    pub in_memory: bool,
+}
+
+/// Analyzes the dump at `hprof`, writing the index columns and `summary.json` into `index_dir`
+/// (which must exist and may hold the dump itself). `progress` gets every [`HeapPhase`] in order;
+/// an `Err` from it, a cancel or any other error removes every index file from `index_dir` and
+/// leaves the rest of the directory alone.
+pub fn analyze(
+    hprof: &Path,
+    index_dir: &Path,
+    opts: AnalyzeOptions,
+    progress: &dyn Fn(HeapPhase, u64, u64) -> Result<(), HeapError>,
+    cancel: &CancellationToken,
+) -> Result<HeapIndex, HeapError> {
+    let store = if opts.in_memory { Store::Ram } else { Store::Dir(index_dir.to_path_buf()) };
+    let result = run(hprof, &store, opts, progress, cancel);
+    // `run` has dropped every column by now: Windows cannot delete a mapped file.
+    if result.is_err() && !matches!(store, Store::Ram) {
+        remove_index_files(index_dir);
+    }
+    result
+}
+
+fn run(
+    hprof: &Path,
+    store: &Store,
+    opts: AnalyzeOptions,
+    report: &dyn Fn(HeapPhase, u64, u64) -> Result<(), HeapError>,
+    cancel: &CancellationToken,
+) -> Result<HeapIndex, HeapError> {
+    let started = Instant::now();
+    let progress = Progress::new(report, cancel);
+    let file = File::open(hprof)?;
+    // SAFETY: nothing writes the dump while it is analyzed or viewed: a capture finishes writing
+    // it before analysis and deletes it only with the snapshot that holds this index.
+    let dump = unsafe { Mmap::map(&file)? };
+    let graph = build_graph(&dump, store, report, cancel)?;
+    let Dominators { idom, preorder, .. } =
+        dominators(&graph.out_start, &graph.out, graph.root, store, report, cancel)?;
+
+    let nodes = u64::from(graph.root) + 1;
+    // Seven walks over the nodes, roughly: sizes, retained, child counts, child sorting, class
+    // counts and the tree walk, which visits every node twice.
+    let mut meter = Meter::start(&progress, HeapPhase::Retained, 7 * nodes)?;
+    let layout = VmLayout::from_vm_flags(opts.vm_flags.as_deref().unwrap_or(""), opts.java_major);
+    let model = ShallowModel::new(&graph, layout, opts.histogram.as_deref(), cancel)?;
+    let (shallow8, retained) = retained::sizes(&graph, &model, &idom, &preorder, store, &mut meter)?;
+    drop(preorder);
+    let (kids_start, kids) = retained::children(&idom, &retained, graph.root, store, &mut meter)?;
+    let tree = retained::Tree { idom, shallow8, retained, kids_start, kids };
+    let (classes, totals) = classes::table(&graph, &tree, store, &mut meter)?;
+    meter.finish()?;
+
+    let mut meter = Meter::start(&progress, HeapPhase::Paths, nodes + graph.out.len() as u64)?;
+    let gc_parent = paths::gc_parents(&graph, store, &mut meter)?;
+    meter.finish()?;
+
+    progress.report(HeapPhase::Saving, 0, 1)?;
+    let HeapGraph {
+        ids,
+        types,
+        offsets,
+        roots,
+        classes: class_info,
+        class_slot,
+        thread_objects,
+        weak_edges_skipped,
+        id_size,
+        ..
+    } = graph;
+    let summary = HeapSummary {
+        objects: totals.reachable_objects + totals.unreachable_objects,
+        reachable_objects: totals.reachable_objects,
+        reachable_bytes: totals.reachable_bytes,
+        unreachable_objects: totals.unreachable_objects,
+        unreachable_bytes: totals.unreachable_bytes,
+        classes: class_info.iter().filter(|c| c.id != 0).count() as u32,
+        weak_edges_skipped,
+        size_source: model.source,
+        analysis_ms: u32::try_from(started.elapsed().as_millis()).unwrap_or(u32::MAX),
+    };
+    if let Store::Dir(dir) = store {
+        write_summary(dir, &summary)?;
+    }
+    let mut root_kinds = HashMap::with_capacity(roots.len());
+    for (object, kind, serial) in roots {
+        root_kinds.entry(object).or_insert((kind, serial));
+    }
+    progress.report(HeapPhase::Saving, 1, 1)?;
+
+    let retained::Tree { idom, shallow8, retained, kids_start, kids } = tree;
+    Ok(HeapIndex {
+        hprof: dump,
+        ids,
+        types,
+        offsets,
+        shallow8,
+        retained,
+        idom,
+        kids_start,
+        kids,
+        gc_parent,
+        classes,
+        class_info,
+        class_slot,
+        roots: root_kinds,
+        thread_objects,
+        id_size,
+        summary,
+    })
+}
+
+/// Removes what an analysis may have written; a file that is not there is fine.
+fn remove_index_files(dir: &Path) {
+    for name in INDEX_FILES {
+        let path = dir.join(name);
+        if let Err(err) = std::fs::remove_file(&path)
+            && err.kind() != std::io::ErrorKind::NotFound
+        {
+            tracing::warn!("could not remove {}: {err}", path.display());
+        }
+    }
+}
+
+/// Steps of one phase: reports progress every [`CHECK_EVERY`] steps, and each loop looks at the
+/// token before its first step.
+pub(super) struct Meter<'a> {
+    progress: &'a Progress<'a>,
+    phase: HeapPhase,
+    /// An estimate; `done` is clamped to it.
+    total: u64,
+    done: u64,
+    next_report: u64,
+}
+
+impl<'a> Meter<'a> {
+    fn start(progress: &'a Progress<'a>, phase: HeapPhase, total: u64) -> Result<Self, HeapError> {
+        progress.report(phase, 0, total)?;
+        Ok(Self { progress, phase, total, done: 0, next_report: CHECK_EVERY })
+    }
+
+    pub(super) fn begin_loop(&self) -> Result<(), HeapError> {
+        self.progress.check()
+    }
+
+    pub(super) fn token(&self) -> &'a CancellationToken {
+        self.progress.token()
+    }
+
+    #[inline]
+    pub(super) fn tick(&mut self) -> Result<(), HeapError> {
+        self.advance(1)
+    }
+
+    pub(super) fn advance(&mut self, steps: u64) -> Result<(), HeapError> {
+        self.done += steps;
+        if self.done < self.next_report {
+            return Ok(());
+        }
+        self.next_report = self.done + CHECK_EVERY;
+        self.progress.report(self.phase, self.done.min(self.total), self.total)
+    }
+
+    fn finish(self) -> Result<(), HeapError> {
+        self.progress.report(self.phase, self.total, self.total)
+    }
+}
+
+#[cfg(test)]
+mod tests;
@@ -0,0 +1,117 @@
+//! The class table: counts and shallow sizes per type slot, retained sizes per type from a walk of
+//! the dominator tree, and the reachable / unreachable totals of the summary.
+
+use super::Meter;
+use super::retained::Tree;
+use crate::column::{NONE, Store};
+use crate::error::HeapError;
+use crate::graph::{HeapGraph, PRIM_SLOTS};
+use crate::hprof::BasicType;
+use crate::index::ClassStat;
+use crate::names::display_name;
+
+/// Instances and arrays only: class objects are left out.
+#[derive(Default)]
+pub(super) struct Totals {
+    pub(super) reachable_objects: u64,
+    pub(super) reachable_bytes: u64,
+    pub(super) unreachable_objects: u64,
+    pub(super) unreachable_bytes: u64,
+}
+
+pub(super) fn table(
+    graph: &HeapGraph,
+    tree: &Tree,
+    store: &Store,
+    meter: &mut Meter,
+) -> Result<(Vec<ClassStat>, Totals), HeapError> {
+    let slots = graph.classes.len() + PRIM_SLOTS as usize;
+    let mut classes: Vec<ClassStat> = (0..slots as u32)
+        .map(|slot| ClassStat { name: slot_name(graph, slot), count: 0, shallow: 0, retained: 0, unreachable_count: 0 })
+        .collect();
+    let mut totals = Totals::default();
+    meter.begin_loop()?;
+    for v in 0..graph.root as usize {
+        meter.tick()?;
+        let slot = graph.types[v];
+        let bytes = u64::from(tree.shallow8[v]) * 8;
+        let stat = &mut classes[slot as usize];
+        let instance = slot != graph.class_slot;
+        if tree.idom[v] == NONE {
+            stat.unreachable_count += 1;
+            if instance {
+                totals.unreachable_objects += 1;
+                totals.unreachable_bytes += bytes;
+            }
+        } else {
+            stat.count += 1;
+            stat.shallow += bytes;
+            if instance {
+                totals.reachable_objects += 1;
+                totals.reachable_bytes += bytes;
+            }
+        }
+    }
+    retained_by_type(graph, tree, &mut classes, store, meter)?;
+    Ok((classes, totals))
+}
+
+/// Adds to each type the retained size of its instances that have no instance of the same type
+/// above them in the dominator tree. The walk keeps no stack: it goes up through `idom` and
+/// resumes a node's children from `cursor`, while `active[t]` counts the instances of type `t` on
+/// the path from the root.
+fn retained_by_type(
+    graph: &HeapGraph,
+    tree: &Tree,
+    classes: &mut [ClassStat],
+    store: &Store,
+    meter: &mut Meter,
+) -> Result<(), HeapError> {
+    let root = graph.root;
+    let mut cursor = store.scratch::<u32>(root as usize + 1, 0)?;
+    let mut active = vec![0u32; classes.len()];
+    cursor[root as usize] = tree.kids_start[root as usize];
+    let mut v = root;
+    meter.begin_loop()?;
+    loop {
+        meter.tick()?;
+        let next = cursor[v as usize];
+        if next < tree.kids_start[v as usize + 1] {
+            cursor[v as usize] = next + 1;
+            let child = tree.kids[next as usize];
+            let slot = graph.types[child as usize] as usize;
+            if active[slot] == 0 {
+                classes[slot].retained += tree.retained[child as usize];
+            }
+            active[slot] += 1;
+            cursor[child as usize] = tree.kids_start[child as usize];
+            v = child;
+        } else if v == root {
+            return Ok(());
+        } else {
+            active[graph.types[v as usize] as usize] -= 1;
+            v = tree.idom[v as usize];
+        }
+    }
+}
+
+fn slot_name(graph: &HeapGraph, slot: u32) -> String {
+    match graph.classes.get(slot as usize) {
+        Some(class) => display_name(&class.name),
+        None => graph.prim_type(slot).map_or_else(String::new, |ty| prim_array_name(ty).to_owned()),
+    }
+}
+
+fn prim_array_name(ty: BasicType) -> &'static str {
+    match ty {
+        BasicType::Object => "java.lang.Object[]",
+        BasicType::Boolean => "boolean[]",
+        BasicType::Char => "char[]",
+        BasicType::Float => "float[]",
+        BasicType::Double => "double[]",
+        BasicType::Byte => "byte[]",
+        BasicType::Short => "short[]",
+        BasicType::Int => "int[]",
+        BasicType::Long => "long[]",
+    }
+}
@@ -0,0 +1,32 @@
+//! Parents on shortest paths from the GC roots, by breadth-first search over strong edges.
+
+use super::Meter;
+use crate::column::{Column, NONE, Store};
+use crate::error::HeapError;
+use crate::files::GC_PARENT_FILE;
+use crate::graph::HeapGraph;
+
+pub(super) fn gc_parents(graph: &HeapGraph, store: &Store, meter: &mut Meter) -> Result<Column<u32>, HeapError> {
+    let root = graph.root;
+    let nodes = root as usize + 1;
+    let mut parent = store.column::<u32>(GC_PARENT_FILE, nodes, NONE)?;
+    let mut queue = store.scratch::<u32>(nodes, 0)?;
+    parent[root as usize] = root;
+    queue[0] = root;
+    let (mut head, mut tail) = (0, 1);
+    meter.begin_loop()?;
+    while head < tail {
+        meter.tick()?;
+        let v = queue[head];
+        head += 1;
+        for &w in graph.out_edges(v) {
+            meter.tick()?;
+            if parent[w as usize] == NONE {
+                parent[w as usize] = v;
+                queue[tail] = w;
+                tail += 1;
+            }
+        }
+    }
+    Ok(parent)
+}
@@ -0,0 +1,209 @@
+//! Shallow and retained sizes, and the dominator tree as child lists sorted by retained size.
+
+use std::cmp::Reverse;
+
+use super::Meter;
+use crate::column::{Column, NONE, Store};
+use crate::error::HeapError;
+use crate::files::{KIDS_FILE, KIDS_START_FILE, RETAINED_FILE, SHALLOW8_FILE};
+use crate::graph::{HeapGraph, accumulate};
+use crate::progress::CHECK_EVERY;
+use crate::shallow::ShallowModel;
+
+/// The dominator tree with sizes; see [`crate::HeapIndex`] for each column.
+pub(super) struct Tree {
+    pub(super) idom: Column<u32>,
+    pub(super) shallow8: Column<u32>,
+    pub(super) retained: Column<u64>,
+    pub(super) kids_start: Column<u32>,
+    pub(super) kids: Column<u32>,
+}
+
+/// `(shallow8, retained)` of every node, the root included. `preorder` is the DFS preorder of the
+/// reachable nodes, from the root.
+pub(super) fn sizes(
+    graph: &HeapGraph,
+    model: &ShallowModel,
+    idom: &[u32],
+    preorder: &[u32],
+    store: &Store,
+    meter: &mut Meter,
+) -> Result<(Column<u32>, Column<u64>), HeapError> {
+    let nodes = idom.len();
+    let mut shallow8 = store.column::<u32>(SHALLOW8_FILE, nodes, 0)?;
+    let mut retained = store.column::<u64>(RETAINED_FILE, nodes, 0)?;
+    meter.begin_loop()?;
+    for v in 0..graph.root {
+        meter.tick()?;
+        let units = u32::try_from(model.size_of(graph, v).div_ceil(8)).unwrap_or(u32::MAX);
+        shallow8[v as usize] = units;
+        if idom[v as usize] != NONE {
+            retained[v as usize] = u64::from(units) * 8;
+        }
+    }
+    // A dominator is a DFS ancestor, so in reverse preorder every node is complete before it is
+    // added to its dominator.
+    meter.begin_loop()?;
+    for &v in preorder.iter().skip(1).rev() {
+        meter.tick()?;
+        let own = retained[v as usize];
+        retained[idom[v as usize] as usize] += own;
+    }
+    Ok((shallow8, retained))
+}
+
+/// `(kids_start, kids)`: the children of every reachable node in the dominator tree, as CSR.
+pub(super) fn children(
+    idom: &[u32],
+    retained: &[u64],
+    root: u32,
+    store: &Store,
+    meter: &mut Meter,
+) -> Result<(Column<u32>, Column<u32>), HeapError> {
+    let nodes = idom.len();
+    // Counted into `starts[d + 1]`; filled through `starts[d]` as a cursor, which leaves it at d's
+    // end until the column is shifted one slot.
+    let mut starts = store.column::<u32>(KIDS_START_FILE, nodes + 1, 0)?;
+    let mut longest = 0;
+    meter.begin_loop()?;
+    for (v, &d) in idom.iter().enumerate() {
+        meter.tick()?;
+        if d != NONE && v != root as usize {
+            let count = &mut starts[d as usize + 1];
+            *count += 1;
+            longest = longest.max(*count);
+        }
+    }
+    let total = accumulate(&mut starts, meter.token())?;
+    let mut kids = store.column::<u32>(KIDS_FILE, total as usize, 0)?;
+    meter.begin_loop()?;
+    for (v, &d) in idom.iter().enumerate() {
+        meter.tick()?;
+        if d != NONE && v != root as usize {
+            let cursor = &mut starts[d as usize];
+            kids[*cursor as usize] = v as u32;
+            *cursor += 1;
+        }
+    }
+    starts.copy_within(0..nodes, 1);
+    starts[0] = 0;
+    sort_children(&mut kids, &starts, retained, longest as usize, CHECK_EVERY as usize, store, meter)?;
+    Ok((starts, kids))
+}
+
+/// Sorts each child list by retained size descending, then by node. A list longer than `run` is
+/// sorted in runs of `run` that are then merged, so no single step takes longer than one run.
+fn sort_children(
+    kids: &mut [u32],
+    starts: &[u32],
+    retained: &[u64],
+    longest: usize,
+    run: usize,
+    store: &Store,
+    meter: &mut Meter,
+) -> Result<(), HeapError> {
+    let key = |&k: &u32| (Reverse(retained[k as usize]), k);
+    let mut buffer = if longest > run { store.scratch::<u32>(longest, 0)? } else { Column::ram(0, 0) };
+    meter.begin_loop()?;
+    for v in 0..starts.len() - 1 {
+        meter.tick()?;
+        let list = &mut kids[starts[v] as usize..starts[v + 1] as usize];
+        if list.len() <= run {
+            list.sort_unstable_by_key(key);
+            meter.advance(list.len() as u64)?;
+            continue;
+        }
+        for chunk in list.chunks_mut(run) {
+            meter.begin_loop()?;
+            chunk.sort_unstable_by_key(key);
+            meter.advance(chunk.len() as u64)?;
+        }
+        let len = list.len();
+        let (mut from, mut to) = (list, &mut buffer[..len]);
+        let mut in_buffer = false;
+        let mut width = run;
+        while width < len {
+            meter.begin_loop()?;
+            for start in (0..len).step_by(2 * width) {
+                let mid = (start + width).min(len);
+                let end = (start + 2 * width).min(len);
+                merge(&from[start..mid], &from[mid..end], &mut to[start..end], &key, meter)?;
+            }
+            std::mem::swap(&mut from, &mut to);
+            in_buffer = !in_buffer;
+            width *= 2;
+        }
+        if in_buffer {
+            to.copy_from_slice(from);
+        }
+    }
+    Ok(())
+}
+
+fn merge<K: Ord>(
+    left: &[u32],
+    right: &[u32],
+    out: &mut [u32],
+    key: &impl Fn(&u32) -> K,
+    meter: &mut Meter,
+) -> Result<(), HeapError> {
+    let (mut i, mut j) = (0, 0);
+    for slot in out.iter_mut() {
+        meter.tick()?;
+        let take_left = j == right.len() || (i < left.len() && key(&left[i]) <= key(&right[j]));
+        if take_left {
+            *slot = left[i];
+            i += 1;
+        } else {
+            *slot = right[j];
+            j += 1;
+        }
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use ide_core::CancellationToken;
+
+    use super::*;
+    use crate::error::HeapError;
+    use crate::progress::{HeapPhase, Progress};
+    use crate::testing::XorShift;
+
+    fn sorted_with_run(lists: &[Vec<u32>], retained: &[u64], run: usize) -> Vec<u32> {
+        let report = |_, _, _| Ok::<(), HeapError>(());
+        let cancel = CancellationToken::new();
+        let progress = Progress::new(&report, &cancel);
+        let mut meter = Meter::start(&progress, HeapPhase::Retained, 1).unwrap();
+        let mut starts = vec![0u32];
+        for list in lists {
+            starts.push(starts.last().unwrap() + list.len() as u32);
+        }
+        let mut kids: Vec<u32> = lists.concat();
+        let longest = lists.iter().map(Vec::len).max().unwrap_or(0);
+        sort_children(&mut kids, &starts, retained, longest, run, &Store::Ram, &mut meter).unwrap();
+        kids
+    }
+
+    #[test]
+    fn long_child_lists_are_merged_from_sorted_runs() {
+        let mut rng = XorShift::new(7);
+        let retained: Vec<u64> = (0..500).map(|_| u64::from(rng.below(40))).collect();
+        let mut nodes: Vec<u32> = (0..500).collect();
+        for i in (1..nodes.len()).rev() {
+            nodes.swap(i, rng.below(i as u32 + 1) as usize);
+        }
+        // Lists below, at and above the run length, and one that ends on a partial run.
+        let lists = vec![nodes[..3].to_vec(), nodes[3..7].to_vec(), nodes[7..400].to_vec(), nodes[400..].to_vec()];
+        for run in [1, 4, 7, 64, 1000] {
+            let mut expected = Vec::new();
+            for list in &lists {
+                let mut list = list.clone();
+                list.sort_by_key(|&k| (Reverse(retained[k as usize]), k));
+                expected.extend(list);
+            }
+            assert_eq!(sorted_with_run(&lists, &retained, run), expected, "run {run}");
+        }
+    }
+}
@@ -0,0 +1,265 @@
+use std::cell::Cell;
+use std::collections::BTreeSet;
+use std::path::PathBuf;
+
+use super::*;
+use crate::column::{Column, NONE};
+use crate::files::INDEX_FILES;
+use crate::hprof::{BasicType, HeapRootKind};
+use crate::index::{ClassStat, HeapSummary};
+use crate::shallow::SizeSource;
+use crate::testing::{MiniHeap, sample};
+
+const DUMP: &str = "dump.hprof";
+const ROOT: Option<&str> = Some("java.lang.Object");
+const NODE_FIELDS: [(&str, BasicType); 2] = [("next", BasicType::Object), ("v", BasicType::Long)];
+
+fn no_progress(_: HeapPhase, _: u64, _: u64) -> Result<(), HeapError> {
+    Ok(())
+}
+
+fn in_memory() -> AnalyzeOptions {
+    AnalyzeOptions { in_memory: true, ..AnalyzeOptions::default() }
+}
+
+/// A temp dir holding `heap` as `dump.hprof`, the way a capture lays out its snapshot dir.
+fn dump_dir(heap: &MiniHeap) -> (tempfile::TempDir, PathBuf) {
+    let dir = tempfile::tempdir().unwrap();
+    let dump = dir.path().join(DUMP);
+    heap.write_to(&dump);
+    (dir, dump)
+}
+
+fn analyze_with(heap: &MiniHeap, opts: AnalyzeOptions) -> HeapIndex {
+    let (dir, dump) = dump_dir(heap);
+    analyze(&dump, dir.path(), opts, &no_progress, &CancellationToken::new()).unwrap()
+}
+
+fn file_names(dir: &Path) -> BTreeSet<String> {
+    std::fs::read_dir(dir).unwrap().map(|e| e.unwrap().file_name().into_string().unwrap()).collect()
+}
+
+fn node(index: &HeapIndex, addr: u64) -> u32 {
+    index.ids.rank(addr).unwrap_or_else(|| panic!("{addr:#x} is not in the dump"))
+}
+
+fn shallow(index: &HeapIndex, addr: u64) -> u64 {
+    u64::from(index.shallow8[node(index, addr) as usize]) * 8
+}
+
+fn retained(index: &HeapIndex, addr: u64) -> u64 {
+    index.retained[node(index, addr) as usize]
+}
+
+fn idom(index: &HeapIndex, addr: u64) -> u32 {
+    index.idom[node(index, addr) as usize]
+}
+
+fn class<'a>(index: &'a HeapIndex, name: &str) -> &'a ClassStat {
+    let mut rows = index.classes.iter().filter(|c| c.name == name);
+    let row = rows.next().unwrap_or_else(|| panic!("no class {name}"));
+    assert!(rows.next().is_none(), "two classes named {name}");
+    row
+}
+
+fn stat(c: &ClassStat) -> (u64, u64, u64, u64) {
+    (c.count, c.shallow, c.retained, c.unreachable_count)
+}
+
+fn kids_of(index: &HeapIndex, v: u32) -> &[u32] {
+    &index.kids[index.kids_start[v as usize] as usize..index.kids_start[v as usize + 1] as usize]
+}
+
+const H: u64 = 0x1000;
+const N1: u64 = 0x1100;
+const N2: u64 = 0x1200;
+const N3: u64 = 0x1300;
+const L: u64 = 0x1400;
+const R: u64 = 0x1500;
+const W: u64 = 0x1600;
+const N4: u64 = 0x1700;
+
+#[test]
+fn sample_retained_sizes() {
+    let index = analyze_with(&sample(), in_memory());
+    let shallows: Vec<u64> = [H, N1, L, W, R].iter().map(|&a| shallow(&index, a)).collect();
+    assert_eq!(shallows, [24, 24, 32, 24, 32]);
+    let retained_sizes: Vec<u64> = [L, N3, N1, N2, H, W].iter().map(|&a| retained(&index, a)).collect();
+    assert_eq!(retained_sizes, [32, 56, 24, 24, 72, 24]);
+    let root = index.root();
+    assert_eq!(
+        [idom(&index, N1), idom(&index, N2), idom(&index, N3), idom(&index, L)],
+        [node(&index, H), node(&index, H), root, node(&index, N3)]
+    );
+    assert_eq!((idom(&index, R), idom(&index, N4)), (NONE, NONE));
+    assert_eq!((retained(&index, R), retained(&index, N4)), (0, 0));
+
+    // Children by retained size descending, ties by node.
+    assert_eq!(kids_of(&index, node(&index, H)), [node(&index, N1), node(&index, N2)]);
+    let top = kids_of(&index, root);
+    assert_eq!(top[..2], [node(&index, H), node(&index, N3)]);
+    assert!(top.windows(2).all(|p| (index.retained[p[0] as usize], p[1]) > (index.retained[p[1] as usize], p[0])));
+    assert_eq!(index.retained[root as usize], top.iter().map(|&k| index.retained[k as usize]).sum::<u64>());
+}
+
+#[test]
+fn sample_class_table() {
+    let index = analyze_with(&sample(), in_memory());
+    assert_eq!(stat(class(&index, "Node")), (3, 72, 104, 1));
+    assert_eq!(stat(class(&index, "Holder")), (1, 24, 72, 0));
+    assert_eq!(stat(class(&index, "long[]")), (1, 32, 32, 0));
+    assert_eq!(stat(class(&index, "java.lang.ref.WeakReference")), (1, 24, 24, 0));
+    let objects = class(&index, "java.lang.Object[]");
+    assert_eq!((objects.count, objects.unreachable_count), (0, 1));
+    // Indexed by type slot, primitive array slots included.
+    assert_eq!(index.classes.len(), index.class_info.len() + crate::graph::PRIM_SLOTS as usize);
+    let slot = index.types[node(&index, H) as usize] as usize;
+    assert_eq!(index.classes[slot].name, "Holder");
+    assert_eq!(index.class_info[slot].name, "Holder");
+}
+
+#[test]
+fn sample_summary() {
+    let index = analyze_with(&sample(), in_memory());
+    let s = &index.summary;
+    assert_eq!(
+        (s.reachable_objects, s.reachable_bytes, s.unreachable_objects, s.unreachable_bytes, s.weak_edges_skipped),
+        (6, 152, 2, 56, 1)
+    );
+    assert_eq!(s.objects, 8);
+    assert_eq!(s.classes, 7);
+    assert_eq!(s.size_source, SizeSource::Layout);
+}
+
+#[test]
+fn sample_gc_parents() {
+    let index = analyze_with(&sample(), in_memory());
+    let parent = |addr| index.gc_parent[node(&index, addr) as usize];
+    let root = index.root();
+    assert_eq!(index.gc_parent[root as usize], root);
+    assert_eq!([parent(H), parent(N3), parent(W)], [root; 3]);
+    assert_eq!([parent(N1), parent(L)], [node(&index, H), node(&index, N3)]);
+    assert_eq!((parent(R), parent(N4)), (NONE, NONE));
+    assert_eq!(index.roots[&node(&index, N3)], (HeapRootKind::JavaFrame, Some(1)));
+}
+
+#[test]
+fn cycle_is_retained_once() {
+    let mut heap = MiniHeap::new();
+    heap.class("Node", ROOT, &NODE_FIELDS)
+        .object(0xA000, "Node", &[("next", 0xB000)])
+        .object(0xB000, "Node", &[("next", 0xA000)])
+        .root(HeapRootKind::JniGlobal, 0xA000);
+    let index = analyze_with(&heap, in_memory());
+    assert_eq!((retained(&index, 0xA000), retained(&index, 0xB000)), (48, 24));
+    assert_eq!(class(&index, "Node").retained, 48);
+}
+
+#[test]
+fn deep_chain_retained_by_class() {
+    // A small stack proves no pass recurses along the chain.
+    std::thread::Builder::new()
+        .stack_size(256 << 10)
+        .spawn(|| {
+            const NODES: u64 = 1_000_000;
+            let addr = |i: u64| 0x10_0000 + i * 0x100;
+            let mut heap = MiniHeap::new();
+            heap.class("Node", ROOT, &NODE_FIELDS);
+            for i in 0..NODES {
+                let next = if i + 1 < NODES { addr(i + 1) } else { 0 };
+                heap.object(addr(i), "Node", &[("next", next)]);
+            }
+            heap.root(HeapRootKind::JniGlobal, addr(0));
+            let index = analyze_with(&heap, in_memory());
+            assert_eq!(retained(&index, addr(0)), 24_000_000);
+            assert_eq!(retained(&index, addr(NODES - 1)), 24);
+            assert_eq!(class(&index, "Node").retained, 24_000_000);
+            assert_eq!(index.summary.reachable_objects, NODES);
+        })
+        .unwrap()
+        .join()
+        .unwrap();
+}
+
+#[test]
+fn histogram_sizes_flow_into_retained() {
+    let histogram = vec![HistogramEntry { name: "Node".to_owned(), instances: 3, bytes: 96 }];
+    let index = analyze_with(&sample(), AnalyzeOptions { histogram: Some(histogram), ..in_memory() });
+    assert_eq!(shallow(&index, N3), 32);
+    assert_eq!(retained(&index, N3), 64);
+    assert_eq!(index.summary.size_source, SizeSource::Jvm);
+}
+
+#[test]
+fn vm_flags_of_an_unknown_jdk_set_the_layout() {
+    // Without compressed oops a Node is 12 + 8 + 8 = 28 → 32.
+    let opts = AnalyzeOptions { vm_flags: Some("-XX:MaxHeapSize=42949672960".to_owned()), ..in_memory() };
+    let index = analyze_with(&sample(), opts);
+    assert_eq!(shallow(&index, N4), 32);
+}
+
+#[test]
+fn cancel_stops_analysis() {
+    let (dir, dump) = dump_dir(&sample());
+    let cancel = CancellationToken::new();
+    let phases = Cell::new(Vec::new());
+    let progress = |phase, _, _| {
+        let mut seen = phases.take();
+        seen.push(phase);
+        phases.set(seen);
+        if phase == HeapPhase::Dominators {
+            cancel.cancel();
+        }
+        Ok(())
+    };
+    let result = analyze(&dump, dir.path(), AnalyzeOptions::default(), &progress, &cancel);
+    assert!(matches!(result, Err(HeapError::Cancelled)), "{:?}", result.err());
+    assert_eq!(phases.take().last(), Some(&HeapPhase::Dominators), "stops at the first check");
+    // The dump the capture analyzes in place stays; every file the analysis wrote is gone.
+    assert_eq!(file_names(dir.path()), BTreeSet::from([DUMP.to_owned()]));
+}
+
+#[test]
+fn progress_error_stops_analysis_and_removes_its_files() {
+    let (dir, dump) = dump_dir(&sample());
+    // The last report, once every index file is written.
+    let progress = |phase, done, _| match (phase, done) {
+        (HeapPhase::Saving, 1) => Err(HeapError::NotEnoughDiskSpace { needed: 2, available: 1 }),
+        _ => Ok(()),
+    };
+    let result = analyze(&dump, dir.path(), AnalyzeOptions::default(), &progress, &CancellationToken::new());
+    assert!(matches!(result, Err(HeapError::NotEnoughDiskSpace { .. })), "{:?}", result.err());
+    assert_eq!(file_names(dir.path()), BTreeSet::from([DUMP.to_owned()]));
+}
+
+#[test]
+fn writes_index_files() {
+    let (dir, dump) = dump_dir(&sample());
+    let index = analyze(&dump, dir.path(), AnalyzeOptions::default(), &no_progress, &CancellationToken::new()).unwrap();
+    let names = file_names(dir.path());
+    for name in [
+        "types.u32",
+        "shallow8.u32",
+        "retained.u64",
+        "idom.u32",
+        "kids_start.u32",
+        "kids.u32",
+        "gc_parent.u32",
+        "offsets.u64",
+        "summary.json",
+    ] {
+        assert!(names.contains(name), "{name} missing from {names:?}");
+    }
+    assert!(names.iter().any(|n| n.starts_with("ids.")), "{names:?}");
+    // Nothing but the index files and the dump: scratch arrays are gone.
+    let unexpected: Vec<_> = names.iter().filter(|n| *n != DUMP && !INDEX_FILES.contains(&n.as_str())).collect();
+    assert!(unexpected.is_empty(), "{unexpected:?}");
+
+    let summary: HeapSummary =
+        serde_json::from_slice(&std::fs::read(dir.path().join("summary.json")).unwrap()).unwrap();
+    assert_eq!(summary, index.summary);
+    drop(index);
+    let retained = Column::<u64>::open(&dir.path().join("retained.u64")).unwrap();
+    let in_ram = analyze_with(&sample(), in_memory());
+    assert_eq!(&*retained, &*in_ram.retained);
+}
@@ -0,0 +1,306 @@
+//! Fixed-width arrays that live in RAM or in memory-mapped files, so the analysis arrays of a
+//! large heap do not count against the process's anonymous memory.
+
+use std::fs::{File, OpenOptions};
+use std::io;
+use std::marker::PhantomData;
+use std::mem::size_of;
+use std::ops::{Deref, DerefMut};
+use std::path::{Path, PathBuf};
+
+use memmap2::{Mmap, MmapMut};
+
+/// Index value meaning "no object" in every `u32` column.
+pub const NONE: u32 = u32::MAX;
+
+/// # Safety
+/// Every bit pattern of `Self` is a valid value and `Self` has no padding, so a mapped file's
+/// bytes can be viewed as `[Self]` and a file can be written from one.
+pub unsafe trait Pod: Copy + 'static {}
+
+unsafe impl Pod for u8 {}
+unsafe impl Pod for u32 {}
+unsafe impl Pod for u64 {}
+
+enum Backing<T> {
+    Ram(Vec<T>),
+    Map(MmapMut),
+    ReadOnly(Mmap),
+}
+
+/// A fixed-length array of `T`. File contents use the machine's byte order: the arrays are a
+/// cache for this machine and are never moved to another one.
+pub struct Column<T: Pod> {
+    backing: Backing<T>,
+    /// Scratch file to delete on drop where it could not be unlinked while mapped.
+    delete_on_drop: Option<PathBuf>,
+    _marker: PhantomData<T>,
+}
+
+fn byte_len<T: Pod>(len: usize) -> io::Result<usize> {
+    len.checked_mul(size_of::<T>())
+        .filter(|&bytes| bytes <= isize::MAX as usize)
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "column is too long"))
+}
+
+fn is_zero<T: Pod>(value: &T) -> bool {
+    // SAFETY: `Pod` types have no padding, so all `size_of::<T>()` bytes are initialized.
+    let bytes = unsafe { std::slice::from_raw_parts((value as *const T).cast::<u8>(), size_of::<T>()) };
+    bytes.iter().all(|&b| b == 0)
+}
+
+impl<T: Pod> Column<T> {
+    fn new(backing: Backing<T>, delete_on_drop: Option<PathBuf>) -> Self {
+        Self { backing, delete_on_drop, _marker: PhantomData }
+    }
+
+    pub fn ram(len: usize, fill: T) -> Self {
+        Self::new(Backing::Ram(vec![fill; len]), None)
+    }
+
+    /// Creates (or truncates) `path` and maps it; the file stays after the column is dropped.
+    pub fn file(path: &Path, len: usize, fill: T) -> io::Result<Self> {
+        let bytes = byte_len::<T>(len)?;
+        let file = OpenOptions::new().read(true).write(true).create(true).truncate(true).open(path)?;
+        match Self::map_new(&file, len, bytes, fill) {
+            Ok(column) => Ok(column),
+            Err(err) => {
+                drop(file);
+                let _ = std::fs::remove_file(path);
+                Err(err)
+            }
+        }
+    }
+
+    /// A temporary column in `dir` that never outlives the process. On unix the file is unlinked
+    /// right after mapping, so even a crash leaves nothing behind; elsewhere (Windows cannot delete
+    /// a mapped file) it is deleted on drop.
+    pub fn scratch(dir: &Path, len: usize, fill: T) -> io::Result<Self> {
+        let bytes = byte_len::<T>(len)?;
+        if bytes == 0 {
+            return Ok(Self::ram(0, fill));
+        }
+        let path = dir.join(format!("scratch-{}.tmp", uuid::Uuid::new_v4().simple()));
+        let file = OpenOptions::new().read(true).write(true).create_new(true).open(&path)?;
+        let mut column = match Self::map_new(&file, len, bytes, fill) {
+            Ok(column) => column,
+            Err(err) => {
+                drop(file);
+                let _ = std::fs::remove_file(&path);
+                return Err(err);
+            }
+        };
+        // If the unlink fails the drop-time delete is the fallback, so the file is never left behind.
+        if !cfg!(unix) || std::fs::remove_file(&path).is_err() {
+            column.delete_on_drop = Some(path);
+        }
+        Ok(column)
+    }
+
+    /// Maps an existing file read-only; writing through the column panics.
+    pub fn open(path: &Path) -> io::Result<Self> {
+        let file = File::open(path)?;
+        let bytes = file.metadata()?.len();
+        if bytes % size_of::<T>() as u64 != 0 {
+            return Err(io::Error::new(
+                io::ErrorKind::InvalidData,
+                format!("{} is {bytes} bytes, not a multiple of {}", path.display(), size_of::<T>()),
+            ));
+        }
+        if bytes == 0 {
+            return Ok(Self::ram(0, zero_value()));
+        }
+        // SAFETY: the file is private to the snapshot directory; nothing truncates it while mapped.
+        let map = unsafe { Mmap::map(&file)? };
+        Ok(Self::new(Backing::ReadOnly(map), None))
+    }
+
+    fn map_new(file: &File, len: usize, bytes: usize, fill: T) -> io::Result<Self> {
+        if bytes == 0 {
+            return Ok(Self::ram(0, fill));
+        }
+        file.set_len(bytes as u64)?;
+        // SAFETY: as in `open`; the file was just created by this process.
+        let map = unsafe { MmapMut::map_mut(file)? };
+        let mut column = Self::new(Backing::Map(map), None);
+        // A fresh file reads as zeros, and writing a zero fill would fault in every page.
+        if !is_zero(&fill) {
+            debug_assert_eq!(column.len(), len);
+            column.fill(fill);
+        }
+        Ok(column)
+    }
+}
+
+fn zero_value<T: Pod>() -> T {
+    // SAFETY: every bit pattern of a `Pod` type is a valid value.
+    unsafe { std::mem::zeroed() }
+}
+
+impl<T: Pod> Deref for Column<T> {
+    type Target = [T];
+
+    fn deref(&self) -> &[T] {
+        match &self.backing {
+            Backing::Ram(v) => v,
+            // SAFETY: maps are page-aligned, `T` is `Pod` and the length is a multiple of its width.
+            Backing::Map(m) => unsafe { std::slice::from_raw_parts(m.as_ptr().cast(), m.len() / size_of::<T>()) },
+            Backing::ReadOnly(m) => unsafe { std::slice::from_raw_parts(m.as_ptr().cast(), m.len() / size_of::<T>()) },
+        }
+    }
+}
+
+impl<T: Pod> DerefMut for Column<T> {
+    fn deref_mut(&mut self) -> &mut [T] {
+        match &mut self.backing {
+            Backing::Ram(v) => v,
+            // SAFETY: as in `deref`; `&mut self` makes the access exclusive.
+            Backing::Map(m) => unsafe {
+                std::slice::from_raw_parts_mut(m.as_mut_ptr().cast(), m.len() / size_of::<T>())
+            },
+            Backing::ReadOnly(_) => panic!("column opened from a file is read-only"),
+        }
+    }
+}
+
+impl<T: Pod> Drop for Column<T> {
+    fn drop(&mut self) {
+        if let Some(path) = self.delete_on_drop.take() {
+            // Unmap before deleting: Windows refuses to delete a mapped file.
+            self.backing = Backing::Ram(Vec::new());
+            let _ = std::fs::remove_file(path);
+        }
+    }
+}
+
+/// Where a computation puts its arrays: result columns are named files in `Dir`, temporary arrays
+/// are scratch columns in the same directory.
+pub enum Store {
+    Ram,
+    Dir(PathBuf),
+}
+
+impl Store {
+    /// `name` is the whole file name (`types.u32`); it is used as given.
+    pub fn column<T: Pod>(&self, name: &str, len: usize, fill: T) -> io::Result<Column<T>> {
+        match self {
+            Store::Ram => Ok(Column::ram(len, fill)),
+            Store::Dir(dir) => Column::file(&dir.join(name), len, fill),
+        }
+    }
+
+    pub fn scratch<T: Pod>(&self, len: usize, fill: T) -> io::Result<Column<T>> {
+        match self {
+            Store::Ram => Ok(Column::ram(len, fill)),
+            Store::Dir(dir) => Column::scratch(dir, len, fill),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn dir_is_empty(dir: &Path) -> bool {
+        std::fs::read_dir(dir).unwrap().next().is_none()
+    }
+
+    #[test]
+    fn file_column_round_trips() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("x.u32");
+        {
+            let mut col = Column::<u32>::file(&path, 4, 7).unwrap();
+            assert_eq!(&*col, &[7, 7, 7, 7]);
+            col.copy_from_slice(&[1, 2, 3, 4]);
+        }
+        let col = Column::<u32>::open(&path).unwrap();
+        assert_eq!(&*col, &[1, 2, 3, 4]);
+    }
+
+    #[test]
+    fn scratch_column_leaves_no_file() {
+        let dir = tempfile::tempdir().unwrap();
+        {
+            let mut col = Column::<u64>::scratch(dir.path(), 1024, 9).unwrap();
+            col[1023] = 5;
+            assert_eq!((col[0], col[1023]), (9, 5));
+            // Windows cannot delete a mapped file, so only unix unlinks while the map is alive.
+            #[cfg(unix)]
+            assert!(dir_is_empty(dir.path()));
+        }
+        assert!(dir_is_empty(dir.path()));
+    }
+
+    #[test]
+    fn zero_fill_and_non_zero_fill() {
+        let dir = tempfile::tempdir().unwrap();
+        let zeros = Column::<u64>::file(&dir.path().join("z.u64"), 3000, 0).unwrap();
+        assert!(zeros.iter().all(|&v| v == 0));
+        let ones = Column::<u8>::file(&dir.path().join("o.u8"), 10_000, 0xAB).unwrap();
+        assert!(ones.iter().all(|&v| v == 0xAB));
+        let ram = Column::<u32>::ram(5, NONE);
+        assert_eq!(&*ram, &[NONE; 5]);
+    }
+
+    #[test]
+    fn empty_columns_work_everywhere() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("e.u32");
+        assert_eq!(Column::<u32>::ram(0, 1).len(), 0);
+        assert_eq!(Column::<u32>::file(&path, 0, 1).unwrap().len(), 0);
+        assert_eq!(Column::<u32>::open(&path).unwrap().len(), 0);
+        assert_eq!(Column::<u32>::scratch(dir.path(), 0, 1).unwrap().len(), 0);
+    }
+
+    #[test]
+    fn open_rejects_a_length_that_is_not_a_multiple_of_the_width() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("odd.u32");
+        std::fs::write(&path, [0u8; 6]).unwrap();
+        let err = Column::<u32>::open(&path).err().unwrap();
+        assert_eq!(err.kind(), io::ErrorKind::InvalidData);
+    }
+
+    #[test]
+    fn file_rejects_a_byte_length_that_overflows() {
+        let dir = tempfile::tempdir().unwrap();
+        let err = Column::<u64>::file(&dir.path().join("big.u64"), usize::MAX, 0).err().unwrap();
+        assert_eq!(err.kind(), io::ErrorKind::InvalidInput);
+        assert!(dir_is_empty(dir.path()));
+    }
+
+    #[test]
+    #[should_panic(expected = "read-only")]
+    fn opened_columns_are_read_only() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("r.u32");
+        drop(Column::<u32>::file(&path, 2, 0).unwrap());
+        let mut col = Column::<u32>::open(&path).unwrap();
+        col[0] = 1;
+    }
+
+    #[test]
+    fn store_names_files_exactly_as_given() {
+        let dir = tempfile::tempdir().unwrap();
+        let store = Store::Dir(dir.path().to_path_buf());
+        let mut col = store.column::<u32>("types.u32", 3, 0).unwrap();
+        col[2] = 8;
+        drop(col);
+        let names: Vec<_> =
+            std::fs::read_dir(dir.path()).unwrap().map(|e| e.unwrap().file_name().into_string().unwrap()).collect();
+        assert_eq!(names, ["types.u32"]);
+        assert_eq!(&*Column::<u32>::open(&dir.path().join("types.u32")).unwrap(), &[0, 0, 8]);
+    }
+
+    #[test]
+    fn store_scratch_is_anonymous_and_ram_store_touches_no_disk() {
+        let dir = tempfile::tempdir().unwrap();
+        let store = Store::Dir(dir.path().to_path_buf());
+        drop(store.scratch::<u32>(16, 1).unwrap());
+        assert!(dir_is_empty(dir.path()));
+        let ram = Store::Ram.column::<u8>("ignored.u8", 4, 2).unwrap();
+        assert_eq!(&*ram, &[2; 4]);
+        assert_eq!(&*Store::Ram.scratch::<u8>(2, 3).unwrap(), &[3, 3]);
+    }
+}
@@ -0,0 +1,59 @@
+//! Class-by-class differences between two snapshots.
+
+use std::collections::HashMap;
+
+use serde::{Deserialize, Serialize};
+
+use crate::index::HeapIndex;
+
+/// One class name in two snapshots; a side without the name has zeros.
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, specta::Type)]
+#[serde(rename_all = "camelCase")]
+pub struct HeapClassDiff {
+    pub name: String,
+    pub base_count: u64,
+    pub count: u64,
+    pub base_shallow: u64,
+    pub shallow: u64,
+    pub base_retained: u64,
+    pub retained: u64,
+    /// Several classes have this name on one side or both, so the sides cannot be matched: each
+    /// side shows the sum over its classes, and there is no difference to show.
+    pub ambiguous: bool,
+}
+
+/// Joined by class name, the only key two dumps share. By the change in retained size
+/// descending (none for an ambiguous name), then by name.
+pub fn compare(base: &HeapIndex, current: &HeapIndex) -> Vec<HeapClassDiff> {
+    let mut by_name: HashMap<String, HeapClassDiff> = HashMap::new();
+    for (index, is_base) in [(base, true), (current, false)] {
+        for row in index.class_rows() {
+            let diff = by_name.entry(row.name).or_insert_with_key(|name| HeapClassDiff {
+                name: name.clone(),
+                base_count: 0,
+                count: 0,
+                base_shallow: 0,
+                shallow: 0,
+                base_retained: 0,
+                retained: 0,
+                ambiguous: false,
+            });
+            let (count, shallow, retained) = if is_base {
+                (&mut diff.base_count, &mut diff.base_shallow, &mut diff.base_retained)
+            } else {
+                (&mut diff.count, &mut diff.shallow, &mut diff.retained)
+            };
+            *count = count.saturating_add(row.count);
+            *shallow = shallow.saturating_add(row.shallow);
+            *retained = retained.saturating_add(row.retained);
+            diff.ambiguous |= row.ambiguous;
+        }
+    }
+    let change = |d: &HeapClassDiff| if d.ambiguous { 0 } else { d.retained.abs_diff(d.base_retained) };
+    let mut diffs: Vec<HeapClassDiff> = by_name.into_values().collect();
+    diffs.sort_by(|a, b| change(b).cmp(&change(a)).then_with(|| a.name.cmp(&b.name)));
+    diffs
+}
+
+#[cfg(test)]
+mod tests;
@@ -0,0 +1,78 @@
+use super::*;
+use crate::hprof::HeapRootKind;
+use crate::test_index::analyzed;
+use crate::testing::{MiniHeap, sample, sample_with_two_more_nodes};
+
+fn diff<'a>(diffs: &'a [HeapClassDiff], name: &str) -> &'a HeapClassDiff {
+    let mut rows = diffs.iter().filter(|d| d.name == name);
+    let row = rows.next().unwrap_or_else(|| panic!("no diff for {name}"));
+    assert!(rows.next().is_none(), "two diffs for {name}");
+    row
+}
+
+fn sort_key(d: &HeapClassDiff) -> u64 {
+    if d.ambiguous { 0 } else { d.retained.abs_diff(d.base_retained) }
+}
+
+fn assert_sorted(diffs: &[HeapClassDiff]) {
+    assert!(diffs.windows(2).all(|p| (sort_key(&p[1]), &p[0].name) < (sort_key(&p[0]), &p[1].name)), "{diffs:#?}");
+}
+
+#[test]
+fn compare_counts_by_class() {
+    let base = analyzed(&sample());
+    let mut current = sample_with_two_more_nodes();
+    current
+        .class("Extra", Some("java.lang.Object"), &[])
+        .object(0x4000, "Extra", &[])
+        .root(HeapRootKind::JniGlobal, 0x4000);
+    let diffs = compare(&base, &analyzed(&current));
+    let node = diff(&diffs, "Node");
+    assert_eq!((node.base_count, node.count), (3, 5));
+    assert_eq!(node.shallow - node.base_shallow, 48);
+    assert_eq!((node.base_retained, node.retained), (104, 152));
+    assert!(!node.ambiguous);
+    let extra = diff(&diffs, "Extra");
+    assert_eq!((extra.base_count, extra.base_shallow, extra.base_retained, extra.count), (0, 0, 0, 1));
+    assert_eq!(diffs[0].name, "Node");
+    let holder = diff(&diffs, "Holder");
+    assert_eq!((holder.base_count, holder.count, holder.base_retained, holder.retained), (1, 1, 72, 72));
+    assert_sorted(&diffs);
+
+    // A class only in the base snapshot.
+    let gone = compare(&analyzed(&current), &base);
+    let extra = diff(&gone, "Extra");
+    assert_eq!((extra.base_count, extra.count, extra.shallow, extra.retained), (1, 0, 0, 0));
+    assert_sorted(&gone);
+}
+
+fn two_foos(extra: bool) -> MiniHeap {
+    let mut heap = sample();
+    heap.class("Foo", Some("java.lang.Object"), &[])
+        .object(0x3000, "Foo", &[])
+        .class("Foo", Some("java.lang.Object"), &[])
+        .object(0x3100, "Foo", &[]);
+    if extra {
+        heap.object(0x3200, "Foo", &[]);
+    }
+    for addr in [0x3000, 0x3100, 0x3200] {
+        heap.root(HeapRootKind::JniGlobal, addr);
+    }
+    heap
+}
+
+#[test]
+fn compare_gives_no_diff_for_a_name_of_several_classes() {
+    let diffs = compare(&analyzed(&two_foos(false)), &analyzed(&two_foos(true)));
+    let foo = diff(&diffs, "Foo");
+    assert!(foo.ambiguous);
+    assert_eq!((foo.base_count, foo.count), (2, 3));
+    assert!(diffs.iter().filter(|d| d.name != "Foo").all(|d| !d.ambiguous));
+    assert_sorted(&diffs);
+
+    // Ambiguous on one side only is still ambiguous.
+    let diffs = compare(&analyzed(&sample()), &analyzed(&two_foos(false)));
+    let foo = diff(&diffs, "Foo");
+    assert!(foo.ambiguous);
+    assert_eq!((foo.base_count, foo.count), (0, 2));
+}
@@ -0,0 +1,137 @@
+//! Disk-space estimates and the per-snapshot directories under the app's cache.
+//!
+//! A snapshot directory holds the index columns and, for a capture, the dump itself. Its `owner`
+//! file names the host process that created it, so a host that starts later can remove the
+//! directories of hosts that died without cleaning up.
+
+use std::fs;
+use std::io::{self, Read};
+use std::path::{Path, PathBuf};
+use std::time::{Duration, SystemTime};
+
+/// Space to leave free on top of an estimate.
+pub const DISK_MARGIN: u64 = 1 << 30;
+
+/// hprof ≈ `bytes + 16 × instances`; index columns and scratch arrays ≈ `90 × instances`.
+const CAPTURE_BYTES_PER_INSTANCE: u64 = 106;
+const INDEX_BYTES_PER_OBJECT: u64 = 90;
+
+const OWNER_FILE: &str = "owner";
+const TMP_SUFFIX: &str = ".tmp";
+/// A directory still being created by a live host has no `owner` for a moment; one that stays
+/// unfinished for longer than this was abandoned.
+const TMP_GRACE: Duration = Duration::from_secs(60 * 60);
+
+/// Bytes of free space that a dump of `heap_bytes` bytes holding `instances` instances needs,
+/// including its index.
+pub fn capture_need(heap_bytes: u64, instances: u64) -> u64 {
+    heap_bytes.saturating_add(CAPTURE_BYTES_PER_INSTANCE.saturating_mul(instances)).saturating_add(DISK_MARGIN)
+}
+
+/// Bytes of free space that indexing a dump of `objects` objects needs.
+pub fn index_need(objects: u64) -> u64 {
+    INDEX_BYTES_PER_OBJECT.saturating_mul(objects).saturating_add(DISK_MARGIN)
+}
+
+/// Bytes available to this process on the volume holding `dir`; `None` where it is not checked
+/// (not unix) or cannot be read, so callers go ahead rather than refuse.
+pub fn free_bytes(dir: &Path) -> Option<u64> {
+    #[cfg(unix)]
+    {
+        use std::os::unix::ffi::OsStrExt;
+
+        let path = std::ffi::CString::new(dir.as_os_str().as_bytes()).ok()?;
+        // SAFETY: `path` is NUL-terminated and `stat` is only read after `statvfs` filled it in.
+        let stat = unsafe {
+            let mut stat = std::mem::MaybeUninit::<libc::statvfs>::zeroed();
+            if libc::statvfs(path.as_ptr(), stat.as_mut_ptr()) != 0 {
+                return None;
+            }
+            stat.assume_init()
+        };
+        // The field widths differ between unix targets.
+        #[allow(clippy::unnecessary_cast)]
+        Some((stat.f_bavail as u64).saturating_mul(stat.f_frsize as u64))
+    }
+    #[cfg(not(unix))]
+    {
+        let _ = dir;
+        None
+    }
+}
+
+/// `<local_data_dir>/cache/heap-snapshots`.
+pub fn snapshots_root(local_data_dir: &Path) -> PathBuf {
+    local_data_dir.join("cache").join("heap-snapshots")
+}
+
+/// Creates `<root>/<uuid>/` holding an `owner` file with this process's pid. The directory is
+/// built as `<uuid>.tmp` and renamed, so another host scanning `root` never sees it without its
+/// owner and removes it as abandoned.
+pub fn create_snapshot_dir(root: &Path) -> io::Result<PathBuf> {
+    fs::create_dir_all(root)?;
+    let name = uuid::Uuid::new_v4().to_string();
+    let building = root.join(format!("{name}{TMP_SUFFIX}"));
+    let dir = root.join(&name);
+    create_private_dir(&building)?;
+    let published =
+        fs::write(building.join(OWNER_FILE), std::process::id().to_string()).and_then(|()| fs::rename(&building, &dir));
+    if let Err(err) = published {
+        let _ = fs::remove_dir_all(&building);
+        return Err(err);
+    }
+    Ok(dir)
+}
+
+/// A dump holds whatever was in the JVM's memory, passwords and tokens included, so on unix only
+/// the owner may enter the directory. Windows keeps the ACL inherited from the user's data dir.
+fn create_private_dir(path: &Path) -> io::Result<()> {
+    #[cfg(unix)]
+    {
+        use std::os::unix::fs::DirBuilderExt;
+        fs::DirBuilder::new().mode(0o700).create(path)
+    }
+    #[cfg(not(unix))]
+    {
+        fs::DirBuilder::new().create(path)
+    }
+}
+
+/// Removes the directories under `root` whose owner file is missing, unreadable or names a pid
+/// for which `alive` is false, and unfinished `*.tmp` ones older than an hour. Only directories
+/// directly under `root` are touched; a symlink there is left alone.
+pub fn remove_stale(root: &Path, alive: impl Fn(u32) -> bool) {
+    remove_stale_at(root, alive, SystemTime::now());
+}
+
+fn remove_stale_at(root: &Path, alive: impl Fn(u32) -> bool, now: SystemTime) {
+    let Ok(entries) = fs::read_dir(root) else { return };
+    for entry in entries.flatten() {
+        let path = entry.path();
+        // `symlink_metadata` does not follow: a link out of `root` is not a directory here.
+        let Ok(meta) = fs::symlink_metadata(&path) else { continue };
+        if !meta.file_type().is_dir() {
+            continue;
+        }
+        let stale = if path.extension().is_some_and(|ext| ext == "tmp") {
+            meta.modified()
+                .ok()
+                .and_then(|modified| now.duration_since(modified).ok())
+                .is_some_and(|age| age >= TMP_GRACE)
+        } else {
+            read_owner(&path).is_none_or(|pid| !alive(pid))
+        };
+        if stale && let Err(err) = fs::remove_dir_all(&path) {
+            tracing::warn!(dir = %path.display(), %err, "could not remove a stale heap snapshot directory");
+        }
+    }
+}
+
+fn read_owner(dir: &Path) -> Option<u32> {
+    let mut text = String::new();
+    fs::File::open(dir.join(OWNER_FILE)).ok()?.take(32).read_to_string(&mut text).ok()?;
+    text.trim().parse().ok()
+}
+
+#[cfg(test)]
+mod tests;
@@ -0,0 +1,137 @@
+use std::time::{Duration, SystemTime};
+
+use super::*;
+
+fn names(dir: &Path) -> Vec<String> {
+    let mut names: Vec<_> =
+        std::fs::read_dir(dir).unwrap().map(|e| e.unwrap().file_name().into_string().unwrap()).collect();
+    names.sort();
+    names
+}
+
+fn dir_with_owner(root: &Path, name: &str, owner: Option<&str>) {
+    let dir = root.join(name);
+    std::fs::create_dir(&dir).unwrap();
+    std::fs::write(dir.join("dump.hprof"), b"x").unwrap();
+    if let Some(owner) = owner {
+        std::fs::write(dir.join(OWNER_FILE), owner).unwrap();
+    }
+}
+
+#[test]
+fn estimates() {
+    assert_eq!(capture_need(664_018_520, 20_087_506), 664_018_520 + 106 * 20_087_506 + (1 << 30));
+    assert_eq!(index_need(10), 900 + (1 << 30));
+}
+
+#[test]
+fn estimates_saturate_instead_of_wrapping() {
+    assert_eq!(capture_need(u64::MAX, u64::MAX), u64::MAX);
+    assert_eq!(index_need(u64::MAX), u64::MAX);
+}
+
+#[cfg(unix)]
+#[test]
+fn not_enough_space_for_estimate() {
+    let dir = tempfile::tempdir().unwrap();
+    let free = free_bytes(dir.path()).expect("statvfs works on a temp dir");
+    assert!(free > 0);
+    assert!(capture_need(u64::MAX / 4, 0) > free);
+}
+
+#[cfg(unix)]
+#[test]
+fn free_bytes_of_a_missing_dir_is_unknown() {
+    let dir = tempfile::tempdir().unwrap();
+    assert_eq!(free_bytes(&dir.path().join("missing")), None);
+}
+
+#[cfg(not(unix))]
+#[test]
+fn free_space_is_not_checked_off_unix() {
+    assert_eq!(free_bytes(Path::new(".")), None);
+}
+
+#[test]
+fn snapshots_root_is_under_the_cache() {
+    let local = Path::new("local");
+    assert_eq!(snapshots_root(local), local.join("cache").join("heap-snapshots"));
+}
+
+#[test]
+fn snapshot_dir_records_its_owner_and_leaves_no_tmp() {
+    let root = tempfile::tempdir().unwrap();
+    let missing_parent = root.path().join("cache").join("heap-snapshots");
+    let dir = create_snapshot_dir(&missing_parent).unwrap();
+    assert_eq!(dir.parent(), Some(missing_parent.as_path()));
+    assert!(uuid::Uuid::parse_str(dir.file_name().unwrap().to_str().unwrap()).is_ok());
+    assert_eq!(std::fs::read_to_string(dir.join(OWNER_FILE)).unwrap(), std::process::id().to_string());
+    assert_eq!(names(&missing_parent), [dir.file_name().unwrap().to_str().unwrap()]);
+    assert_ne!(create_snapshot_dir(&missing_parent).unwrap(), dir);
+}
+
+#[cfg(unix)]
+#[test]
+fn snapshot_dir_is_private_to_the_owner() {
+    use std::os::unix::fs::PermissionsExt;
+
+    let root = tempfile::tempdir().unwrap();
+    let dir = create_snapshot_dir(root.path()).unwrap();
+    assert_eq!(std::fs::metadata(&dir).unwrap().permissions().mode() & 0o777, 0o700);
+}
+
+#[test]
+fn stale_dirs_removed_on_start() {
+    let root = tempfile::tempdir().unwrap();
+    dir_with_owner(root.path(), "a", Some("999999"));
+    dir_with_owner(root.path(), "b", Some(&std::process::id().to_string()));
+    dir_with_owner(root.path(), "c", None);
+    remove_stale(root.path(), |pid| pid == std::process::id());
+    assert_eq!(names(root.path()), ["b"]);
+}
+
+#[test]
+fn owner_that_is_not_a_pid_is_stale() {
+    let root = tempfile::tempdir().unwrap();
+    dir_with_owner(root.path(), "empty", Some(""));
+    dir_with_owner(root.path(), "word", Some("not-a-pid"));
+    dir_with_owner(root.path(), "negative", Some("-1"));
+    dir_with_owner(root.path(), "padded", Some(&format!(" {}\n", std::process::id())));
+    remove_stale(root.path(), |pid| pid == std::process::id());
+    assert_eq!(names(root.path()), ["padded"]);
+}
+
+#[test]
+fn stale_scan_tolerates_a_missing_root_and_leaves_loose_files_alone() {
+    let root = tempfile::tempdir().unwrap();
+    remove_stale(&root.path().join("missing"), |_| false);
+    std::fs::write(root.path().join("notes.txt"), b"keep").unwrap();
+    remove_stale(root.path(), |_| false);
+    assert_eq!(names(root.path()), ["notes.txt"]);
+}
+
+#[test]
+fn unfinished_dir_is_left_alone_for_an_hour() {
+    let root = tempfile::tempdir().unwrap();
+    // Another host may be between `create_dir` and the rename of its own snapshot dir.
+    dir_with_owner(root.path(), "young.tmp", None);
+    let now = SystemTime::now();
+    remove_stale_at(root.path(), |_| false, now);
+    assert_eq!(names(root.path()), ["young.tmp"]);
+    remove_stale_at(root.path(), |_| false, now + TMP_GRACE - Duration::from_secs(60));
+    assert_eq!(names(root.path()), ["young.tmp"]);
+    remove_stale_at(root.path(), |_| false, now + TMP_GRACE + Duration::from_secs(60));
+    assert!(names(root.path()).is_empty());
+}
+
+#[cfg(unix)]
+#[test]
+fn stale_scan_never_follows_a_symlink_out_of_the_root() {
+    let outside = tempfile::tempdir().unwrap();
+    std::fs::write(outside.path().join("precious"), b"x").unwrap();
+    let root = tempfile::tempdir().unwrap();
+    std::os::unix::fs::symlink(outside.path(), root.path().join("link")).unwrap();
+    remove_stale(root.path(), |_| false);
+    remove_stale_at(root.path(), |_| false, SystemTime::now() + 2 * TMP_GRACE);
+    assert_eq!(names(outside.path()), ["precious"]);
+}
@@ -0,0 +1,247 @@
+//! Dominator tree of the object graph, by Lengauer–Tarjan with path compression and no recursion:
+//! the DFS "stack" is the chain of DFS parents, and `eval` reverses the links of the path it
+//! compresses instead of keeping a stack.
+//!
+//! Between the DFS and the final translation every array is indexed by DFS number, so the root is
+//! 0 and a vertex's DFS ancestors all have smaller numbers.
+
+use ide_core::CancellationToken;
+
+use crate::column::{Column, NONE, Store};
+use crate::error::HeapError;
+use crate::files::IDOM_FILE;
+use crate::graph::accumulate;
+use crate::progress::{CHECK_EVERY, HeapPhase, Progress};
+
+const PHASE: HeapPhase = HeapPhase::Dominators;
+
+pub struct Dominators {
+    /// Immediate dominator of each node, saved as `idom.u32` in the store; [`NONE`] for
+    /// unreachable nodes, `root` for `root`.
+    pub idom: Column<u32>,
+    /// Nodes in DFS preorder from `root`, `reachable` entries.
+    pub preorder: Column<u32>,
+    pub reachable: u32,
+}
+
+/// The dominators of the graph whose node `v` has targets `out[out_start[v]..out_start[v + 1]]`,
+/// from `root` (which must be a node). `progress` gets `Dominators` from `(0, total)` to
+/// `(total, total)`; an `Err` from it stops the computation with that error.
+pub fn dominators(
+    out_start: &[u32],
+    out: &[u32],
+    root: u32,
+    store: &Store,
+    progress: &dyn Fn(HeapPhase, u64, u64) -> Result<(), HeapError>,
+    cancel: &CancellationToken,
+) -> Result<Dominators, HeapError> {
+    compute(out_start, out, root, store, progress, cancel).map(|(dominators, _)| dominators)
+}
+
+/// [`dominators`], plus the number of steps of work it took.
+fn compute(
+    out_start: &[u32],
+    out: &[u32],
+    root: u32,
+    store: &Store,
+    progress: &dyn Fn(HeapPhase, u64, u64) -> Result<(), HeapError>,
+    cancel: &CancellationToken,
+) -> Result<(Dominators, u64), HeapError> {
+    let nodes = out_start.len().saturating_sub(1);
+    if nodes >= NONE as usize {
+        return Err(HeapError::TooLarge);
+    }
+    assert!((root as usize) < nodes, "root {root} is not one of the {nodes} nodes");
+    // Three walks of at most `nodes` vertices each: DFS, semidominators, immediate dominators.
+    let span = nodes as u64;
+    let mut work = Work::new(Progress::new(progress, cancel), 3 * span);
+    work.report(0)?;
+
+    // DFS: `pre` maps nodes to DFS numbers and `vertex` back; `cursor[d]` is the next edge of
+    // DFS vertex d to follow.
+    let mut pre = store.scratch::<u32>(nodes, NONE)?;
+    let mut vertex = store.scratch::<u32>(nodes, 0)?;
+    let mut parent = store.scratch::<u32>(nodes, 0)?;
+    let mut cursor = store.scratch::<u32>(nodes, 0)?;
+    pre[root as usize] = 0;
+    vertex[0] = root;
+    cursor[0] = out_start[root as usize];
+    let (mut reachable, mut d) = (1u32, 0u32);
+    loop {
+        work.tick()?;
+        let edge = cursor[d as usize];
+        if edge < out_start[vertex[d as usize] as usize + 1] {
+            cursor[d as usize] = edge + 1;
+            let w = out[edge as usize] as usize;
+            if pre[w] == NONE {
+                let next = reachable as usize;
+                pre[w] = reachable;
+                vertex[next] = w as u32;
+                parent[next] = d;
+                cursor[next] = out_start[w];
+                d = reachable;
+                reachable += 1;
+                work.done = u64::from(reachable);
+            }
+        } else if d == 0 {
+            break;
+        } else {
+            d = parent[d as usize];
+        }
+    }
+    let r = reachable as usize;
+    work.report(span)?;
+
+    // Predecessors by DFS number, as CSR: counted into `starts[w + 1]`, then filled through
+    // `starts[w]` as a cursor, which leaves it at w's end until the column is shifted one slot.
+    let mut semi = cursor;
+    let mut label = store.scratch::<u32>(r, 0)?;
+    let mut starts = store.scratch::<u32>(r + 1, 0)?;
+    for d in 0..r {
+        work.tick()?;
+        semi[d] = d as u32;
+        label[d] = d as u32;
+        for &w in targets(out_start, out, vertex[d]) {
+            work.tick()?;
+            starts[pre[w as usize] as usize + 1] += 1;
+        }
+    }
+    let edges = accumulate(&mut starts, cancel)?;
+    let mut preds = store.scratch::<u32>(edges as usize, 0)?;
+    for d in 0..r {
+        work.tick()?;
+        for &w in targets(out_start, out, vertex[d]) {
+            work.tick()?;
+            let slot = &mut starts[pre[w as usize] as usize];
+            preds[*slot as usize] = d as u32;
+            *slot += 1;
+        }
+    }
+    starts.copy_within(0..r, 1);
+    starts[0] = 0;
+    drop(pre);
+    work.report(span)?;
+
+    // Semidominators in reverse preorder, linking each finished vertex to its parent. Vertex w
+    // waits in its semidominator's bucket until a child of that semidominator is linked; then
+    // `eval` gives w's idom, or a vertex whose idom is also w's, fixed up in preorder below.
+    // `idom[v]` is the next vertex in v's bucket until v leaves the bucket.
+    let mut ancestor = store.scratch::<u32>(r, NONE)?;
+    let mut bucket = store.scratch::<u32>(r, NONE)?;
+    let mut idom = store.scratch::<u32>(r, NONE)?;
+    for w in (1..r).rev() {
+        work.done = span + (r - w) as u64;
+        work.tick()?;
+        let mut s = semi[w];
+        for &v in &preds[starts[w] as usize..starts[w + 1] as usize] {
+            work.tick()?;
+            let u = eval(v, &mut ancestor, &mut label, &semi, &mut work)?;
+            s = s.min(semi[u as usize]);
+        }
+        semi[w] = s;
+        idom[w] = bucket[s as usize];
+        bucket[s as usize] = w as u32;
+        let p = parent[w];
+        ancestor[w] = p;
+        let mut v = std::mem::replace(&mut bucket[p as usize], NONE);
+        while v != NONE {
+            work.tick()?;
+            let next = idom[v as usize];
+            let u = eval(v, &mut ancestor, &mut label, &semi, &mut work)?;
+            idom[v as usize] = if semi[u as usize] < semi[v as usize] { u } else { p };
+            v = next;
+        }
+    }
+    drop((label, starts, preds, ancestor, bucket, parent));
+    work.report(2 * span)?;
+
+    // Immediate dominators, translated back to nodes; in preorder, so a deferred vertex's
+    // stand-in already has its final idom.
+    let mut node_idom = store.column::<u32>(IDOM_FILE, nodes, NONE)?;
+    let mut preorder = store.scratch::<u32>(r, 0)?;
+    idom[0] = 0;
+    node_idom[root as usize] = root;
+    preorder[0] = root;
+    for w in 1..r {
+        work.done = 2 * span + w as u64;
+        work.tick()?;
+        if idom[w] != semi[w] {
+            idom[w] = idom[idom[w] as usize];
+        }
+        node_idom[vertex[w] as usize] = vertex[idom[w] as usize];
+        preorder[w] = vertex[w];
+    }
+    work.report(3 * span)?;
+    Ok((Dominators { idom: node_idom, preorder, reachable }, work.steps))
+}
+
+fn targets<'a>(out_start: &[u32], out: &'a [u32], v: u32) -> &'a [u32] {
+    let v = v as usize;
+    &out[out_start[v] as usize..out_start[v + 1] as usize]
+}
+
+/// Counts steps of work and reports progress, which also looks at the cancellation token, at
+/// every phase boundary and after every [`CHECK_EVERY`] steps, whichever loop they are in.
+struct Work<'a> {
+    progress: Progress<'a>,
+    total: u64,
+    /// Progress to report at the next check; the loops keep it current.
+    done: u64,
+    steps: u64,
+    next_report: u64,
+}
+
+impl<'a> Work<'a> {
+    fn new(progress: Progress<'a>, total: u64) -> Self {
+        Self { progress, total, done: 0, steps: 0, next_report: CHECK_EVERY }
+    }
+
+    fn report(&mut self, done: u64) -> Result<(), HeapError> {
+        self.done = done;
+        self.next_report = self.steps + CHECK_EVERY;
+        self.progress.report(PHASE, done, self.total)
+    }
+
+    #[inline]
+    fn tick(&mut self) -> Result<(), HeapError> {
+        self.steps += 1;
+        if self.steps < self.next_report { Ok(()) } else { self.report(self.done) }
+    }
+}
+
+/// The vertex of smallest semidominator on the forest path from `v` up to, not including, its
+/// tree's root (`v` itself when it is a root), compressing that path so the root becomes the
+/// ancestor of every vertex on it. The walk up reverses the `ancestor` links so the walk down can
+/// fold labels nearest-root first, as the recursive compress does, without a stack.
+fn eval(v: u32, ancestor: &mut [u32], label: &mut [u32], semi: &[u32], work: &mut Work) -> Result<u32, HeapError> {
+    let at = |x: u32| x as usize;
+    if ancestor[at(v)] == NONE {
+        return Ok(v);
+    }
+    let (mut x, mut below) = (v, NONE);
+    while ancestor[at(ancestor[at(x)])] != NONE {
+        work.tick()?;
+        let up = ancestor[at(x)];
+        ancestor[at(x)] = below;
+        below = x;
+        x = up;
+    }
+    // `x` is the root's child, already compressed; `below` is the vertex under it.
+    let root = ancestor[at(x)];
+    let mut best = label[at(x)];
+    while below != NONE {
+        work.tick()?;
+        let y = at(below);
+        if semi[at(best)] < semi[at(label[y])] {
+            label[y] = best;
+        } else {
+            best = label[y];
+        }
+        below = ancestor[y];
+        ancestor[y] = root;
+    }
+    Ok(label[at(v)])
+}
+
+#[cfg(test)]
+mod tests;
@@ -0,0 +1,348 @@
+use std::cell::RefCell;
+use std::time::Instant;
+
+use proptest::collection::vec;
+use proptest::prelude::*;
+
+use super::*;
+use crate::NONE;
+use crate::graph::build_graph;
+use crate::testing::{XorShift, sample};
+
+fn no_progress(_: HeapPhase, _: u64, _: u64) -> Result<(), HeapError> {
+    Ok(())
+}
+
+/// CSR of `edges` over `n` nodes, targets in edge order.
+fn csr(n: usize, edges: &[(u32, u32)]) -> (Vec<u32>, Vec<u32>) {
+    let mut start = vec![0u32; n + 1];
+    for &(from, _) in edges {
+        start[from as usize + 1] += 1;
+    }
+    for v in 0..n {
+        start[v + 1] += start[v];
+    }
+    let mut cursor = start.clone();
+    let mut out = vec![0u32; edges.len()];
+    for &(from, to) in edges {
+        out[cursor[from as usize] as usize] = to;
+        cursor[from as usize] += 1;
+    }
+    (start, out)
+}
+
+fn run(n: usize, edges: &[(u32, u32)], root: u32) -> Dominators {
+    let (start, out) = csr(n, edges);
+    dominators(&start, &out, root, &Store::Ram, &no_progress, &CancellationToken::new()).unwrap()
+}
+
+fn preds(n: usize, edges: &[(u32, u32)]) -> Vec<Vec<u32>> {
+    let mut preds = vec![Vec::new(); n];
+    for &(from, to) in edges {
+        preds[to as usize].push(from);
+    }
+    preds
+}
+
+/// Immediate dominators from iterated `Dom` sets (one `u64` bitset per node, `n <= 64`).
+fn dataflow_idom(n: usize, edges: &[(u32, u32)], root: u32) -> Vec<u32> {
+    let (start, out) = csr(n, edges);
+    let mut reachable = vec![false; n];
+    let mut todo = vec![root];
+    reachable[root as usize] = true;
+    while let Some(v) = todo.pop() {
+        for &w in &out[start[v as usize] as usize..start[v as usize + 1] as usize] {
+            if !std::mem::replace(&mut reachable[w as usize], true) {
+                todo.push(w);
+            }
+        }
+    }
+    let preds = preds(n, edges);
+    let mut dom = vec![u64::MAX; n];
+    dom[root as usize] = 1 << root;
+    let mut changed = true;
+    while changed {
+        changed = false;
+        for v in (0..n).filter(|&v| reachable[v] && v != root as usize) {
+            let meet =
+                preds[v].iter().filter(|&&p| reachable[p as usize]).fold(u64::MAX, |acc, &p| acc & dom[p as usize]);
+            let next = meet | 1 << v;
+            if next != dom[v] {
+                dom[v] = next;
+                changed = true;
+            }
+        }
+    }
+    (0..n)
+        .map(|v| {
+            if !reachable[v] {
+                NONE
+            } else if v == root as usize {
+                root
+            } else {
+                // The strict dominator dominated by all the others has the most dominators itself.
+                let strict = dom[v] & !(1 << v);
+                (0..n as u32).filter(|&d| strict & 1 << d != 0).max_by_key(|&d| dom[d as usize].count_ones()).unwrap()
+            }
+        })
+        .collect()
+}
+
+/// Cooper–Harvey–Kennedy iteration over reverse postorder from an explicit-stack DFS.
+fn chk_idom(n: usize, edges: &[(u32, u32)], root: u32) -> Vec<u32> {
+    let (start, out) = csr(n, edges);
+    let mut post_number = vec![NONE; n];
+    let mut postorder = Vec::new();
+    let mut seen = vec![false; n];
+    let mut stack = vec![(root, start[root as usize])];
+    seen[root as usize] = true;
+    while let Some(&mut (v, ref mut next)) = stack.last_mut() {
+        if *next < start[v as usize + 1] {
+            let w = out[*next as usize];
+            *next += 1;
+            if !std::mem::replace(&mut seen[w as usize], true) {
+                stack.push((w, start[w as usize]));
+            }
+        } else {
+            post_number[v as usize] = postorder.len() as u32;
+            postorder.push(v);
+            stack.pop();
+        }
+    }
+    let preds = preds(n, edges);
+    let mut idom = vec![NONE; n];
+    idom[root as usize] = root;
+    let intersect = |idom: &[u32], mut a: u32, mut b: u32| {
+        while a != b {
+            while post_number[a as usize] < post_number[b as usize] {
+                a = idom[a as usize];
+            }
+            while post_number[b as usize] < post_number[a as usize] {
+                b = idom[b as usize];
+            }
+        }
+        a
+    };
+    let mut changed = true;
+    while changed {
+        changed = false;
+        for &v in postorder.iter().rev().filter(|&&v| v != root) {
+            let mut new = NONE;
+            for &p in &preds[v as usize] {
+                if idom[p as usize] != NONE {
+                    new = if new == NONE { p } else { intersect(&idom, p, new) };
+                }
+            }
+            if new != idom[v as usize] {
+                idom[v as usize] = new;
+                changed = true;
+            }
+        }
+    }
+    idom
+}
+
+#[test]
+fn diamond() {
+    let d = run(5, &[(0, 1), (0, 2), (1, 3), (2, 3), (3, 4)], 0);
+    assert_eq!(&*d.idom, &[0, 0, 0, 0, 3]);
+    assert_eq!(d.reachable, 5);
+    assert_eq!(&*d.preorder, &[0, 1, 3, 4, 2]);
+}
+
+#[test]
+fn unreachable_is_none() {
+    let d = run(6, &[(0, 1), (0, 2), (1, 3), (2, 3), (3, 4), (5, 3)], 0);
+    assert_eq!(&*d.idom, &[0, 0, 0, 0, 3, NONE]);
+    assert_eq!(d.reachable, 5);
+    assert_eq!(d.preorder.len(), 5);
+}
+
+#[test]
+fn cycle_back_to_root() {
+    let d = run(3, &[(0, 1), (1, 2), (2, 1), (2, 0)], 0);
+    assert_eq!(&*d.idom, &[0, 0, 1]);
+}
+
+#[test]
+fn root_need_not_be_node_zero() {
+    let d = run(4, &[(3, 1), (1, 0), (3, 2), (2, 0), (0, 3)], 3);
+    assert_eq!(&*d.idom, &[3, 3, 3, 3]);
+    assert_eq!(d.preorder[0], 3);
+}
+
+#[test]
+fn sample_heap_dominators() {
+    let bytes = sample().bytes();
+    let g = build_graph(&bytes, &Store::Ram, &no_progress, &CancellationToken::new()).unwrap();
+    let d = dominators(&g.out_start, &g.out, g.root, &Store::Ram, &no_progress, &CancellationToken::new()).unwrap();
+    let node = |id: u64| g.ids.rank(id).unwrap();
+    let idom_of = |id: u64| d.idom[node(id) as usize];
+    assert_eq!(idom_of(0x1100), node(0x1000), "N1 → H");
+    assert_eq!(idom_of(0x1200), node(0x1000), "N2 → H");
+    assert_eq!(idom_of(0x1300), g.root, "N3 → root");
+    assert_eq!(idom_of(0x1400), node(0x1300), "L → N3");
+    assert_eq!(idom_of(0x1500), NONE, "R is unreachable");
+    assert_eq!(idom_of(0x1700), NONE, "N4 only through the weak referent");
+    assert_eq!(d.idom[g.root as usize], g.root);
+}
+
+#[test]
+fn idom_is_written_to_the_store() {
+    let dir = tempfile::tempdir().unwrap();
+    let (start, out) = csr(3, &[(0, 1), (1, 2)]);
+    let store = Store::Dir(dir.path().to_path_buf());
+    let d = dominators(&start, &out, 0, &store, &no_progress, &CancellationToken::new()).unwrap();
+    drop(d);
+    assert_eq!(&*Column::<u32>::open(&dir.path().join("idom.u32")).unwrap(), &[0, 0, 1]);
+    let names: Vec<_> = std::fs::read_dir(dir.path()).unwrap().map(|e| e.unwrap().file_name()).collect();
+    assert_eq!(names, ["idom.u32"], "scratch columns leave no files");
+}
+
+#[test]
+fn reports_dominators_progress_and_stops_on_cancel() {
+    let (start, out) = csr(3, &[(0, 1), (1, 2)]);
+    let calls = RefCell::new(Vec::new());
+    let record = |phase, done, total| {
+        calls.borrow_mut().push((phase, done, total));
+        Ok(())
+    };
+    dominators(&start, &out, 0, &Store::Ram, &record, &CancellationToken::new()).unwrap();
+    let calls = calls.into_inner();
+    let total = calls[0].2;
+    assert_eq!(calls.first(), Some(&(HeapPhase::Dominators, 0, total)));
+    assert_eq!(calls.last(), Some(&(HeapPhase::Dominators, total, total)));
+    assert!(calls.windows(2).all(|w| w[0].1 <= w[1].1), "progress never goes back: {calls:?}");
+
+    let cancel = CancellationToken::new();
+    let cancel_on_first = |_, _, _| {
+        cancel.cancel();
+        Ok(())
+    };
+    let result = dominators(&start, &out, 0, &Store::Ram, &cancel_on_first, &cancel);
+    assert!(matches!(result, Err(HeapError::Cancelled)));
+
+    let refuse = |_, _, _| Err(HeapError::TooLarge);
+    let result = dominators(&start, &out, 0, &Store::Ram, &refuse, &CancellationToken::new());
+    assert!(matches!(result, Err(HeapError::TooLarge)));
+}
+
+#[test]
+fn cancel_during_idom_phase() {
+    // Long enough that the last phase alone passes a check interval.
+    const NODES: u32 = (CHECK_EVERY + CHECK_EVERY / 2) as u32;
+    let edges: Vec<_> = (1..NODES).map(|v| (v - 1, v)).collect();
+    let (start, out) = csr(NODES as usize, &edges);
+    let span = u64::from(NODES);
+    let cancel = CancellationToken::new();
+    let calls = RefCell::new(Vec::new());
+    let cancel_past_semidominators = |_, done, _| {
+        calls.borrow_mut().push(done);
+        if done > 2 * span {
+            cancel.cancel();
+        }
+        Ok(())
+    };
+    let result = dominators(&start, &out, 0, &Store::Ram, &cancel_past_semidominators, &cancel);
+    assert!(matches!(result, Err(HeapError::Cancelled)));
+    let calls = calls.into_inner();
+    assert!(calls.contains(&(2 * span)), "the semidominator phase finished: {calls:?}");
+    let last = *calls.last().unwrap();
+    assert!(last > 2 * span && last < 3 * span, "stopped inside the idom phase: {calls:?}");
+}
+
+proptest! {
+    #[test]
+    fn matches_dataflow_on_random_graphs(n in 2usize..64, edges in vec((0u32..64, 0u32..64), 0..192)) {
+        let edges: Vec<_> = edges.into_iter().map(|(a, b)| (a % n as u32, b % n as u32)).collect();
+        let d = run(n, &edges, 0);
+        prop_assert_eq!(d.idom.to_vec(), dataflow_idom(n, &edges, 0));
+        prop_assert_eq!(d.preorder.len(), d.reachable as usize);
+    }
+}
+
+#[test]
+fn matches_chk_on_a_large_random_graph() {
+    const NODES: u32 = 200_000;
+    const EXTRA_EDGES: u32 = 100_000;
+    let mut rng = XorShift::new(0x5EED_D0D0);
+    let mut edges: Vec<_> = (1..NODES).map(|v| (rng.below(v), v)).collect();
+    edges.extend((0..EXTRA_EDGES).map(|_| (rng.below(NODES), rng.below(NODES))));
+    // Visit edges in a scrambled order so the DFS tree differs from the generating tree.
+    for i in (1..edges.len()).rev() {
+        edges.swap(i, rng.below(i as u32 + 1) as usize);
+    }
+    let started = Instant::now();
+    let d = run(NODES as usize, &edges, 0);
+    let elapsed = started.elapsed();
+    let expected = chk_idom(NODES as usize, &edges, 0);
+    let mismatches = d.idom.iter().zip(&expected).filter(|(a, b)| a != b).count();
+    assert_eq!(mismatches, 0);
+    assert_eq!(d.reachable, NODES);
+    eprintln!("dominators of {NODES} random nodes: {elapsed:?}");
+}
+
+/// A doubly linked list whose items an array list also holds: super-root 0 → list 1 → node 0,
+/// 0 → array list 2 → element array 3 → every item; node j → item j, next, prev.
+fn shared_list_and_array(k: u32) -> (usize, Vec<(u32, u32)>) {
+    let node = |j: u32| 4 + 2 * j;
+    let item = |j: u32| 5 + 2 * j;
+    let mut edges = vec![(0, 1), (0, 2), (2, 3), (1, node(0))];
+    for j in 0..k {
+        edges.push((node(j), item(j)));
+        if j + 1 < k {
+            edges.push((node(j), node(j + 1)));
+        }
+        if j > 0 {
+            edges.push((node(j), node(j - 1)));
+        }
+        edges.push((3, item(j)));
+    }
+    ((4 + 2 * k) as usize, edges)
+}
+
+#[test]
+fn shared_list_and_array_is_near_linear() {
+    let (n, edges) = shared_list_and_array(29);
+    assert_eq!(run(n, &edges, 0).idom.to_vec(), dataflow_idom(n, &edges, 0));
+    let (n, edges) = shared_list_and_array(3_000);
+    assert_eq!(run(n, &edges, 0).idom.to_vec(), chk_idom(n, &edges, 0));
+
+    const K: u32 = 200_000;
+    let (n, edges) = shared_list_and_array(K);
+    let (start, out) = csr(n, &edges);
+    let started = Instant::now();
+    let (d, steps) = compute(&start, &out, 0, &Store::Ram, &no_progress, &CancellationToken::new()).unwrap();
+    let elapsed = started.elapsed();
+    let item = |j: u32| (5 + 2 * j) as usize;
+    assert!((0..K).all(|j| d.idom[item(j)] == 0), "every item is held by both collections");
+    let m = edges.len() as u64;
+    let bound = (n as u64 + m) * u64::from(n.ilog2());
+    eprintln!("shared list k={K}: {steps} steps (n={n}, m={m}, bound {bound}), {elapsed:?}");
+    assert!(steps <= bound, "{steps} steps for n={n}, m={m}");
+}
+
+#[test]
+fn deep_chain_has_linear_dominators() {
+    // A small stack proves the algorithm does not recurse along the chain.
+    std::thread::Builder::new()
+        .stack_size(256 << 10)
+        .spawn(|| {
+            const NODES: u32 = 1_000_000;
+            let mut edges: Vec<_> = (1..NODES).map(|v| (v - 1, v)).collect();
+            edges.push((0, NODES - 1));
+            // Makes `eval` compress the whole chain at once.
+            edges.push((NODES - 1, 1));
+            let started = Instant::now();
+            let d = run(NODES as usize, &edges, 0);
+            eprintln!("dominators of a {NODES}-node chain: {:?}", started.elapsed());
+            assert_eq!(d.reachable, NODES);
+            assert_eq!(d.idom[0], 0);
+            assert_eq!(d.idom[NODES as usize - 1], 0);
+            let wrong = (1..NODES - 1).filter(|&v| d.idom[v as usize] != v - 1).count();
+            assert_eq!(wrong, 0);
+        })
+        .unwrap()
+        .join()
+        .unwrap();
+}
@@ -0,0 +1,26 @@
+/// Offsets (`at`) are byte positions in the `.hprof` file.
+#[derive(Debug, thiserror::Error)]
+pub enum HeapError {
+    #[error("the dump ends early at byte {at}")]
+    Truncated { at: u64 },
+    #[error("not a heap dump (no JAVA PROFILE header)")]
+    NotHprof,
+    #[error("unsupported identifier size {0}")]
+    UnsupportedIdSize(u32),
+    #[error("unknown heap sub-record tag {tag:#04x} at byte {at}")]
+    UnknownSubRecord { tag: u8, at: u64 },
+    #[error("unknown value type {ty} at byte {at}")]
+    UnknownType { ty: u8, at: u64 },
+    #[error("class {id:#x} used at byte {at} is not in the dump")]
+    MissingClass { id: u64, at: u64 },
+    #[error("the heap has too many objects to analyze")]
+    TooLarge,
+    #[error("the dump's class and field names are too long; it is not a dump a JVM wrote")]
+    NamesTooLarge,
+    #[error("not enough disk space: {needed} bytes needed, {available} free")]
+    NotEnoughDiskSpace { needed: u64, available: u64 },
+    #[error("cancelled")]
+    Cancelled,
+    #[error("{0}")]
+    Io(#[from] std::io::Error),
+}
@@ -0,0 +1,34 @@
+//! Names of the files an analysis writes into its index directory. They live in one list so a
+//! failed analysis removes exactly these and never the dump or anything else sharing the directory.
+
+pub(crate) const IDS_BITS_FILE: &str = "ids.bits.u64";
+pub(crate) const IDS_WRANK_FILE: &str = "ids.wrank.u32";
+pub(crate) const IDS_TABLE_FILE: &str = "ids.table.u32";
+pub(crate) const IDS_CHUNKS_FILE: &str = "ids.chunks.u32";
+pub(crate) const IDS_SORTED_FILE: &str = "ids.sorted.u64";
+pub(crate) const TYPES_FILE: &str = "types.u32";
+pub(crate) const OFFSETS_FILE: &str = "offsets.u64";
+pub(crate) const IDOM_FILE: &str = "idom.u32";
+pub(crate) const SHALLOW8_FILE: &str = "shallow8.u32";
+pub(crate) const RETAINED_FILE: &str = "retained.u64";
+pub(crate) const KIDS_START_FILE: &str = "kids_start.u32";
+pub(crate) const KIDS_FILE: &str = "kids.u32";
+pub(crate) const GC_PARENT_FILE: &str = "gc_parent.u32";
+pub(crate) const SUMMARY_FILE: &str = "summary.json";
+
+pub(crate) const INDEX_FILES: &[&str] = &[
+    IDS_BITS_FILE,
+    IDS_WRANK_FILE,
+    IDS_TABLE_FILE,
+    IDS_CHUNKS_FILE,
+    IDS_SORTED_FILE,
+    TYPES_FILE,
+    OFFSETS_FILE,
+    IDOM_FILE,
+    SHALLOW8_FILE,
+    RETAINED_FILE,
+    KIDS_START_FILE,
+    KIDS_FILE,
+    GC_PARENT_FILE,
+    SUMMARY_FILE,
+];
@@ -0,0 +1,274 @@
+//! The object graph of a heap dump in CSR form: one node per object, plus a super-root.
+
+mod classes;
+mod link;
+
+use std::collections::hash_map::Entry;
+use std::collections::{HashMap, HashSet};
+
+use ide_core::CancellationToken;
+
+use self::classes::{ClassTable, IdHash};
+use self::link::Linker;
+use crate::column::{Column, Store};
+use crate::error::HeapError;
+use crate::files::{OFFSETS_FILE, TYPES_FILE};
+use crate::hprof::{BasicType, HeapRecord, HeapRootKind, Record, parse_header, scan, scan_segment};
+use crate::ids::{IdMap, IdMapBuilder};
+use crate::progress::{CHECK_EVERY, HeapPhase, Progress};
+
+/// Number of primitive array type slots, one per non-object [`BasicType`].
+pub const PRIM_SLOTS: u32 = 8;
+
+pub struct ClassInfo {
+    pub id: u64,
+    /// The hprof name with `/` turned into `.` (`java.lang.String`, `[Ljava.lang.Object;`).
+    pub name: String,
+    /// Index in [`HeapGraph::classes`]. Chains always end: a super cycle in the dump is cut where it closes.
+    pub super_index: Option<u32>,
+    /// This class's own instance fields, in dump order.
+    pub fields: Vec<(String, BasicType)>,
+    /// `(name, object id)` of reference statics; a null static has object id 0.
+    pub statics: Vec<(String, u64)>,
+    pub static_bytes: u64,
+    /// Index in `fields` of `java.lang.ref.Reference.referent`, on that class only.
+    pub weak_field: Option<usize>,
+}
+
+pub struct HeapGraph {
+    pub ids: IdMap,
+    /// Type slot of each object: class index for instances and object arrays,
+    /// [`Self::class_slot`] for class objects, [`Self::prim_slot`] for primitive arrays.
+    pub types: Column<u32>,
+    /// Array length, 0 for instances and classes.
+    pub lens: Column<u32>,
+    /// File offset of each object's record, for [`crate::record_at`].
+    pub offsets: Column<u64>,
+    /// `n + 2` entries: node `v`'s targets are `out[out_start[v]..out_start[v + 1]]`.
+    pub out_start: Column<u32>,
+    pub out: Column<u32>,
+    /// The super-root, node `n`, with an edge to every rooted object.
+    pub root: u32,
+    /// `(object, kind, thread serial)` of every root record whose object is in the dump.
+    pub roots: Vec<(u32, HeapRootKind, Option<u32>)>,
+    /// Class dumps in dump order, then a synthetic `java.lang.Class` when the dump has none.
+    pub classes: Vec<ClassInfo>,
+    /// Type slot of class objects (`java.lang.Class`).
+    pub class_slot: u32,
+    /// Thread serial → thread object, from `ROOT THREAD OBJECT` records.
+    pub thread_objects: HashMap<u32, u32>,
+    /// `Reference.referent` edges left out of the graph.
+    pub weak_edges_skipped: u64,
+    pub id_size: u8,
+    pub timestamp_ms: u64,
+}
+
+impl HeapGraph {
+    pub fn prim_slot(&self, ty: BasicType) -> u32 {
+        self.classes.len() as u32 + (ty as u32 - BasicType::Boolean as u32)
+    }
+
+    /// The element type of primitive array slot `slot`, or `None` for a class slot.
+    pub fn prim_type(&self, slot: u32) -> Option<BasicType> {
+        let offset = slot.checked_sub(self.classes.len() as u32).filter(|&o| o < PRIM_SLOTS)?;
+        BasicType::from_u8(offset as u8 + BasicType::Boolean as u8)
+    }
+
+    pub fn out_edges(&self, v: u32) -> &[u32] {
+        let v = v as usize;
+        &self.out[self.out_start[v] as usize..self.out_start[v + 1] as usize]
+    }
+}
+
+/// Builds the graph of the dump in `buf`. Strong edges only: `Reference.referent` is counted in
+/// [`HeapGraph::weak_edges_skipped`] instead. `progress` gets `Reading` by bytes, then
+/// `Linking` by objects, starting with `(Linking, 0, objects)` once the object count is known;
+/// an `Err` from it stops the build with that error.
+pub fn build_graph(
+    buf: &[u8],
+    store: &Store,
+    progress: &dyn Fn(HeapPhase, u64, u64) -> Result<(), HeapError>,
+    cancel: &CancellationToken,
+) -> Result<HeapGraph, HeapError> {
+    let progress = Progress::new(progress, cancel);
+    let header = parse_header(buf)?;
+    let id_size = header.id_size;
+    let len = buf.len() as u64;
+
+    // Pass 1: names, classes, roots and every object id.
+    progress.report(HeapPhase::Reading, 0, len)?;
+    let mut names: IdHash<&[u8]> = IdHash::default();
+    let mut class_names: IdHash<u64> = IdHash::default();
+    let mut dumps = Vec::new();
+    let mut class_by_id: IdHash<u32> = IdHash::default();
+    let mut raw_roots = Vec::new();
+    let mut segments = Vec::new();
+    let mut builder = IdMapBuilder::new();
+    let (mut step, mut pos) = (0, header.body_start as u64);
+    scan(buf, &header, &mut |record| {
+        progress.step(step, HeapPhase::Reading, pos, len)?;
+        step += 1;
+        match record {
+            Record::Utf8 { id, text } => {
+                names.insert(id, text);
+            }
+            Record::LoadClass { class_id, name_id } => {
+                class_names.insert(class_id, name_id);
+            }
+            Record::Segment(range) => {
+                pos = range.start as u64;
+                segments.push(range);
+            }
+            Record::Heap { at, record } => {
+                pos = at as u64;
+                match record {
+                    HeapRecord::Root { kind, id, thread_serial } => raw_roots.push((id, kind, thread_serial)),
+                    HeapRecord::Class(dump) => {
+                        builder.insert(dump.id);
+                        if let Entry::Vacant(slot) = class_by_id.entry(dump.id) {
+                            slot.insert(dumps.len() as u32);
+                            dumps.push(dump);
+                        }
+                    }
+                    HeapRecord::Instance { id, .. }
+                    | HeapRecord::ObjectArray { id, .. }
+                    | HeapRecord::PrimitiveArray { id, .. } => builder.insert(id),
+                }
+            }
+        }
+        Ok(())
+    })?;
+    progress.report(HeapPhase::Reading, len, len)?;
+
+    let ids = builder.build(store, cancel)?;
+    let table = classes::resolve(dumps, class_by_id, &names, &class_names, id_size, buf.len())?;
+    drop((names, class_names));
+    let n = ids.len();
+    let objects = u64::from(n);
+    progress.report(HeapPhase::Linking, 0, objects)?;
+
+    let mut roots = Vec::new();
+    let mut root_targets = Vec::new();
+    let mut thread_objects = HashMap::new();
+    let (mut seen_records, mut seen_targets) = (HashSet::new(), HashSet::new());
+    for (i, &(id, kind, serial)) in raw_roots.iter().enumerate() {
+        progress.step(i as u64, HeapPhase::Linking, 0, objects)?;
+        let Some(object) = ids.rank(id).filter(|_| id != 0) else { continue };
+        if seen_records.insert((object, kind as u8, serial)) {
+            roots.push((object, kind, serial));
+        }
+        if seen_targets.insert(object) {
+            root_targets.push(object);
+        }
+        if let (HeapRootKind::ThreadObject, Some(serial)) = (kind, serial) {
+            thread_objects.entry(serial).or_insert(object);
+        }
+    }
+    drop((raw_roots, seen_records, seen_targets));
+
+    let linker = Linker::new(&ids, &table, id_size);
+    let nodes = n as usize;
+    let mut types = store.column::<u32>(TYPES_FILE, nodes, 0)?;
+    let mut lens = store.scratch::<u32>(nodes, 0)?;
+    let mut offsets = store.column::<u64>(OFFSETS_FILE, nodes, 0)?;
+    let mut out_start = store.scratch::<u32>(nodes + 2, 0)?;
+    // Both passes visit every object record once; report them as one walk over the objects.
+    let mut visited = 0u64;
+    let tick = |visited: &mut u64| {
+        let result = progress.step(*visited, HeapPhase::Linking, (*visited / 2).min(objects), objects);
+        *visited += 1;
+        result
+    };
+
+    // Pass 2: node columns and out-degrees, counted into `out_start[v + 1]`.
+    let mut weak_edges_skipped = 0;
+    for segment in &segments {
+        scan_segment(buf, segment.clone(), id_size, &mut |at, record| {
+            let Some(node) = linker.node(at, &record)? else { return Ok(()) };
+            tick(&mut visited)?;
+            let v = node.index as usize;
+            types[v] = node.ty;
+            lens[v] = node.len;
+            offsets[v] = at as u64;
+            let mut degree = 0u64;
+            linker.edges(&record, &mut |_, weak| {
+                if weak {
+                    weak_edges_skipped += 1;
+                } else {
+                    degree += 1;
+                }
+            });
+            let count = &mut out_start[v + 1];
+            *count = u32::try_from(u64::from(*count) + degree).map_err(|_| HeapError::TooLarge)?;
+            Ok(())
+        })?;
+    }
+    out_start[nodes + 1] = u32::try_from(root_targets.len()).map_err(|_| HeapError::TooLarge)?;
+    let edges = accumulate(&mut out_start, cancel)?;
+
+    // Pass 3: `out_start[v]` is node v's write cursor; once filled it holds v's end, so the
+    // column is shifted one slot right to become the starts again.
+    let mut out = store.scratch::<u32>(edges as usize, 0)?;
+    for segment in &segments {
+        scan_segment(buf, segment.clone(), id_size, &mut |at, record| {
+            let Some(node) = linker.node(at, &record)? else { return Ok(()) };
+            tick(&mut visited)?;
+            let v = node.index as usize;
+            let mut cursor = out_start[v] as usize;
+            linker.edges(&record, &mut |target, weak| {
+                if !weak && let Some(slot) = out.get_mut(cursor) {
+                    *slot = target;
+                    cursor += 1;
+                }
+            });
+            out_start[v] = cursor as u32;
+            Ok(())
+        })?;
+    }
+    let first_root_edge = out_start[nodes] as usize;
+    for (slot, &target) in out[first_root_edge..].iter_mut().zip(&root_targets) {
+        *slot = target;
+    }
+    out_start[nodes] = edges;
+    out_start.copy_within(0..nodes + 1, 1);
+    out_start[0] = 0;
+    progress.report(HeapPhase::Linking, objects, objects)?;
+
+    drop(linker);
+    let ClassTable { classes, class_slot, .. } = table;
+    Ok(HeapGraph {
+        ids,
+        types,
+        lens,
+        offsets,
+        out_start,
+        out,
+        root: n,
+        roots,
+        classes,
+        class_slot,
+        thread_objects,
+        weak_edges_skipped,
+        id_size,
+        timestamp_ms: header.timestamp_ms,
+    })
+}
+
+/// Turns per-node counts into running totals in place (`counts[k]` becomes the sum of
+/// `counts[..=k]`) and returns the last one; past `u32::MAX` the graph is `TooLarge`.
+pub(crate) fn accumulate(counts: &mut [u32], cancel: &CancellationToken) -> Result<u32, HeapError> {
+    let mut sum = 0u32;
+    for chunk in counts.chunks_mut(CHECK_EVERY as usize) {
+        if cancel.is_cancelled() {
+            return Err(HeapError::Cancelled);
+        }
+        for count in chunk {
+            sum = sum.checked_add(*count).ok_or(HeapError::TooLarge)?;
+            *count = sum;
+        }
+    }
+    Ok(sum)
+}
+
+#[cfg(test)]
+mod tests;
@@ -0,0 +1,188 @@
+//! Class dumps resolved into [`ClassInfo`]s, plus where the references sit in instance data.
+
+use std::collections::HashMap;
+use std::hash::BuildHasherDefault;
+
+use super::{ClassInfo, PRIM_SLOTS};
+use crate::column::NONE;
+use crate::error::HeapError;
+use crate::hprof::{BasicType, ClassDump};
+use crate::ids::ChunkHasher;
+
+/// Keyed by hprof id (an address).
+pub(super) type IdHash<V> = HashMap<u64, V, BuildHasherDefault<ChunkHasher>>;
+
+const REFERENCE: &str = "java.lang.ref.Reference";
+const JAVA_LANG_CLASS: &str = "java.lang.Class";
+/// Not computed yet, in [`ClassTable::fielded`].
+const UNSET: u32 = NONE - 1;
+
+pub(super) struct ClassTable {
+    pub(super) classes: Vec<ClassInfo>,
+    /// Class id → class index.
+    pub(super) by_id: IdHash<u32>,
+    pub(super) class_slot: u32,
+    /// Per class: `(offset, is the referent)` of each reference field in the class's own part of
+    /// the instance data.
+    pub(super) refs: Vec<Vec<(usize, bool)>>,
+    /// Per class: bytes of its own part of the instance data.
+    pub(super) own_bytes: Vec<usize>,
+    /// Per class: the nearest class up its chain (itself included) whose own part is not empty, or
+    /// `NONE`. Walking these instead of every super keeps a walk bounded by the data it reads.
+    pub(super) fielded: Vec<u32>,
+}
+
+/// The longest name a JVM gives a class or field (a constant pool UTF8 has a `u2` length).
+const MAX_NAME_BYTES: usize = 65_535;
+/// Name bytes a dump may decode beyond 4 per byte of the file.
+const NAME_BUDGET_FLOOR: usize = 64 << 20;
+
+/// Names decoded from the UTF8 records within a budget: every field of every class may name the
+/// same record, so a small crafted file could otherwise ask for terabytes and abort the IDE.
+struct Names<'m, 'b> {
+    records: &'m IdHash<&'b [u8]>,
+    left: usize,
+}
+
+impl Names<'_, '_> {
+    /// Empty for an id with no record.
+    fn text(&mut self, id: u64) -> Result<String, HeapError> {
+        let Some(bytes) = self.records.get(&id) else { return Ok(String::new()) };
+        if bytes.len() > MAX_NAME_BYTES || bytes.len() > self.left {
+            return Err(HeapError::NamesTooLarge);
+        }
+        self.left -= bytes.len();
+        Ok(String::from_utf8_lossy(bytes).into_owned())
+    }
+}
+
+/// `dumps` holds one dump per class id, and `by_id` its position; `names` are the UTF8 records
+/// and `class_names` the LOAD CLASS class id → name id pairs. `file_len` bounds the name bytes.
+pub(super) fn resolve(
+    dumps: Vec<ClassDump>,
+    by_id: IdHash<u32>,
+    names: &IdHash<&[u8]>,
+    class_names: &IdHash<u64>,
+    id_size: u8,
+    file_len: usize,
+) -> Result<ClassTable, HeapError> {
+    // Room for the synthetic java.lang.Class and the primitive slots in a `u32` type id.
+    if dumps.len() > (u32::MAX - PRIM_SLOTS - 1) as usize {
+        return Err(HeapError::TooLarge);
+    }
+    let mut names = Names { records: names, left: file_len.saturating_mul(4).saturating_add(NAME_BUDGET_FLOOR) };
+    let mut supers: Vec<Option<u32>> =
+        dumps.iter().map(|d| if d.super_id == 0 { None } else { by_id.get(&d.super_id).copied() }).collect();
+    cut_super_cycles(&mut supers);
+
+    let mut classes = Vec::with_capacity(dumps.len() + 1);
+    let mut refs = Vec::with_capacity(dumps.len() + 1);
+    let mut own_bytes = Vec::with_capacity(dumps.len() + 1);
+    for (dump, super_index) in dumps.into_iter().zip(supers) {
+        let name = match class_names.get(&dump.id) {
+            Some(&name_id) if names.records.contains_key(&name_id) => names.text(name_id)?.replace('/', "."),
+            _ => format!("{:#x}", dump.id),
+        };
+        let fields =
+            dump.fields.iter().map(|&(n, ty)| Ok((names.text(n)?, ty))).collect::<Result<Vec<_>, HeapError>>()?;
+        let weak_field = if name == REFERENCE {
+            fields.iter().position(|(n, ty)| n == "referent" && *ty == BasicType::Object)
+        } else {
+            None
+        };
+        let mut offset = 0;
+        let mut class_refs = Vec::new();
+        for (k, &(_, ty)) in fields.iter().enumerate() {
+            if ty == BasicType::Object {
+                class_refs.push((offset, weak_field == Some(k)));
+            }
+            offset += ty.hprof_size(id_size);
+        }
+        refs.push(class_refs);
+        own_bytes.push(offset);
+        classes.push(ClassInfo {
+            id: dump.id,
+            name,
+            super_index,
+            fields,
+            statics: dump
+                .statics
+                .iter()
+                .map(|&(n, object)| Ok((names.text(n)?, object)))
+                .collect::<Result<_, HeapError>>()?,
+            static_bytes: dump.static_bytes,
+            weak_field,
+        });
+    }
+
+    let class_slot = match classes.iter().position(|c| c.name == JAVA_LANG_CLASS) {
+        Some(slot) => slot as u32,
+        None => {
+            classes.push(ClassInfo {
+                id: 0,
+                name: JAVA_LANG_CLASS.to_owned(),
+                super_index: None,
+                fields: Vec::new(),
+                statics: Vec::new(),
+                static_bytes: 0,
+                weak_field: None,
+            });
+            refs.push(Vec::new());
+            own_bytes.push(0);
+            classes.len() as u32 - 1
+        }
+    };
+    let fielded = nearest_fielded(&classes, &own_bytes);
+    Ok(ClassTable { classes, by_id, class_slot, refs, own_bytes, fielded })
+}
+
+/// A dump may claim `A extends B extends A`; the edge that closes a cycle is dropped so every
+/// walk up the chain ends. Iterative, so a long chain cannot overflow the stack.
+fn cut_super_cycles(supers: &mut [Option<u32>]) {
+    const NEW: u8 = 0;
+    const ON_PATH: u8 = 1;
+    const DONE: u8 = 2;
+    let mut state = vec![NEW; supers.len()];
+    let mut path = Vec::new();
+    for start in 0..supers.len() {
+        let mut c = start;
+        while state[c] == NEW {
+            state[c] = ON_PATH;
+            path.push(c);
+            match supers[c] {
+                Some(s) if state[s as usize] == ON_PATH => {
+                    supers[c] = None;
+                    break;
+                }
+                Some(s) => c = s as usize,
+                None => break,
+            }
+        }
+        for p in path.drain(..) {
+            state[p] = DONE;
+        }
+    }
+}
+
+fn nearest_fielded(classes: &[ClassInfo], own_bytes: &[usize]) -> Vec<u32> {
+    let mut fielded: Vec<u32> =
+        own_bytes.iter().enumerate().map(|(c, &b)| if b > 0 { c as u32 } else { UNSET }).collect();
+    let mut path = Vec::new();
+    for start in 0..classes.len() {
+        let mut c = start;
+        let found = loop {
+            if fielded[c] != UNSET {
+                break fielded[c];
+            }
+            path.push(c);
+            match classes[c].super_index {
+                Some(s) => c = s as usize,
+                None => break NONE,
+            }
+        };
+        for p in path.drain(..) {
+            fielded[p] = found;
+        }
+    }
+    fielded
+}
@@ -0,0 +1,107 @@
+//! Turns one heap sub-record into its node (index, type, length) and its outgoing edges. Both
+//! linking passes go through here, so the degrees counted in the first match the edges written
+//! in the second.
+
+use super::classes::ClassTable;
+use crate::column::NONE;
+use crate::error::HeapError;
+use crate::hprof::{BasicType, HeapRecord};
+use crate::ids::IdMap;
+
+pub(super) struct Node {
+    pub(super) index: u32,
+    pub(super) ty: u32,
+    pub(super) len: u32,
+}
+
+pub(super) struct Linker<'a> {
+    ids: &'a IdMap,
+    table: &'a ClassTable,
+    /// Per class: the node of its class object, `NONE` for the synthetic `java.lang.Class`.
+    class_objects: Vec<u32>,
+    id_size: u8,
+}
+
+impl<'a> Linker<'a> {
+    pub(super) fn new(ids: &'a IdMap, table: &'a ClassTable, id_size: u8) -> Self {
+        let class_objects =
+            table.classes.iter().map(|c| if c.id == 0 { NONE } else { ids.rank(c.id).unwrap_or(NONE) }).collect();
+        Self { ids, table, class_objects, id_size }
+    }
+
+    /// `None` for root records. `at` is the record's offset, for the error of an instance or array
+    /// whose class is not in the dump.
+    pub(super) fn node(&self, at: usize, record: &HeapRecord<'_>) -> Result<Option<Node>, HeapError> {
+        let class_index = |class_id: u64| {
+            self.table.by_id.get(&class_id).copied().ok_or(HeapError::MissingClass { id: class_id, at: at as u64 })
+        };
+        let (id, ty, len) = match record {
+            HeapRecord::Root { .. } => return Ok(None),
+            HeapRecord::Class(c) => (c.id, self.table.class_slot, 0),
+            HeapRecord::Instance { id, class_id, .. } => (*id, class_index(*class_id)?, 0),
+            HeapRecord::ObjectArray { id, class_id, len, .. } => (*id, class_index(*class_id)?, *len),
+            HeapRecord::PrimitiveArray { id, ty, len, .. } => {
+                let prim_base = self.table.classes.len() as u32;
+                (*id, prim_base + (*ty as u32 - BasicType::Boolean as u32), *len)
+            }
+        };
+        Ok(self.ids.rank(id).map(|index| Node { index, ty, len }))
+    }
+
+    /// Calls `edge(target, is_referent)` for every reference to an object in the dump, in field
+    /// order; `is_referent` marks `Reference.referent`, which callers leave out of the graph.
+    pub(super) fn edges(&self, record: &HeapRecord<'_>, edge: &mut impl FnMut(u32, bool)) {
+        let mut to = |id: u64, weak: bool| {
+            if let Some(target) = self.target(id) {
+                edge(target, weak);
+            }
+        };
+        match record {
+            HeapRecord::Root { .. } | HeapRecord::PrimitiveArray { .. } => {}
+            HeapRecord::Class(c) => {
+                to(c.super_id, false);
+                to(c.loader_id, false);
+                for &(_, object) in &c.statics {
+                    to(object, false);
+                }
+            }
+            HeapRecord::ObjectArray { data, .. } => {
+                for item in data.chunks_exact(usize::from(self.id_size)) {
+                    to(read_id(item, 0, self.id_size).unwrap_or(0), false);
+                }
+            }
+            HeapRecord::Instance { class_id, data, .. } => {
+                let Some(&class) = self.table.by_id.get(class_id) else { return };
+                let mut base = 0;
+                let mut next = self.table.fielded[class as usize];
+                'chain: while next != NONE && base < data.len() {
+                    let c = next as usize;
+                    for &(offset, weak) in &self.table.refs[c] {
+                        let Some(id) = read_id(data, base + offset, self.id_size) else { break 'chain };
+                        to(id, weak);
+                    }
+                    base += self.table.own_bytes[c];
+                    next = self.table.classes[c].super_index.map_or(NONE, |s| self.table.fielded[s as usize]);
+                }
+                let object = self.class_objects[class as usize];
+                if object != NONE {
+                    edge(object, false);
+                }
+            }
+        }
+    }
+
+    /// Id 0 is null; ids not in the dump have no node.
+    fn target(&self, id: u64) -> Option<u32> {
+        if id == 0 { None } else { self.ids.rank(id) }
+    }
+}
+
+/// The big-endian id at `at`, or `None` when `data` ends first.
+fn read_id(data: &[u8], at: usize, id_size: u8) -> Option<u64> {
+    let bytes = data.get(at..at.checked_add(usize::from(id_size))?)?;
+    Some(match *bytes {
+        [a, b, c, d] => u64::from(u32::from_be_bytes([a, b, c, d])),
+        _ => u64::from_be_bytes(bytes.try_into().ok()?),
+    })
+}
@@ -0,0 +1,479 @@
+use std::cell::RefCell;
+
+use proptest::prelude::*;
+
+use super::*;
+use crate::hprof::{HeapRecord, record_at};
+use crate::testing::{HprofWriter, MiniHeap, sample};
+
+const H: u64 = 0x1000;
+const N1: u64 = 0x1100;
+const N2: u64 = 0x1200;
+const N3: u64 = 0x1300;
+const L: u64 = 0x1400;
+const R: u64 = 0x1500;
+const W: u64 = 0x1600;
+const N4: u64 = 0x1700;
+
+fn no_progress(_: HeapPhase, _: u64, _: u64) -> Result<(), HeapError> {
+    Ok(())
+}
+
+fn build(bytes: &[u8]) -> Result<HeapGraph, HeapError> {
+    build_graph(bytes, &Store::Ram, &no_progress, &CancellationToken::new())
+}
+
+fn index(g: &HeapGraph, id: u64) -> u32 {
+    g.ids.rank(id).unwrap_or_else(|| panic!("{id:#x} is not in the graph"))
+}
+
+fn class(g: &HeapGraph, name: &str) -> (u32, u64) {
+    let slot = g.classes.iter().position(|c| c.name == name).unwrap_or_else(|| panic!("no class {name}"));
+    (slot as u32, g.classes[slot].id)
+}
+
+/// Targets of `id`'s edges, as ids, in edge order.
+fn out_of(g: &HeapGraph, id: u64) -> Vec<u64> {
+    g.out_edges(index(g, id)).iter().map(|&t| g.ids.id_at(t)).collect()
+}
+
+fn root_targets(g: &HeapGraph) -> Vec<u64> {
+    g.out_edges(g.root).iter().map(|&t| g.ids.id_at(t)).collect()
+}
+
+fn assert_csr(g: &HeapGraph) {
+    let n = g.ids.len();
+    assert_eq!(g.root, n);
+    assert_eq!(g.out_start.len(), n as usize + 2);
+    assert_eq!(g.out_start[0], 0);
+    assert!(g.out_start.windows(2).all(|w| w[0] <= w[1]), "out_start is not monotone");
+    assert_eq!(g.out_start[n as usize + 1] as usize, g.out.len());
+    assert!(g.out.iter().all(|&t| t < n), "edge to a node outside the objects");
+    assert_eq!((g.types.len(), g.lens.len(), g.offsets.len()), (n as usize, n as usize, n as usize));
+    let slots = g.classes.len() as u32 + PRIM_SLOTS;
+    assert!(g.types.iter().all(|&t| t < slots), "type outside the slots");
+    for c in &g.classes {
+        let mut steps = 0;
+        let mut cursor = c.super_index;
+        while let Some(s) = cursor {
+            steps += 1;
+            assert!(steps <= g.classes.len(), "super chain of {} does not end", c.name);
+            cursor = g.classes[s as usize].super_index;
+        }
+    }
+}
+
+#[test]
+fn edges_follow_fields_and_skip_weak_referents() {
+    let g = build(&sample().bytes()).unwrap();
+    assert_csr(&g);
+    let (_, holder) = class(&g, "Holder");
+    let (_, node) = class(&g, "Node");
+    let (_, weak) = class(&g, "java.lang.ref.WeakReference");
+    assert_eq!(out_of(&g, H), [N1, N2, holder]);
+    assert_eq!(out_of(&g, N1), [N3, node]);
+    assert_eq!(out_of(&g, N3), [L, node]);
+    assert_eq!(out_of(&g, W), [weak]);
+    assert_eq!(out_of(&g, N4), [node]);
+    assert_eq!(out_of(&g, L), Vec::<u64>::new());
+    assert_eq!(g.weak_edges_skipped, 1);
+    let (reference, _) = class(&g, "java.lang.ref.Reference");
+    assert_eq!(g.classes[reference as usize].weak_field, Some(0));
+    assert!(g.classes.iter().filter(|c| c.name != "java.lang.ref.Reference").all(|c| c.weak_field.is_none()));
+}
+
+#[test]
+fn object_arrays_skip_null_and_unknown() {
+    let mut heap = sample();
+    heap.array_of(0x1800, "[Ljava.lang.Object;", &[N2, 0, 0x9990, N1, 0xDEAD_BEEF]);
+    let g = build(&heap.bytes()).unwrap();
+    assert_csr(&g);
+    assert_eq!(out_of(&g, R), [N1]);
+    assert_eq!(out_of(&g, 0x1800), [N2, N1]);
+}
+
+#[test]
+fn super_root_points_at_each_root_once() {
+    let mut heap = sample();
+    heap.root(HeapRootKind::JniGlobal, H).root(HeapRootKind::MonitorUsed, H).root(HeapRootKind::JniLocal, 0x9990);
+    let g = build(&heap.bytes()).unwrap();
+    assert_csr(&g);
+    let targets = root_targets(&g);
+    assert_eq!(targets.iter().filter(|&&t| t == H).count(), 1);
+    assert!(targets.contains(&N3) && targets.contains(&W));
+    assert!(!targets.contains(&N1) && !targets.contains(&R) && !targets.contains(&N4));
+    let classes = g.classes.iter().filter(|c| c.id != 0).count();
+    assert_eq!(targets.len(), 3 + classes, "H, N3, W and every sticky class");
+    let mut sorted = targets.clone();
+    sorted.sort_unstable();
+    sorted.dedup();
+    assert_eq!(sorted.len(), targets.len());
+
+    let h = index(&g, H);
+    let of_h: Vec<_> = g.roots.iter().filter(|r| r.0 == h).map(|r| (r.1, r.2)).collect();
+    assert_eq!(of_h, [(HeapRootKind::JniGlobal, None), (HeapRootKind::MonitorUsed, None)]);
+    assert!(g.roots.contains(&(index(&g, N3), HeapRootKind::JavaFrame, Some(1))));
+    assert!(g.roots.iter().all(|r| r.0 < g.root));
+}
+
+#[test]
+fn class_dump_edges_cover_statics() {
+    let mut w = HprofWriter::new();
+    w.string(1, "java/lang/Object")
+        .load_class(0x100, 1)
+        .class(0x100, 0, &[], &[])
+        .string(2, "Config")
+        .string(3, "INSTANCE")
+        .string(4, "EMPTY")
+        .string(5, "LOST")
+        .load_class(0x200, 2)
+        .class_with_loader(0x200, 0x100, 0x1200, &[], &[(3, N1), (4, 0), (5, 0x9990)])
+        .instance(N1, 0x100, &[])
+        .instance(0x1200, 0x100, &[]);
+    let g = build(&w.finish()).unwrap();
+    assert_csr(&g);
+    assert_eq!(out_of(&g, 0x200), [0x100, 0x1200, N1]);
+    assert_eq!(out_of(&g, 0x100), Vec::<u64>::new());
+    let (config, _) = class(&g, "Config");
+    let info = &g.classes[config as usize];
+    let statics: Vec<_> = info.statics.iter().map(|(n, id)| (n.as_str(), *id)).collect();
+    assert_eq!(statics, [("INSTANCE", N1), ("EMPTY", 0), ("LOST", 0x9990)]);
+    assert_eq!(info.super_index, Some(class(&g, "java.lang.Object").0));
+}
+
+#[test]
+fn missing_class_is_an_error() {
+    let mut w = HprofWriter::new();
+    w.string(1, "java/lang/Object").load_class(0x100, 1).class(0x100, 0, &[], &[]).instance(N1, 0x999, &[]);
+    let bytes = w.finish();
+    let at = instance_offset(&bytes, N1);
+    let err = build(&bytes).err().unwrap();
+    assert!(matches!(err, HeapError::MissingClass { id: 0x999, at: a } if a == at as u64), "{err:?}");
+
+    let mut w = HprofWriter::new();
+    w.obj_array(0x10, 0x999, &[]);
+    assert!(matches!(build(&w.finish()), Err(HeapError::MissingClass { id: 0x999, .. })));
+}
+
+fn instance_offset(bytes: &[u8], id: u64) -> usize {
+    let header = crate::hprof::parse_header(bytes).unwrap();
+    let mut found = None;
+    crate::hprof::scan(bytes, &header, &mut |r| {
+        if let crate::hprof::Record::Heap { at, record: HeapRecord::Instance { id: i, .. } } = r
+            && i == id
+        {
+            found = Some(at);
+        }
+        Ok(())
+    })
+    .unwrap();
+    found.unwrap()
+}
+
+#[test]
+fn cancel_during_linking_returns_cancelled() {
+    let bytes = sample().bytes();
+    let cancelled = CancellationToken::new();
+    cancelled.cancel();
+    let result = build_graph(&bytes, &Store::Ram, &no_progress, &cancelled);
+    assert!(matches!(result, Err(HeapError::Cancelled)));
+
+    let cancel = CancellationToken::new();
+    let phases = RefCell::new(Vec::new());
+    let result = build_graph(
+        &bytes,
+        &Store::Ram,
+        &|phase, done, total| {
+            phases.borrow_mut().push((phase, done, total));
+            if phase == HeapPhase::Linking {
+                cancel.cancel();
+            }
+            Ok(())
+        },
+        &cancel,
+    );
+    assert!(matches!(result, Err(HeapError::Cancelled)));
+    let phases = phases.into_inner();
+    assert_eq!(phases.first(), Some(&(HeapPhase::Reading, 0, bytes.len() as u64)));
+    assert_eq!(phases.last(), Some(&(HeapPhase::Linking, 0, 15)), "stops at the first Linking report");
+}
+
+#[test]
+fn progress_error_stops_the_build() {
+    let bytes = sample().bytes();
+    let fail = |phase, _, total| {
+        if phase == HeapPhase::Linking {
+            Err(HeapError::NotEnoughDiskSpace { needed: total * 90, available: 0 })
+        } else {
+            Ok(())
+        }
+    };
+    let result = build_graph(&bytes, &Store::Ram, &fail, &CancellationToken::new());
+    assert!(matches!(result, Err(HeapError::NotEnoughDiskSpace { needed: 1350, available: 0 })));
+}
+
+#[test]
+fn progress_reads_bytes_then_links_objects() {
+    let bytes = sample().bytes();
+    let phases = RefCell::new(Vec::new());
+    let g = build_graph(
+        &bytes,
+        &Store::Ram,
+        &|phase, done, total| {
+            phases.borrow_mut().push((phase, done, total));
+            Ok(())
+        },
+        &CancellationToken::new(),
+    )
+    .unwrap();
+    let phases = phases.into_inner();
+    let len = bytes.len() as u64;
+    let objects = u64::from(g.root);
+    assert_eq!(objects, 15, "7 classes and 8 objects");
+    assert_eq!(phases.first(), Some(&(HeapPhase::Reading, 0, len)));
+    assert!(phases.contains(&(HeapPhase::Reading, len, len)));
+    let linking: Vec<_> = phases.iter().filter(|p| p.0 == HeapPhase::Linking).collect();
+    assert_eq!(linking.first(), Some(&&(HeapPhase::Linking, 0, objects)));
+    assert_eq!(linking.iter().filter(|p| ***p == (HeapPhase::Linking, 0, objects)).count(), 1, "{linking:?}");
+    assert_eq!(linking.last(), Some(&&(HeapPhase::Linking, objects, objects)));
+    assert!(phases.iter().all(|&(_, done, total)| done <= total));
+}
+
+#[test]
+fn types_lengths_and_offsets() {
+    let bytes = sample().bytes();
+    let g = build(&bytes).unwrap();
+    let (holder, holder_id) = class(&g, "Holder");
+    let (node, _) = class(&g, "Node");
+    let (array, _) = class(&g, "[Ljava.lang.Object;");
+    let (java_lang_class, _) = class(&g, "java.lang.Class");
+    assert_eq!(g.class_slot, java_lang_class);
+    let ty = |id| g.types[index(&g, id) as usize];
+    let len = |id| g.lens[index(&g, id) as usize];
+    assert_eq!([ty(H), ty(N1), ty(R), ty(holder_id)], [holder, node, array, java_lang_class]);
+    assert_eq!(ty(L), g.prim_slot(BasicType::Long));
+    assert_eq!(g.prim_type(ty(L)), Some(BasicType::Long));
+    assert_eq!(g.prim_type(holder), None);
+    assert_eq!(g.prim_type(g.classes.len() as u32 + PRIM_SLOTS), None);
+    assert_eq!(g.prim_slot(BasicType::Boolean), g.classes.len() as u32);
+    assert_eq!(g.prim_slot(BasicType::Long), g.classes.len() as u32 + PRIM_SLOTS - 1);
+    assert_eq!([len(H), len(L), len(R), len(holder_id)], [0, 2, 3, 0]);
+    for i in 0..g.root {
+        let id = match record_at(&bytes, g.offsets[i as usize] as usize, g.id_size).unwrap() {
+            HeapRecord::Class(c) => c.id,
+            HeapRecord::Instance { id, .. } | HeapRecord::ObjectArray { id, .. } => id,
+            HeapRecord::PrimitiveArray { id, .. } => id,
+            other => panic!("offset of object {i} points at {other:?}"),
+        };
+        assert_eq!(id, g.ids.id_at(i));
+    }
+    assert_eq!((g.id_size, g.timestamp_ms), (8, 1_700_000_000_000));
+    let holder_info = &g.classes[holder as usize];
+    let fields: Vec<_> = holder_info.fields.iter().map(|(n, t)| (n.as_str(), *t)).collect();
+    assert_eq!(fields, [("a", BasicType::Object), ("b", BasicType::Object)]);
+    assert_eq!(holder_info.super_index, Some(class(&g, "java.lang.Object").0));
+    assert_eq!(g.classes[class(&g, "java.lang.Object").0 as usize].super_index, None);
+}
+
+#[test]
+fn dump_without_java_lang_class_gets_a_synthetic_slot() {
+    let mut w = HprofWriter::new();
+    w.string(1, "java/lang/Object").load_class(0x100, 1).class(0x100, 0, &[], &[]).instance(N1, 0x100, &[]);
+    let g = build(&w.finish()).unwrap();
+    assert_csr(&g);
+    assert_eq!(g.classes.len(), 2);
+    let synthetic = &g.classes[g.class_slot as usize];
+    assert_eq!((synthetic.id, synthetic.name.as_str(), synthetic.super_index), (0, "java.lang.Class", None));
+    assert_eq!(g.types[index(&g, 0x100) as usize], g.class_slot);
+    assert_eq!(g.types[index(&g, N1) as usize], 0);
+}
+
+#[test]
+fn classes_without_a_name_record_get_their_address() {
+    let mut w = HprofWriter::new();
+    w.class(0x100, 0, &[(9, BasicType::Object)], &[])
+        .instance(N1, 0x100, &N2.to_be_bytes())
+        .instance(N2, 0x100, &[0; 8]);
+    let g = build(&w.finish()).unwrap();
+    assert_eq!(g.classes[0].name, "0x100");
+    assert_eq!(g.classes[0].fields, [(String::new(), BasicType::Object)]);
+    assert_eq!(out_of(&g, N1), [N2, 0x100]);
+}
+
+#[test]
+fn thread_objects_map_serials_to_threads() {
+    let mut heap = sample();
+    heap.class("java.lang.Thread", Some("java.lang.Object"), &[])
+        .object(0x2000, "java.lang.Thread", &[])
+        .object(0x2100, "java.lang.Thread", &[])
+        .root_in_thread(HeapRootKind::ThreadObject, 0x2000, 1)
+        .root_in_thread(HeapRootKind::ThreadObject, 0x2100, 7)
+        .root_in_thread(HeapRootKind::ThreadObject, 0x9990, 9);
+    let g = build(&heap.bytes()).unwrap();
+    let expected: HashMap<u32, u32> = [(1, index(&g, 0x2000)), (7, index(&g, 0x2100))].into();
+    assert_eq!(g.thread_objects, expected);
+    assert!(g.roots.contains(&(index(&g, 0x2100), HeapRootKind::ThreadObject, Some(7))));
+}
+
+#[test]
+fn inherited_fields_follow_the_class_first_layout() {
+    let mut heap = MiniHeap::new();
+    heap.class("A", Some("java.lang.Object"), &[("a", BasicType::Object), ("n", BasicType::Int)])
+        .class("Empty", Some("A"), &[])
+        .class("B", Some("Empty"), &[("flag", BasicType::Byte), ("b", BasicType::Object)])
+        .object(0x10, "B", &[("a", 0x20), ("b", 0x30), ("flag", 1)])
+        .object(0x20, "A", &[])
+        .object(0x30, "A", &[]);
+    let g = build(&heap.bytes()).unwrap();
+    assert_csr(&g);
+    assert_eq!(out_of(&g, 0x10), [0x30, 0x20, class(&g, "B").1]);
+}
+
+#[test]
+fn super_class_cycles_are_cut() {
+    let mut w = HprofWriter::new();
+    w.string(1, "A")
+        .string(2, "B")
+        .string(3, "S")
+        .string(4, "f")
+        .load_class(0x100, 1)
+        .load_class(0x200, 2)
+        .load_class(0x300, 3)
+        .class(0x100, 0x200, &[(4, BasicType::Object)], &[])
+        .class(0x200, 0x100, &[(4, BasicType::Object)], &[])
+        .class(0x300, 0x300, &[], &[])
+        .instance(0x10, 0x100, &[[0; 7].as_slice(), &[0x20], &[0; 7], &[0x30]].concat())
+        .instance(0x20, 0x300, &[])
+        .instance(0x30, 0x300, &[0; 64]);
+    let g = build(&w.finish()).unwrap();
+    assert_csr(&g);
+    assert_eq!(out_of(&g, 0x10), [0x20, 0x30, 0x100]);
+    assert_eq!(out_of(&g, 0x30), [0x300]);
+}
+
+#[test]
+fn short_instance_data_stops_at_the_last_whole_field() {
+    let mut w = HprofWriter::new();
+    w.string(1, "P")
+        .string(2, "x")
+        .load_class(0x100, 1)
+        .class(0x100, 0, &[(2, BasicType::Object), (2, BasicType::Object)], &[])
+        .instance(0x10, 0x100, &[[0; 7].as_slice(), &[0x20], &[0; 5]].concat())
+        .instance(0x20, 0x100, &[]);
+    let g = build(&w.finish()).unwrap();
+    assert_csr(&g);
+    assert_eq!(out_of(&g, 0x10), [0x20, 0x100]);
+}
+
+#[test]
+fn id_size_4_dump_links() {
+    let mut w = HprofWriter::with_id_size(4);
+    w.string(1, "Node")
+        .string(2, "next")
+        .load_class(0x100, 1)
+        .class(0x100, 0, &[(2, BasicType::Object)], &[(2, 0x20)])
+        .instance(0x10, 0x100, &0x20u32.to_be_bytes())
+        .instance(0x20, 0x100, &[0; 4])
+        .obj_array(0x30, 0x100, &[0x10, 0, 0x20])
+        .root(HeapRootKind::JniGlobal, 0x30);
+    let g = build(&w.finish()).unwrap();
+    assert_csr(&g);
+    assert_eq!(g.id_size, 4);
+    assert_eq!(out_of(&g, 0x10), [0x20, 0x100]);
+    assert_eq!(out_of(&g, 0x30), [0x10, 0x20]);
+    assert_eq!(out_of(&g, 0x100), [0x20]);
+    assert_eq!(root_targets(&g), [0x30]);
+}
+
+#[test]
+fn dir_store_keeps_types_and_offsets_only() {
+    let dir = tempfile::tempdir().unwrap();
+    // The id map's own files (`ids.*`) depend on its layout; MiniHeap's class ids are unaligned.
+    let names = || {
+        let mut names: Vec<_> = std::fs::read_dir(dir.path())
+            .unwrap()
+            .map(|e| e.unwrap().file_name().into_string().unwrap())
+            .filter(|name| !name.starts_with("ids."))
+            .collect();
+        names.sort();
+        names
+    };
+    let g =
+        build_graph(&sample().bytes(), &Store::Dir(dir.path().to_path_buf()), &no_progress, &CancellationToken::new())
+            .unwrap();
+    assert_eq!(out_of(&g, H).len(), 3);
+    let kept = ["offsets.u64", "types.u32"];
+    assert!(dir.path().join("ids.sorted.u64").exists());
+    #[cfg(unix)]
+    assert_eq!(names(), kept);
+    drop(g);
+    assert_eq!(names(), kept);
+}
+
+#[test]
+fn empty_dump_has_only_the_super_root() {
+    let g = build(&HprofWriter::new().finish()).unwrap();
+    assert_csr(&g);
+    assert_eq!((g.root, g.out.len(), g.classes.len()), (0, 0, 1));
+    assert_eq!(g.class_slot, 0);
+}
+
+#[test]
+fn edge_count_past_u32_is_too_large() {
+    let cancel = CancellationToken::new();
+    let mut counts = [0, 2, 0, 3, 1];
+    assert_eq!(accumulate(&mut counts, &cancel).unwrap(), 6);
+    assert_eq!(counts, [0, 2, 2, 5, 6]);
+    let mut full = [0, u32::MAX - 1, 1];
+    assert_eq!(accumulate(&mut full, &cancel).unwrap(), u32::MAX);
+    let mut over = [0, u32::MAX, 1];
+    assert!(matches!(accumulate(&mut over, &cancel), Err(HeapError::TooLarge)));
+    cancel.cancel();
+    assert!(matches!(accumulate(&mut [0, 1], &cancel), Err(HeapError::Cancelled)));
+}
+
+/// A class with `fields` instance fields, every one named by the same `name_len`-byte UTF8 record.
+fn shared_long_field_names(name_len: usize, fields: usize) -> Vec<u8> {
+    let mut w = HprofWriter::new();
+    let long = "f".repeat(name_len);
+    w.string(1, "java/lang/Object").load_class(0x100, 1).class(0x100, 0, &[], &[]);
+    w.string(2, "Wide").string(3, &long).load_class(0x200, 2);
+    w.class(0x200, 0x100, &vec![(3, BasicType::Int); fields], &[]);
+    w.finish()
+}
+
+#[test]
+fn names_that_would_decode_far_larger_than_the_file_are_refused() {
+    // 128 MiB of names from a file of about 70 KB.
+    let bytes = shared_long_field_names(32 << 10, 4000);
+    assert!(bytes.len() < 100 << 10, "{}", bytes.len());
+    let err = build(&bytes).err().expect("refused");
+    assert!(matches!(err, HeapError::NamesTooLarge), "{err:?}");
+}
+
+#[test]
+fn a_name_longer_than_the_jvm_allows_is_refused() {
+    let mut w = HprofWriter::new();
+    w.string(1, &"C".repeat(70_000)).load_class(0x100, 1).class(0x100, 0, &[], &[]);
+    assert!(matches!(build(&w.finish()), Err(HeapError::NamesTooLarge)));
+}
+
+#[test]
+fn real_sized_names_shared_by_many_fields_are_fine() {
+    let g = build(&shared_long_field_names(40, 4000)).unwrap();
+    assert_eq!(g.classes[class(&g, "Wide").0 as usize].fields.len(), 4000);
+}
+
+proptest! {
+    #[test]
+    fn build_graph_never_panics_on_mutated_dumps(flips in proptest::collection::vec((0usize..4096, any::<u8>()), 1..12)) {
+        let mut bytes = sample().bytes();
+        for (pos, value) in flips {
+            let at = pos % bytes.len();
+            bytes[at] = value;
+        }
+        if let Ok(g) = build(&bytes) {
+            assert_csr(&g);
+        }
+    }
+}
@@ -0,0 +1,945 @@
+//! Reader for HotSpot `.hprof` heap dumps (format 1.0.1 / 1.0.2, 4- or 8-byte ids).
+
+use std::ops::Range;
+
+use serde::{Deserialize, Serialize};
+
+use crate::error::HeapError;
+use crate::reader::Reader;
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+#[repr(u8)]
+pub enum BasicType {
+    Object = 2,
+    Boolean = 4,
+    Char = 5,
+    Float = 6,
+    Double = 7,
+    Byte = 8,
+    Short = 9,
+    Int = 10,
+    Long = 11,
+}
+
+impl BasicType {
+    pub fn from_u8(v: u8) -> Option<Self> {
+        Some(match v {
+            2 => Self::Object,
+            4 => Self::Boolean,
+            5 => Self::Char,
+            6 => Self::Float,
+            7 => Self::Double,
+            8 => Self::Byte,
+            9 => Self::Short,
+            10 => Self::Int,
+            11 => Self::Long,
+            _ => return None,
+        })
+    }
+
+    /// Bytes one value takes in the dump; `id_size` is the dump's identifier size.
+    pub fn hprof_size(self, id_size: u8) -> usize {
+        match self {
+            Self::Object => usize::from(id_size),
+            Self::Boolean | Self::Byte => 1,
+            Self::Char | Self::Short => 2,
+            Self::Float | Self::Int => 4,
+            Self::Double | Self::Long => 8,
+        }
+    }
+
+    /// Bytes one value takes in the JVM heap; `ref_size` is 4 with compressed oops, else 8.
+    pub fn heap_size(self, ref_size: u8) -> u64 {
+        match self {
+            Self::Object => u64::from(ref_size),
+            other => other.hprof_size(ref_size) as u64,
+        }
+    }
+
+    /// JVM class name of an array of this type (`[I`); `Object` arrays have no single name, so
+    /// they report the most general one.
+    pub fn array_name(self) -> &'static str {
+        match self {
+            Self::Object => "[Ljava.lang.Object;",
+            Self::Boolean => "[Z",
+            Self::Char => "[C",
+            Self::Float => "[F",
+            Self::Double => "[D",
+            Self::Byte => "[B",
+            Self::Short => "[S",
+            Self::Int => "[I",
+            Self::Long => "[J",
+        }
+    }
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize, specta::Type)]
+#[serde(rename_all = "camelCase")]
+pub enum HeapRootKind {
+    Unknown,
+    JniGlobal,
+    JniLocal,
+    JavaFrame,
+    NativeStack,
+    StickyClass,
+    ThreadBlock,
+    MonitorUsed,
+    ThreadObject,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct HprofHeader {
+    pub version: String,
+    pub id_size: u8,
+    pub timestamp_ms: u64,
+    /// Offset of the first record.
+    pub body_start: usize,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct ClassDump {
+    pub id: u64,
+    pub super_id: u64,
+    pub loader_id: u64,
+    /// `(name id, object id)` of reference statics; a null static has object id 0.
+    pub statics: Vec<(u64, u64)>,
+    /// Heap bytes taken by the primitive statics.
+    pub static_bytes: u64,
+    /// Instance fields in dump order, `(name id, type)`.
+    pub fields: Vec<(u64, BasicType)>,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub enum HeapRecord<'a> {
+    /// `thread_serial` is present for the kinds that carry one.
+    Root {
+        kind: HeapRootKind,
+        id: u64,
+        thread_serial: Option<u32>,
+    },
+    Class(ClassDump),
+    /// `data` is the field values, class first, then each superclass.
+    Instance {
+        id: u64,
+        class_id: u64,
+        data: &'a [u8],
+    },
+    /// `data` is `len` ids.
+    ObjectArray {
+        id: u64,
+        class_id: u64,
+        len: u32,
+        data: &'a [u8],
+    },
+    /// `data` is `len` big-endian elements.
+    PrimitiveArray {
+        id: u64,
+        ty: BasicType,
+        len: u32,
+        data: &'a [u8],
+    },
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub enum Record<'a> {
+    Utf8 {
+        id: u64,
+        text: &'a [u8],
+    },
+    LoadClass {
+        class_id: u64,
+        name_id: u64,
+    },
+    /// Byte range of a heap-dump segment's sub-records; its contents follow as `Heap` records.
+    Segment(Range<usize>),
+    /// `at` is the offset of the sub-record's tag, for [`record_at`].
+    Heap {
+        at: usize,
+        record: HeapRecord<'a>,
+    },
+}
+
+const MAGIC_1_0_2: &[u8] = b"JAVA PROFILE 1.0.2\0";
+const MAGIC_1_0_1: &[u8] = b"JAVA PROFILE 1.0.1\0";
+
+const TAG_UTF8: u8 = 0x01;
+const TAG_LOAD_CLASS: u8 = 0x02;
+const TAG_HEAP_DUMP: u8 = 0x0C;
+const TAG_HEAP_DUMP_SEGMENT: u8 = 0x1C;
+const TAG_HEAP_DUMP_END: u8 = 0x2C;
+/// Tag (u1), timestamp (u4) and body length (u4).
+const RECORD_HEADER_LEN: usize = 9;
+
+fn check_id_size(id_size: u8) -> Result<(), HeapError> {
+    match id_size {
+        4 | 8 => Ok(()),
+        other => Err(HeapError::UnsupportedIdSize(u32::from(other))),
+    }
+}
+
+fn truncated(buf: &[u8]) -> HeapError {
+    HeapError::Truncated { at: buf.len() as u64 }
+}
+
+pub fn parse_header(buf: &[u8]) -> Result<HprofHeader, HeapError> {
+    let magic =
+        [MAGIC_1_0_2, MAGIC_1_0_1].into_iter().find(|magic| buf.starts_with(magic)).ok_or(HeapError::NotHprof)?;
+    let mut r = Reader::new(buf, magic.len(), 8);
+    let id_size = r.u4()?;
+    let id_size = u8::try_from(id_size).map_err(|_| HeapError::UnsupportedIdSize(id_size))?;
+    check_id_size(id_size)?;
+    let timestamp_ms = r.u8()?;
+    let version = String::from_utf8_lossy(&magic[..magic.len() - 1]).into_owned();
+    Ok(HprofHeader { version, id_size, timestamp_ms, body_start: r.pos() })
+}
+
+/// Walks every top-level record and, for heap-dump segments, the sub-records inside them.
+/// A record whose length runs past the end of `buf` is `Truncated`, and so are segments that no
+/// HEAP DUMP END follows: HotSpot writes whole segments, so a dump it did not finish ends after one.
+pub fn scan<'a>(
+    buf: &'a [u8],
+    header: &HprofHeader,
+    visit: &mut dyn FnMut(Record<'a>) -> Result<(), HeapError>,
+) -> Result<(), HeapError> {
+    let id_size = header.id_size;
+    check_id_size(id_size)?;
+    if header.body_start > buf.len() {
+        return Err(truncated(buf));
+    }
+    let mut pos = header.body_start;
+    let mut unended_segments = false;
+    while pos < buf.len() {
+        let mut head = Reader::new(buf, pos, id_size);
+        let tag = head.u1()?;
+        head.skip(4)?; // timestamp
+        let len = head.u4()? as usize;
+        let start = pos + RECORD_HEADER_LEN;
+        let end = start.checked_add(len).filter(|&end| end <= buf.len()).ok_or_else(|| truncated(buf))?;
+        let body = &buf[..end];
+        let mut r = Reader::new(body, start, id_size);
+        match tag {
+            TAG_UTF8 => {
+                let id = r.id()?;
+                let text = r.bytes(end - r.pos())?;
+                visit(Record::Utf8 { id, text })?;
+            }
+            TAG_LOAD_CLASS => {
+                r.skip(4)?; // class serial
+                let class_id = r.id()?;
+                r.skip(4)?; // stack trace serial
+                let name_id = r.id()?;
+                visit(Record::LoadClass { class_id, name_id })?;
+            }
+            TAG_HEAP_DUMP | TAG_HEAP_DUMP_SEGMENT => {
+                // A single HEAP DUMP (JDK 8, under 2 GB) has no end record.
+                unended_segments |= tag == TAG_HEAP_DUMP_SEGMENT;
+                visit(Record::Segment(start..end))?;
+                walk_segment(buf, start..end, id_size, &mut |at, record| visit(Record::Heap { at, record }))?;
+            }
+            TAG_HEAP_DUMP_END => unended_segments = false,
+            _ => {}
+        }
+        pos = end;
+    }
+    if unended_segments {
+        return Err(truncated(buf));
+    }
+    Ok(())
+}
+
+/// Visits the sub-records of one segment (the range from [`Record::Segment`]) with their offsets.
+/// The last sub-record must end exactly at the segment's end.
+pub fn scan_segment<'a>(
+    buf: &'a [u8],
+    segment: Range<usize>,
+    id_size: u8,
+    visit: &mut dyn FnMut(usize, HeapRecord<'a>) -> Result<(), HeapError>,
+) -> Result<(), HeapError> {
+    check_id_size(id_size)?;
+    walk_segment(buf, segment, id_size, visit)
+}
+
+fn walk_segment<'a>(
+    buf: &'a [u8],
+    segment: Range<usize>,
+    id_size: u8,
+    visit: &mut dyn FnMut(usize, HeapRecord<'a>) -> Result<(), HeapError>,
+) -> Result<(), HeapError> {
+    if segment.start > segment.end || segment.end > buf.len() {
+        return Err(truncated(buf));
+    }
+    let mut r = Reader::new(&buf[..segment.end], segment.start, id_size);
+    while r.pos() < segment.end {
+        let at = r.pos();
+        let record = read_sub_record(&mut r)?;
+        visit(at, record)?;
+    }
+    Ok(())
+}
+
+/// Reads the sub-record whose tag is at `at`, as given by [`Record::Heap`].
+pub fn record_at(buf: &[u8], at: usize, id_size: u8) -> Result<HeapRecord<'_>, HeapError> {
+    check_id_size(id_size)?;
+    read_sub_record(&mut Reader::new(buf, at, id_size))
+}
+
+fn read_sub_record<'a>(r: &mut Reader<'a>) -> Result<HeapRecord<'a>, HeapError> {
+    let at = r.pos();
+    let tag = r.u1()?;
+    // (kind, carries a thread serial, bytes after it)
+    let root = match tag {
+        0xFF => Some((HeapRootKind::Unknown, false, 0)),
+        0x01 => Some((HeapRootKind::JniGlobal, false, usize::from(r.id_size()))),
+        0x02 => Some((HeapRootKind::JniLocal, true, 4)),
+        0x03 => Some((HeapRootKind::JavaFrame, true, 4)),
+        0x04 => Some((HeapRootKind::NativeStack, true, 0)),
+        0x05 => Some((HeapRootKind::StickyClass, false, 0)),
+        0x06 => Some((HeapRootKind::ThreadBlock, true, 0)),
+        0x07 => Some((HeapRootKind::MonitorUsed, false, 0)),
+        0x08 => Some((HeapRootKind::ThreadObject, true, 4)),
+        _ => None,
+    };
+    if let Some((kind, has_serial, trailing)) = root {
+        let id = r.id()?;
+        let thread_serial = if has_serial { Some(r.u4()?) } else { None };
+        r.skip(trailing)?;
+        return Ok(HeapRecord::Root { kind, id, thread_serial });
+    }
+    match tag {
+        0x20 => read_class_dump(r).map(HeapRecord::Class),
+        0x21 => {
+            let id = r.id()?;
+            r.skip(4)?; // stack trace serial
+            let class_id = r.id()?;
+            let len = r.u4()? as usize;
+            Ok(HeapRecord::Instance { id, class_id, data: r.bytes(len)? })
+        }
+        0x22 => {
+            let id = r.id()?;
+            r.skip(4)?;
+            let len = r.u4()?;
+            let class_id = r.id()?;
+            let size = (len as usize).checked_mul(usize::from(r.id_size())).ok_or_else(|| truncated(r.buf()))?;
+            Ok(HeapRecord::ObjectArray { id, class_id, len, data: r.bytes(size)? })
+        }
+        0x23 => {
+            let id = r.id()?;
+            r.skip(4)?;
+            let len = r.u4()?;
+            let ty = read_type(r)?;
+            if ty == BasicType::Object {
+                return Err(HeapError::UnknownType { ty: ty as u8, at: (r.pos() - 1) as u64 });
+            }
+            let size = (len as usize).checked_mul(ty.hprof_size(r.id_size())).ok_or_else(|| truncated(r.buf()))?;
+            Ok(HeapRecord::PrimitiveArray { id, ty, len, data: r.bytes(size)? })
+        }
+        tag => Err(HeapError::UnknownSubRecord { tag, at: at as u64 }),
+    }
+}
+
+fn read_type(r: &mut Reader<'_>) -> Result<BasicType, HeapError> {
+    let at = r.pos();
+    let ty = r.u1()?;
+    BasicType::from_u8(ty).ok_or(HeapError::UnknownType { ty, at: at as u64 })
+}
+
+fn read_class_dump(r: &mut Reader<'_>) -> Result<ClassDump, HeapError> {
+    let id_size = r.id_size();
+    let id = r.id()?;
+    r.skip(4)?; // stack trace serial
+    let super_id = r.id()?;
+    let loader_id = r.id()?;
+    r.skip(4 * usize::from(id_size))?; // signers, protection domain, 2 reserved
+    r.skip(4)?; // instance size
+
+    for _ in 0..r.u2()? {
+        r.skip(2)?; // constant pool index
+        let ty = read_type(r)?;
+        r.skip(ty.hprof_size(id_size))?;
+    }
+
+    let mut statics = Vec::new();
+    let mut static_bytes = 0;
+    for _ in 0..r.u2()? {
+        let name = r.id()?;
+        let ty = read_type(r)?;
+        if ty == BasicType::Object {
+            statics.push((name, r.id()?));
+        } else {
+            r.skip(ty.hprof_size(id_size))?;
+            static_bytes += ty.heap_size(id_size);
+        }
+    }
+
+    let mut fields = Vec::new();
+    for _ in 0..r.u2()? {
+        let name = r.id()?;
+        fields.push((name, read_type(r)?));
+    }
+    Ok(ClassDump { id, super_id, loader_id, statics, static_bytes, fields })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::testing::{HprofWriter, MiniHeap};
+    use proptest::prelude::*;
+
+    fn collect(bytes: &[u8]) -> Result<Vec<Record<'_>>, HeapError> {
+        let header = parse_header(bytes)?;
+        let mut out = Vec::new();
+        scan(bytes, &header, &mut |r| {
+            out.push(r);
+            Ok(())
+        })?;
+        Ok(out)
+    }
+
+    fn segment_of(bytes: &[u8]) -> Range<usize> {
+        collect(bytes)
+            .unwrap()
+            .into_iter()
+            .find_map(|r| if let Record::Segment(s) = r { Some(s) } else { None })
+            .expect("a segment")
+    }
+
+    fn all_kinds(w: &mut HprofWriter) {
+        w.string(1, "Node")
+            .string(2, "next")
+            .load_class(0x100, 1)
+            .class(0x100, 0, &[(2, BasicType::Object), (3, BasicType::Long)], &[(2, 0x10)])
+            .instance(0x10, 0x100, &[0u8; 16])
+            .obj_array(0x20, 0x100, &[0x10, 0])
+            .prim_array(0x30, BasicType::Int, 3)
+            .root(HeapRootKind::JavaFrame, 0x10)
+            .root(HeapRootKind::StickyClass, 0x100);
+    }
+
+    #[test]
+    fn parses_header() {
+        let header = parse_header(&MiniHeap::new().bytes()).unwrap();
+        assert_eq!((header.version.as_str(), header.id_size, header.body_start), ("JAVA PROFILE 1.0.2", 8, 31));
+
+        let mut old = b"JAVA PROFILE 1.0.1\0".to_vec();
+        old.extend_from_slice(&[0, 0, 0, 4, 0, 0, 0, 1, 0, 0, 0, 2]);
+        let header = parse_header(&old).unwrap();
+        assert_eq!(
+            (header.version.as_str(), header.id_size, header.timestamp_ms),
+            ("JAVA PROFILE 1.0.1", 4, (1 << 32) + 2)
+        );
+
+        let bad_size = b"JAVA PROFILE 1.0.2\0\0\0\0\x03\0\0\0\0\0\0\0\0";
+        assert!(matches!(parse_header(bad_size), Err(HeapError::UnsupportedIdSize(3))));
+        assert!(matches!(parse_header(b"GIF89a"), Err(HeapError::NotHprof)));
+        assert!(matches!(parse_header(b""), Err(HeapError::NotHprof)));
+        assert!(matches!(parse_header(b"JAVA PROFILE 1.0.3\0\0\0\0\x08"), Err(HeapError::NotHprof)));
+        assert!(matches!(parse_header(b"JAVA PROFILE 1.0.2\0\0\0"), Err(HeapError::Truncated { at: 21 })));
+    }
+
+    #[test]
+    fn scans_every_record_kind() {
+        let mut w = HprofWriter::new();
+        all_kinds(&mut w);
+        let bytes = w.finish();
+        let records = collect(&bytes).unwrap();
+
+        let (mut utf8, mut loads, mut segments) = (0, 0, 0);
+        let (mut classes, mut instances, mut obj_arrays, mut prim_arrays, mut roots) = (0, 0, 0, 0, 0);
+        for record in &records {
+            match record {
+                Record::Utf8 { .. } => utf8 += 1,
+                Record::LoadClass { class_id, name_id } => {
+                    loads += 1;
+                    assert_eq!((*class_id, *name_id), (0x100, 1));
+                }
+                Record::Segment(_) => segments += 1,
+                Record::Heap { record, .. } => match record {
+                    HeapRecord::Class(c) => {
+                        classes += 1;
+                        assert_eq!((c.id, c.super_id, c.loader_id), (0x100, 0, 0));
+                        assert_eq!(c.statics, [(2, 0x10)]);
+                        assert_eq!(c.fields, [(2, BasicType::Object), (3, BasicType::Long)]);
+                    }
+                    HeapRecord::Instance { id, class_id, data } => {
+                        instances += 1;
+                        assert_eq!((*id, *class_id, data.len()), (0x10, 0x100, 16));
+                    }
+                    HeapRecord::ObjectArray { id, class_id, len, data } => {
+                        obj_arrays += 1;
+                        assert_eq!((*id, *class_id, *len, data.len()), (0x20, 0x100, 2, 16));
+                    }
+                    HeapRecord::PrimitiveArray { id, ty, len, data } => {
+                        prim_arrays += 1;
+                        assert_eq!((*id, *ty, *len, data.len()), (0x30, BasicType::Int, 3, 12));
+                    }
+                    HeapRecord::Root { kind, id, thread_serial } => {
+                        roots += 1;
+                        match kind {
+                            HeapRootKind::JavaFrame => assert_eq!((*id, *thread_serial), (0x10, Some(1))),
+                            HeapRootKind::StickyClass => assert_eq!((*id, *thread_serial), (0x100, None)),
+                            other => panic!("unexpected root {other:?}"),
+                        }
+                    }
+                },
+            }
+        }
+        assert_eq!((utf8, loads, segments), (2, 1, 1));
+        assert_eq!((classes, instances, obj_arrays, prim_arrays, roots), (1, 1, 1, 1, 2));
+        assert!(matches!(&records[3], Record::Segment(s) if s.end <= bytes.len()), "segment precedes its contents");
+    }
+
+    #[test]
+    fn root_kinds_carry_the_thread_serial_they_have() {
+        let kinds = [
+            (HeapRootKind::Unknown, false),
+            (HeapRootKind::JniGlobal, false),
+            (HeapRootKind::JniLocal, true),
+            (HeapRootKind::JavaFrame, true),
+            (HeapRootKind::NativeStack, true),
+            (HeapRootKind::StickyClass, false),
+            (HeapRootKind::ThreadBlock, true),
+            (HeapRootKind::MonitorUsed, false),
+            (HeapRootKind::ThreadObject, true),
+        ];
+        let mut w = HprofWriter::new();
+        for (i, (kind, _)) in kinds.iter().enumerate() {
+            w.root_in_thread(*kind, 0x40 + i as u64, 7);
+            w.root(*kind, 0x80 + i as u64);
+        }
+        let bytes = w.finish();
+        let mut seen = Vec::new();
+        for r in collect(&bytes).unwrap() {
+            if let Record::Heap { record: HeapRecord::Root { kind, id, thread_serial }, .. } = r {
+                seen.push((kind, id, thread_serial));
+            }
+        }
+        assert_eq!(seen.len(), kinds.len() * 2);
+        for (i, (kind, carries)) in kinds.iter().enumerate() {
+            let explicit = seen[2 * i];
+            let default = seen[2 * i + 1];
+            assert_eq!(explicit, (*kind, 0x40 + i as u64, carries.then_some(7)));
+            assert_eq!(default, (*kind, 0x80 + i as u64, carries.then_some(1)));
+        }
+    }
+
+    #[test]
+    fn primitive_arrays_expose_their_contents() {
+        let mut w = HprofWriter::new();
+        w.prim_bytes(0x1, BasicType::Char, &[0, b'h', 0, 0xE9]).prim_bytes(0x2, BasicType::Byte, b"v42");
+        let bytes = w.finish();
+        let arrays: Vec<_> = collect(&bytes)
+            .unwrap()
+            .into_iter()
+            .filter_map(|r| match r {
+                Record::Heap { record: HeapRecord::PrimitiveArray { ty, len, data, .. }, .. } => {
+                    Some((ty, len, data.to_vec()))
+                }
+                _ => None,
+            })
+            .collect();
+        assert_eq!(arrays, [(BasicType::Char, 2, vec![0, b'h', 0, 0xE9]), (BasicType::Byte, 3, b"v42".to_vec())]);
+    }
+
+    #[test]
+    fn mini_heap_builds_strings_and_defines_object_and_class() {
+        let mut heap = MiniHeap::new();
+        heap.string(0x10, "v42", true).string(0x20, "hé", false);
+        let bytes = heap.bytes();
+        let records = collect(&bytes).unwrap();
+        let names: Vec<_> = records
+            .iter()
+            .filter_map(|r| {
+                if let Record::Utf8 { text, .. } = r { Some(String::from_utf8_lossy(text).into_owned()) } else { None }
+            })
+            .collect();
+        assert!(names.iter().any(|n| n == "java/lang/Object") && names.iter().any(|n| n == "java/lang/Class"));
+        let arrays: Vec<_> = records
+            .iter()
+            .filter_map(|r| match r {
+                Record::Heap { record: HeapRecord::PrimitiveArray { data, .. }, .. } => Some(data.to_vec()),
+                _ => None,
+            })
+            .collect();
+        assert_eq!(arrays, [b"v42".to_vec(), vec![b'h', 0, 0xE9, 0]]);
+        let coders: Vec<_> = records
+            .iter()
+            .filter_map(|r| match r {
+                Record::Heap { record: HeapRecord::Instance { id, data, .. }, .. } => Some((*id, data[8])),
+                _ => None,
+            })
+            .collect();
+        assert_eq!(coders, [(0x10, 0), (0x20, 1)]);
+    }
+
+    #[test]
+    fn mini_heap_lays_out_fields_class_first_then_super() {
+        let mut heap = MiniHeap::new();
+        heap.class("A", Some("java.lang.Object"), &[("a", BasicType::Int)])
+            .class("B", Some("A"), &[("b", BasicType::Object), ("c", BasicType::Byte)])
+            .object(0x10, "B", &[("a", 7), ("b", 0x99), ("c", 0x1FF)]);
+        let bytes = heap.bytes();
+        let data = collect(&bytes)
+            .unwrap()
+            .into_iter()
+            .find_map(|r| match r {
+                Record::Heap { record: HeapRecord::Instance { data, .. }, .. } => Some(data.to_vec()),
+                _ => None,
+            })
+            .unwrap();
+        assert_eq!(data, [0, 0, 0, 0, 0, 0, 0, 0x99, 0xFF, 0, 0, 0, 7]);
+    }
+
+    #[test]
+    fn record_at_reads_back_an_instance() {
+        let mut w = HprofWriter::new();
+        all_kinds(&mut w);
+        let bytes = w.finish();
+        let at = collect(&bytes)
+            .unwrap()
+            .into_iter()
+            .find_map(|r| match r {
+                Record::Heap { at, record: HeapRecord::Instance { .. } } => Some(at),
+                _ => None,
+            })
+            .unwrap();
+        match record_at(&bytes, at, 8).unwrap() {
+            HeapRecord::Instance { id, class_id, data } => assert_eq!((id, class_id, data.len()), (0x10, 0x100, 16)),
+            other => panic!("expected an instance, got {other:?}"),
+        }
+        assert!(matches!(record_at(&bytes, bytes.len(), 8), Err(HeapError::Truncated { .. })));
+        assert!(matches!(record_at(&bytes, at, 3), Err(HeapError::UnsupportedIdSize(3))));
+    }
+
+    #[test]
+    fn scan_segment_visits_the_same_records_scan_does() {
+        let mut w = HprofWriter::new();
+        all_kinds(&mut w);
+        let bytes = w.finish();
+        let from_scan: Vec<_> = collect(&bytes)
+            .unwrap()
+            .into_iter()
+            .filter_map(|r| if let Record::Heap { at, record } = r { Some((at, record)) } else { None })
+            .collect();
+        let mut from_segment = Vec::new();
+        scan_segment(&bytes, segment_of(&bytes), 8, &mut |at, record| {
+            from_segment.push((at, record));
+            Ok(())
+        })
+        .unwrap();
+        assert_eq!(from_scan.len(), 6);
+        assert_eq!(from_scan, from_segment);
+    }
+
+    #[test]
+    fn visitor_errors_stop_the_scan() {
+        let bytes = MiniHeap::new().bytes();
+        let header = parse_header(&bytes).unwrap();
+        let mut calls = 0;
+        let result = scan(&bytes, &header, &mut |_| {
+            calls += 1;
+            Err(HeapError::Cancelled)
+        });
+        assert!(matches!(result, Err(HeapError::Cancelled)));
+        assert_eq!(calls, 1);
+    }
+
+    #[test]
+    fn both_heap_dump_tags_are_segments() {
+        let bytes = MiniHeap::new().bytes();
+        let segment = segment_of(&bytes);
+        let mut old_style = bytes.clone();
+        old_style[segment.start - 9] = 0x0C;
+        assert_eq!(segment_of(&old_style), segment);
+    }
+
+    #[test]
+    fn id_size_4_dump_parses() {
+        let mut w = HprofWriter::with_id_size(4);
+        w.string(1, "Node")
+            .load_class(0x100, 1)
+            .class(0x100, 0, &[(2, BasicType::Object), (3, BasicType::Long)], &[(2, 0x10)])
+            .instance(0x10, 0x100, &[0u8; 12])
+            .obj_array(0x20, 0x100, &[0x10, 0, 0xFFFF_FFFF])
+            .prim_array(0x30, BasicType::Short, 5)
+            .root(HeapRootKind::ThreadObject, 0x10);
+        let bytes = w.finish();
+        let header = parse_header(&bytes).unwrap();
+        assert_eq!((header.id_size, header.body_start), (4, 31));
+        let records: Vec<_> = collect(&bytes)
+            .unwrap()
+            .into_iter()
+            .filter_map(|r| if let Record::Heap { record, .. } = r { Some(record) } else { None })
+            .collect();
+        assert_eq!(records.len(), 5);
+        match &records[0] {
+            HeapRecord::Class(c) => {
+                assert_eq!((c.id, c.statics.as_slice(), c.fields.len()), (0x100, &[(2, 0x10)][..], 2));
+            }
+            other => panic!("expected the class, got {other:?}"),
+        }
+        assert!(matches!(&records[1], HeapRecord::Instance { id: 0x10, class_id: 0x100, data } if data.len() == 12));
+        assert!(
+            matches!(&records[2], HeapRecord::ObjectArray { len: 3, data, .. } if data.len() == 12 && data[8..] == [0xFF; 4])
+        );
+        assert!(
+            matches!(&records[3], HeapRecord::PrimitiveArray { ty: BasicType::Short, len: 5, data, .. } if data.len() == 10)
+        );
+        assert!(matches!(
+            &records[4],
+            HeapRecord::Root { kind: HeapRootKind::ThreadObject, id: 0x10, thread_serial: Some(1) }
+        ));
+    }
+
+    fn push_id(out: &mut Vec<u8>, v: u64) {
+        out.extend_from_slice(&v.to_be_bytes());
+    }
+
+    /// A class dump written by hand: loader, a constant-pool entry, and statics of each shape.
+    fn class_dump_bytes(last_field_type: u8) -> Vec<u8> {
+        let mut b = vec![0x20];
+        push_id(&mut b, 0x500);
+        b.extend_from_slice(&0u32.to_be_bytes());
+        for id in [0x400, 0x77, 0, 0, 0, 0] {
+            push_id(&mut b, id); // super, loader, signers, domain, 2 reserved
+        }
+        b.extend_from_slice(&24u32.to_be_bytes());
+        b.extend_from_slice(&[0, 1, 0, 7, 10, 0, 0, 0, 99]); // constant pool: one int
+        b.extend_from_slice(&[0, 3]);
+        push_id(&mut b, 0xA);
+        b.push(2);
+        push_id(&mut b, 0xBEEF); // object static
+        push_id(&mut b, 0xB);
+        b.push(10);
+        b.extend_from_slice(&5u32.to_be_bytes()); // int static
+        push_id(&mut b, 0xC);
+        b.push(11);
+        b.extend_from_slice(&6u64.to_be_bytes()); // long static
+        b.extend_from_slice(&[0, 2]);
+        push_id(&mut b, 0xD);
+        b.push(2);
+        push_id(&mut b, 0xE);
+        b.push(last_field_type);
+        b
+    }
+
+    #[test]
+    fn class_dump_reads_loader_statics_and_skips_the_constant_pool() {
+        let bytes = class_dump_bytes(11);
+        let HeapRecord::Class(c) = record_at(&bytes, 0, 8).unwrap() else { panic!("expected a class") };
+        assert_eq!((c.id, c.super_id, c.loader_id), (0x500, 0x400, 0x77));
+        assert_eq!(c.statics, [(0xA, 0xBEEF)]);
+        assert_eq!(c.static_bytes, 12);
+        assert_eq!(c.fields, [(0xD, BasicType::Object), (0xE, BasicType::Long)]);
+    }
+
+    #[test]
+    fn unknown_value_type_reports_its_offset() {
+        let bytes = class_dump_bytes(0x63);
+        let err = record_at(&bytes, 0, 8).unwrap_err();
+        assert!(matches!(err, HeapError::UnknownType { ty: 0x63, at } if at == bytes.len() as u64 - 1), "{err:?}");
+    }
+
+    #[test]
+    fn truncated_dump_reports_offset() {
+        let mut bytes = MiniHeap::new().bytes();
+        let body_start = parse_header(&bytes).unwrap().body_start as u64;
+        bytes.truncate(bytes.len() - 5);
+        let err = collect(&bytes).unwrap_err();
+        assert!(matches!(err, HeapError::Truncated { at } if at > body_start && at == bytes.len() as u64), "{err:?}");
+        assert!(err.to_string().contains(&bytes.len().to_string()));
+    }
+
+    /// HotSpot flushes whole segments, so a JVM killed while dumping leaves a file that ends
+    /// right after one, without HEAP DUMP END.
+    #[test]
+    fn a_dump_the_jvm_did_not_finish_ends_early() {
+        let mut bytes = MiniHeap::new().bytes();
+        bytes.truncate(bytes.len() - RECORD_HEADER_LEN);
+        let err = collect(&bytes).unwrap_err();
+        assert!(matches!(err, HeapError::Truncated { at } if at == bytes.len() as u64), "{err:?}");
+        assert!(err.to_string().starts_with("the dump ends early"), "{err}");
+    }
+
+    #[test]
+    fn every_prefix_is_ok_or_truncated_at_its_length() {
+        let mut heap = MiniHeap::new();
+        heap.class("N", Some("java.lang.Object"), &[("next", BasicType::Object)])
+            .object(0x10, "N", &[("next", 0x20)])
+            .string(0x20, "x", true)
+            .root(HeapRootKind::ThreadObject, 0x10);
+        let bytes = heap.bytes();
+        let header = parse_header(&bytes).unwrap();
+        let segment = segment_of(&bytes);
+        for cut in header.body_start..bytes.len() {
+            let prefix = &bytes[..cut];
+            match scan(prefix, &header, &mut |_| Ok(())) {
+                // Only a prefix without heap data is a complete (empty) dump.
+                Ok(()) => assert!(cut < segment.start, "cut at {cut} has heap data but no end"),
+                Err(HeapError::Truncated { at }) => assert_eq!(at, cut as u64, "cut at {cut}"),
+                Err(other) => panic!("cut at {cut}: {other:?}"),
+            }
+        }
+        let past_end = HprofHeader { body_start: bytes.len() + 1, ..header };
+        assert!(matches!(scan(&bytes, &past_end, &mut |_| Ok(())), Err(HeapError::Truncated { .. })));
+    }
+
+    #[test]
+    fn sub_record_overrunning_its_segment_is_truncated() {
+        let mut bytes = MiniHeap::new().bytes();
+        let segment = segment_of(&bytes);
+        let len_at = segment.start - 4;
+        let shrunk = (segment.len() as u32 - 3).to_be_bytes();
+        bytes[len_at..segment.start].copy_from_slice(&shrunk);
+        let err = collect(&bytes).unwrap_err();
+        assert!(matches!(err, HeapError::Truncated { at } if at == segment.end as u64 - 3), "{err:?}");
+    }
+
+    #[test]
+    fn unknown_sub_record_is_an_error() {
+        let mut bytes = MiniHeap::new().bytes();
+        let segment = segment_of(&bytes);
+        bytes[segment.start] = 0x99;
+        let err = collect(&bytes).unwrap_err();
+        assert!(matches!(err, HeapError::UnknownSubRecord { tag: 0x99, at } if at == segment.start as u64), "{err:?}");
+        assert!(err.to_string().contains("0x99"));
+    }
+
+    #[test]
+    fn array_lengths_larger_than_the_dump_do_not_allocate_or_panic() {
+        let mut b = vec![0x23];
+        push_id(&mut b, 1);
+        b.extend_from_slice(&0u32.to_be_bytes());
+        b.extend_from_slice(&u32::MAX.to_be_bytes());
+        b.push(11);
+        assert!(matches!(record_at(&b, 0, 8), Err(HeapError::Truncated { .. })));
+        b[0] = 0x22;
+        assert!(matches!(record_at(&b, 0, 8), Err(HeapError::Truncated { .. })));
+    }
+
+    #[test]
+    fn primitive_array_of_objects_is_an_unknown_type() {
+        let mut b = vec![0x23];
+        push_id(&mut b, 1);
+        b.extend_from_slice(&[0, 0, 0, 0, 0, 0, 0, 0]);
+        b.push(2);
+        assert!(matches!(record_at(&b, 0, 8), Err(HeapError::UnknownType { ty: 2, .. })));
+    }
+
+    #[test]
+    fn basic_type_sizes_and_names() {
+        assert_eq!(BasicType::from_u8(10), Some(BasicType::Int));
+        assert_eq!([0, 1, 3, 12, 255].map(BasicType::from_u8), [None; 5]);
+        assert_eq!((BasicType::Object.hprof_size(4), BasicType::Object.hprof_size(8)), (4, 8));
+        assert_eq!((BasicType::Object.heap_size(4), BasicType::Long.heap_size(4)), (4, 8));
+        assert_eq!((BasicType::Char.hprof_size(8), BasicType::Boolean.heap_size(8)), (2, 1));
+        assert_eq!(
+            (BasicType::Int.array_name(), BasicType::Boolean.array_name(), BasicType::Long.array_name()),
+            ("[I", "[Z", "[J")
+        );
+    }
+
+    fn sample() -> Vec<u8> {
+        let mut heap = MiniHeap::new();
+        heap.class("Node", Some("java.lang.Object"), &[("next", BasicType::Object), ("v", BasicType::Long)])
+            .class("[LNode;", Some("java.lang.Object"), &[])
+            .object(0x10, "Node", &[("next", 0x20), ("v", 5)])
+            .object(0x20, "Node", &[])
+            .array_of(0x30, "[LNode;", &[0x10, 0x20, 0])
+            .prim(0x40, BasicType::Int, 3)
+            .string(0x50, "hello", true)
+            .root(HeapRootKind::ThreadObject, 0x10)
+            .root(HeapRootKind::JniGlobal, 0x30);
+        heap.bytes()
+    }
+
+    fn with_header(id_size: u8, body: &[u8]) -> Vec<u8> {
+        let mut bytes = MAGIC_1_0_2.to_vec();
+        bytes.extend_from_slice(&u32::from(id_size).to_be_bytes());
+        bytes.extend_from_slice(&1_700_000_000_000u64.to_be_bytes());
+        bytes.extend_from_slice(body);
+        bytes
+    }
+
+    proptest! {
+        #[test]
+        fn parser_never_panics(bytes in proptest::collection::vec(any::<u8>(), 0..4096)) {
+            let _ = parse_header(&bytes).and_then(|h| scan(&bytes, &h, &mut |_| Ok(())));
+        }
+
+        /// Random bytes alone fail the magic check, so this puts a valid header in front to make
+        /// `scan` and `scan_segment` run on arbitrary record streams.
+        #[test]
+        fn random_bytes_after_a_valid_header_never_panic(
+            id_size in prop_oneof![Just(4u8), Just(8u8)],
+            body in proptest::collection::vec(any::<u8>(), 0..4096),
+        ) {
+            let bytes = with_header(id_size, &body);
+            let header = parse_header(&bytes).unwrap();
+            prop_assert_eq!(header.id_size, id_size);
+            let _ = scan(&bytes, &header, &mut |_| Ok(()));
+            let _ = scan_segment(&bytes, header.body_start..bytes.len(), id_size, &mut |_, _| Ok(()));
+            for at in header.body_start..bytes.len() {
+                let _ = record_at(&bytes, at, id_size);
+            }
+        }
+
+        /// Same, but every record has a correct length so the random bytes reach the heap-dump
+        /// segments and the sub-record parsers instead of stopping at the first bad length.
+        #[test]
+        fn random_records_after_a_valid_header_never_panic(
+            id_size in prop_oneof![Just(4u8), Just(8u8)],
+            records in proptest::collection::vec(
+                (
+                    prop_oneof![Just(0x01u8), Just(0x02), Just(0x0C), Just(0x1C), Just(0x2C), any::<u8>()],
+                    prop_oneof![
+                        Just(0x20u8), Just(0x21), Just(0x22), Just(0x23), Just(0x01), Just(0x02), Just(0x03),
+                        Just(0x04), Just(0x05), Just(0x06), Just(0x07), Just(0x08), Just(0xFF), any::<u8>()
+                    ],
+                    proptest::collection::vec(any::<u8>(), 0..96),
+                ),
+                0..12,
+            ),
+        ) {
+            let mut body = Vec::new();
+            for (tag, first, rest) in records {
+                let mut record_body = vec![first];
+                record_body.extend_from_slice(&rest);
+                body.push(tag);
+                body.extend_from_slice(&0u32.to_be_bytes());
+                body.extend_from_slice(&(record_body.len() as u32).to_be_bytes());
+                body.extend_from_slice(&record_body);
+            }
+            let bytes = with_header(id_size, &body);
+            let header = parse_header(&bytes).unwrap();
+            let mut segments = Vec::new();
+            let _ = scan(&bytes, &header, &mut |r| {
+                if let Record::Segment(s) = r {
+                    segments.push(s);
+                }
+                Ok(())
+            });
+            for s in segments {
+                let _ = scan_segment(&bytes, s, id_size, &mut |_, _| Ok(()));
+            }
+        }
+
+        #[test]
+        fn mutated_dump_never_panics(flips in proptest::collection::vec((0usize..2048, any::<u8>()), 1..16)) {
+            let pristine = sample();
+            let header = parse_header(&pristine).unwrap();
+            let segment = segment_of(&pristine);
+            let mut bytes = pristine.clone();
+            for (pos, value) in flips {
+                let at = pos % bytes.len();
+                bytes[at] = value;
+            }
+            let _ = scan(&bytes, &header, &mut |_| Ok(()));
+            for id_size in [4, 8] {
+                let _ = scan_segment(&bytes, segment.clone(), id_size, &mut |_, _| Ok(()));
+            }
+            for at in 0..bytes.len() {
+                let _ = record_at(&bytes, at, 8);
+            }
+        }
+    }
+}
@@ -0,0 +1,327 @@
+//! Maps hprof object ids (addresses) to dense `u32` indices in ascending address order.
+//!
+//! Object ids are addresses and objects are at least 8 bytes apart, so one bit per 8-byte slot
+//! ranks ids at about a byte per object. Addresses are grouped in 1 MiB chunks of
+//! [`WORDS_PER_CHUNK`] words; only chunks holding an id get words, and a dense table maps a chunk
+//! to its first word. The rank of an id is the number of set bits before its word plus the set
+//! bits below it in the word. Ids that are not multiples of 8, or that span too wide an address
+//! range for the table, use a sorted array and binary search instead.
+
+mod builder;
+
+pub(crate) use builder::ChunkHasher;
+pub use builder::IdMapBuilder;
+
+use crate::column::{Column, NONE};
+
+const SLOT_SHIFT: u32 = 3;
+const CHUNK_SLOT_BITS: u32 = 17;
+const WORD_BITS: u64 = 64;
+const WORDS_PER_CHUNK: usize = (1 << CHUNK_SLOT_BITS) / WORD_BITS as usize;
+
+// The bit table is matched in hot paths; boxing the big variant would only add
+// an indirection to save stack space `IdMap` never holds by value in bulk.
+#[allow(clippy::large_enum_variant)]
+enum Repr {
+    Bits {
+        /// Chunk number of the first table entry.
+        min_chunk: u64,
+        words: Column<u64>,
+        /// Set bits before each word.
+        wrank: Column<u32>,
+        /// Per chunk from `min_chunk` up: index of its first word, or `NONE` when it has no id.
+        table: Column<u32>,
+        /// Per chunk that has words, in word order: its offset from `min_chunk`.
+        chunk_offsets: Column<u32>,
+        len: u32,
+    },
+    Sorted(Column<u64>),
+}
+
+/// Ids ranked in ascending address order: the index of an object is its position in the order.
+pub struct IdMap {
+    repr: Repr,
+}
+
+impl IdMap {
+    pub fn len(&self) -> u32 {
+        match &self.repr {
+            Repr::Bits { len, .. } => *len,
+            Repr::Sorted(ids) => ids.len() as u32,
+        }
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.len() == 0
+    }
+
+    /// The index of `id`, or `None` when it is not in the map.
+    pub fn rank(&self, id: u64) -> Option<u32> {
+        match &self.repr {
+            Repr::Sorted(ids) => ids.binary_search(&id).ok().map(|r| r as u32),
+            Repr::Bits { min_chunk, words, wrank, table, .. } => {
+                if id & 7 != 0 {
+                    return None;
+                }
+                let slot = id >> SLOT_SHIFT;
+                let offset = (slot >> CHUNK_SLOT_BITS).checked_sub(*min_chunk)?;
+                let first = *table.get(usize::try_from(offset).ok()?)?;
+                if first == NONE {
+                    return None;
+                }
+                let w = first as usize + (slot >> 6) as usize % WORDS_PER_CHUNK;
+                let bit = slot & 63;
+                let word = words[w];
+                if word >> bit & 1 == 0 {
+                    return None;
+                }
+                Some(wrank[w] + (word & ((1u64 << bit) - 1)).count_ones())
+            }
+        }
+    }
+
+    /// The id with index `rank`. Panics when `rank >= len()`.
+    pub fn id_at(&self, rank: u32) -> u64 {
+        match &self.repr {
+            Repr::Sorted(ids) => ids[rank as usize],
+            Repr::Bits { words, wrank, len, .. } => {
+                assert!(rank < *len, "rank {rank} out of range for {len} ids");
+                let w = wrank.partition_point(|&before| before <= rank) - 1;
+                let mut word = words[w];
+                for _ in 0..rank - wrank[w] {
+                    word &= word - 1;
+                }
+                (self.word_first_slot(w) + u64::from(word.trailing_zeros())) << SLOT_SHIFT
+            }
+        }
+    }
+
+    /// All ids in ascending order.
+    pub fn ids(&self) -> impl Iterator<Item = u64> + '_ {
+        Ids { map: self, next: 0, slot_base: 0, rest: 0 }
+    }
+
+    /// Bytes of all the columns, in RAM or on disk.
+    pub fn column_bytes(&self) -> u64 {
+        match &self.repr {
+            Repr::Sorted(ids) => std::mem::size_of_val(&**ids) as u64,
+            Repr::Bits { words, wrank, table, chunk_offsets, .. } => {
+                (std::mem::size_of_val(&**words)
+                    + std::mem::size_of_val(&**wrank)
+                    + std::mem::size_of_val(&**table)
+                    + std::mem::size_of_val(&**chunk_offsets)) as u64
+            }
+        }
+    }
+
+    /// Slot number of bit 0 of word `w` (bitmap layout only).
+    fn word_first_slot(&self, w: usize) -> u64 {
+        let Repr::Bits { min_chunk, chunk_offsets, .. } = &self.repr else {
+            unreachable!("sorted layout has no words")
+        };
+        let chunk = min_chunk + u64::from(chunk_offsets[w / WORDS_PER_CHUNK]);
+        (chunk << CHUNK_SLOT_BITS) + (w % WORDS_PER_CHUNK) as u64 * WORD_BITS
+    }
+}
+
+struct Ids<'a> {
+    map: &'a IdMap,
+    /// Next word (bitmap layout) or next id (sorted layout).
+    next: usize,
+    slot_base: u64,
+    /// Bits of the current word not yielded yet.
+    rest: u64,
+}
+
+impl Iterator for Ids<'_> {
+    type Item = u64;
+
+    fn next(&mut self) -> Option<u64> {
+        match &self.map.repr {
+            Repr::Sorted(ids) => {
+                let id = *ids.get(self.next)?;
+                self.next += 1;
+                Some(id)
+            }
+            Repr::Bits { words, .. } => loop {
+                if self.rest != 0 {
+                    let bit = u64::from(self.rest.trailing_zeros());
+                    self.rest &= self.rest - 1;
+                    return Some((self.slot_base + bit) << SLOT_SHIFT);
+                }
+                self.rest = *words.get(self.next)?;
+                self.slot_base = self.map.word_first_slot(self.next);
+                self.next += 1;
+            },
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use ide_core::CancellationToken;
+
+    use super::*;
+    use crate::column::Store;
+
+    /// Deterministic generator so failures replay; the crate has no `rand`.
+    struct XorShift(u64);
+
+    impl XorShift {
+        fn next(&mut self) -> u64 {
+            self.0 ^= self.0 << 13;
+            self.0 ^= self.0 >> 7;
+            self.0 ^= self.0 << 17;
+            self.0
+        }
+    }
+
+    fn build(ids: &[u64], store: &Store) -> IdMap {
+        let mut builder = IdMapBuilder::new();
+        for &id in ids {
+            builder.insert(id);
+        }
+        builder.build(store, &CancellationToken::new()).unwrap()
+    }
+
+    fn assert_matches_model(map: &IdMap, model: &[u64]) {
+        assert_eq!(map.len() as usize, model.len());
+        assert_eq!(map.ids().collect::<Vec<_>>(), model);
+        for (rank, &id) in model.iter().enumerate() {
+            assert_eq!(map.rank(id), Some(rank as u32), "rank of {id:#x}");
+            assert_eq!(map.id_at(rank as u32), id, "id at {rank}");
+            for near in [id.wrapping_sub(8), id.wrapping_add(8), id.wrapping_add(1), id.wrapping_add(4)] {
+                if model.binary_search(&near).is_err() {
+                    assert_eq!(map.rank(near), None, "{near:#x} was never inserted");
+                }
+            }
+        }
+    }
+
+    #[test]
+    fn rank_follows_address_order() {
+        let map = build(&[0x2000, 0x1000, 0x1010, 0x7_0000_0000], &Store::Ram);
+        assert_eq!(map.len(), 4);
+        assert_eq!(
+            [0x1000, 0x1010, 0x2000, 0x7_0000_0000].map(|id| map.rank(id)),
+            [Some(0), Some(1), Some(2), Some(3)]
+        );
+        assert_eq!(map.rank(0x1008), None);
+        assert_eq!(map.rank(0), None);
+        assert_eq!(map.id_at(2), 0x2000);
+        assert_eq!(map.ids().collect::<Vec<_>>(), [0x1000, 0x1010, 0x2000, 0x7_0000_0000]);
+    }
+
+    #[test]
+    fn unaligned_ids_use_sorted_fallback() {
+        let map = build(&[0x2000, 0x1004, 0x1000, 0x1010, 0x7_0000_0000], &Store::Ram);
+        assert_eq!(
+            [0x1000, 0x1004, 0x1010, 0x2000, 0x7_0000_0000].map(|id| map.rank(id)),
+            [Some(0), Some(1), Some(2), Some(3), Some(4)]
+        );
+        assert_eq!(map.rank(0x1008), None);
+        assert_eq!(map.id_at(1), 0x1004);
+    }
+
+    #[test]
+    fn sparse_wide_span_uses_fallback() {
+        let map = build(&[0x1000 + (1 << 50), 0x1000], &Store::Ram);
+        assert_eq!(map.rank(0x1000), Some(0));
+        assert_eq!(map.rank(0x1000 + (1 << 50)), Some(1));
+        assert_eq!(map.rank(0x1008), None);
+        assert!(map.column_bytes() < 64, "a two-id map must not allocate a chunk table");
+    }
+
+    #[test]
+    fn ten_million_ids_cost_about_a_byte_each() {
+        let mut builder = IdMapBuilder::new();
+        for i in 0..10_000_000u64 {
+            builder.insert(0x7_0000_0000 + i * 32);
+        }
+        let map = builder.build(&Store::Ram, &CancellationToken::new()).unwrap();
+        assert_eq!(map.len(), 10_000_000);
+        assert!(map.column_bytes() < 16_000_000, "{} bytes", map.column_bytes());
+        assert_eq!(map.rank(0x7_0000_0000 + 9_999_999 * 32), Some(9_999_999));
+        assert_eq!(map.id_at(5_000_000), 0x7_0000_0000 + 5_000_000 * 32);
+    }
+
+    #[test]
+    fn empty_map_and_duplicates() {
+        let map = build(&[], &Store::Ram);
+        assert!(map.is_empty());
+        assert_eq!(map.rank(0x1000), None);
+        assert_eq!(map.ids().count(), 0);
+        let map = build(&[0x1000, 0x1000, 0x1008, 0x1000], &Store::Ram);
+        assert_eq!(map.len(), 2);
+        assert_eq!(map.rank(0x1008), Some(1));
+    }
+
+    #[test]
+    fn chunk_boundaries_and_gaps_between_chunks() {
+        let ids = [0, 0xF_FFF8, 0x10_0000, 0x10_0008, 0x50_0000, 0x1_0000_0000];
+        let mut model = ids.to_vec();
+        model.sort_unstable();
+        let map = build(&ids, &Store::Ram);
+        assert_matches_model(&map, &model);
+    }
+
+    #[test]
+    fn ids_outside_the_populated_chunks_are_absent() {
+        let ids = [0x30_0000, 0x30_0008, 0x70_0000];
+        let map = build(&ids, &Store::Ram);
+        assert_matches_model(&map, &ids);
+        // Chunk 5 lies between the lowest and highest chunk but holds nothing.
+        assert_eq!(map.rank(0x50_0000), None);
+        // Chunk 1 is below the lowest populated chunk.
+        assert_eq!(map.rank(0x10_0000), None);
+        // Chunk 9 is past the last table entry.
+        assert_eq!(map.rank(0x900_0000), None);
+        assert_eq!(map.rank(u64::MAX - 7), None);
+    }
+
+    #[test]
+    fn highest_address_round_trips() {
+        let map = build(&[0xFFFF_FFFF_FFFF_FFF8], &Store::Ram);
+        assert_eq!(map.rank(0xFFFF_FFFF_FFFF_FFF8), Some(0));
+        assert_eq!(map.id_at(0), 0xFFFF_FFFF_FFFF_FFF8);
+        let map = build(&[0, 0xFFFF_FFFF_FFFF_FFF8], &Store::Ram);
+        assert_matches_model(&map, &[0, 0xFFFF_FFFF_FFFF_FFF8]);
+    }
+
+    #[test]
+    fn random_ids_match_a_sorted_model_in_ram_and_on_disk() {
+        for seed in 1..=8u64 {
+            let mut rng = XorShift(seed.wrapping_mul(0x9E37_79B9_7F4A_7C15) | 1);
+            let aligned = seed % 4 != 0;
+            let wide = seed % 3 == 0;
+            let ids: Vec<u64> = (0..3000)
+                .map(|_| {
+                    let r = rng.next();
+                    let base = if wide { (r >> 20) << 12 } else { 0x6_0000_0000 + (r % (6 << 20)) };
+                    if aligned { base & !7 } else { base }
+                })
+                .collect();
+            let mut model = ids.clone();
+            model.sort_unstable();
+            model.dedup();
+            let dir = tempfile::tempdir().unwrap();
+            for store in [Store::Ram, Store::Dir(dir.path().to_path_buf())] {
+                assert_matches_model(&build(&ids, &store), &model);
+            }
+        }
+    }
+
+    #[test]
+    fn dir_store_keeps_the_index_columns_under_stable_names() {
+        let dir = tempfile::tempdir().unwrap();
+        let map = build(&[0x1000, 0x2000], &Store::Dir(dir.path().to_path_buf()));
+        let mut names: Vec<_> =
+            std::fs::read_dir(dir.path()).unwrap().map(|e| e.unwrap().file_name().into_string().unwrap()).collect();
+        names.sort();
+        assert_eq!(names, ["ids.bits.u64", "ids.chunks.u32", "ids.table.u32", "ids.wrank.u32"]);
+        assert_eq!(map.rank(0x2000), Some(1));
+        let dir = tempfile::tempdir().unwrap();
+        build(&[0x1004], &Store::Dir(dir.path().to_path_buf()));
+        assert!(dir.path().join("ids.sorted.u64").exists());
+    }
+}
@@ -0,0 +1,409 @@
+//! Collects object ids in any order and lays them out as an [`IdMap`].
+
+use std::cmp::Reverse;
+use std::collections::{BinaryHeap, HashMap};
+use std::hash::{BuildHasherDefault, Hasher};
+
+use ide_core::CancellationToken;
+
+use super::{CHUNK_SLOT_BITS, IdMap, Repr, SLOT_SHIFT, WORD_BITS, WORDS_PER_CHUNK};
+use crate::column::{NONE, Store};
+use crate::error::HeapError;
+use crate::files::{IDS_BITS_FILE, IDS_CHUNKS_FILE, IDS_SORTED_FILE, IDS_TABLE_FILE, IDS_WRANK_FILE};
+
+/// Widest chunk range the dense table covers (256 MiB of table at most).
+const MAX_CHUNK_SPAN: u64 = 1 << 26;
+/// Chunks whose first word index still fits in the table's `u32`.
+const MAX_CHUNKS: usize = (NONE as usize - 1) / WORDS_PER_CHUNK;
+/// Largest object count: `NONE` and the one after it stay free as sentinels for the columns above.
+const MAX_IDS: u64 = NONE as u64 - 2;
+
+const CHUNK_BYTES: u64 = (WORDS_PER_CHUNK * size_of::<u64>()) as u64;
+/// The bitmaps may take this much before any id is inserted, then this many bytes per distinct id.
+/// A heap that fits in 4 GiB of address space never reaches the base allowance whatever the
+/// insertion order; sparse input (one id per chunk) is moved to plain `u64` ids, 8 bytes each,
+/// after at most `BITMAP_BASE_BYTES + BITMAP_BYTES_PER_ID * ids` of bitmaps, so builder memory is
+/// bounded by the number of ids rather than by the address range they spread over.
+const BITMAP_BASE_BYTES: u64 = 64 << 20;
+const BITMAP_BYTES_PER_ID: u64 = 32;
+
+/// Longest stretch of work between two looks at the cancellation token.
+const CHECK_EVERY: usize = 1 << 20;
+
+/// Chunk numbers are already well spread by a multiplicative hash; SipHash would dominate the
+/// per-insert cost on a 100M-object heap. The product keeps the low zero bits of its input, so
+/// the high half is folded down for the bucket index. The graph builder keys class and string
+/// ids (addresses) with it for the same reason.
+#[derive(Default)]
+pub(crate) struct ChunkHasher(u64);
+
+impl Hasher for ChunkHasher {
+    fn finish(&self) -> u64 {
+        self.0 ^ (self.0 >> 32)
+    }
+
+    fn write(&mut self, bytes: &[u8]) {
+        for &b in bytes {
+            self.write_u64(self.0.rotate_left(8) ^ u64::from(b));
+        }
+    }
+
+    fn write_u64(&mut self, n: u64) {
+        self.0 = n.wrapping_mul(0x9E37_79B9_7F4A_7C15);
+    }
+}
+
+struct BitChunks {
+    /// Chunk number → position in `words`.
+    index: HashMap<u64, usize, BuildHasherDefault<ChunkHasher>>,
+    words: Vec<Box<[u64]>>,
+    min: u64,
+    max: u64,
+    /// The chunk of the previous insert; ids usually arrive in runs.
+    last: Option<(u64, usize)>,
+    distinct: u64,
+}
+
+impl BitChunks {
+    fn new() -> Self {
+        Self { index: HashMap::default(), words: Vec::new(), min: u64::MAX, max: 0, last: None, distinct: 0 }
+    }
+
+    /// `false` when the id's chunk does not fit the dense table or the bitmap budget; the id is
+    /// then not recorded and the caller moves to plain ids.
+    fn insert(&mut self, id: u64) -> bool {
+        let slot = id >> SLOT_SHIFT;
+        let chunk = slot >> CHUNK_SLOT_BITS;
+        let at = match self.last {
+            Some((c, at)) if c == chunk => at,
+            _ => match self.index.get(&chunk) {
+                Some(&at) => at,
+                None => {
+                    let (min, max) = (self.min.min(chunk), self.max.max(chunk));
+                    let bitmap_bytes = (self.words.len() as u64 + 1) * CHUNK_BYTES;
+                    if self.words.len() >= MAX_CHUNKS
+                        || max - min >= MAX_CHUNK_SPAN
+                        || bitmap_bytes > BITMAP_BASE_BYTES + BITMAP_BYTES_PER_ID * self.distinct
+                    {
+                        return false;
+                    }
+                    (self.min, self.max) = (min, max);
+                    self.words.push(vec![0u64; WORDS_PER_CHUNK].into_boxed_slice());
+                    self.index.insert(chunk, self.words.len() - 1);
+                    self.words.len() - 1
+                }
+            },
+        };
+        self.last = Some((chunk, at));
+        let word = &mut self.words[at][(slot >> 6) as usize % WORDS_PER_CHUNK];
+        let bit = 1u64 << (slot & 63);
+        if *word & bit == 0 {
+            *word |= bit;
+            self.distinct += 1;
+        }
+        true
+    }
+
+    /// `(chunk number, position in words)` in ascending address order.
+    fn sorted_chunks(&self) -> Vec<(u64, usize)> {
+        let mut chunks: Vec<_> = self.index.iter().map(|(&chunk, &at)| (chunk, at)).collect();
+        chunks.sort_unstable();
+        chunks
+    }
+
+    /// The recorded ids, ascending.
+    fn ids(&self) -> BitIds<'_> {
+        BitIds { chunks: self.sorted_chunks(), words: &self.words, chunk: 0, word: 0, base: 0, rest: 0 }
+    }
+}
+
+struct BitIds<'a> {
+    chunks: Vec<(u64, usize)>,
+    words: &'a [Box<[u64]>],
+    chunk: usize,
+    word: usize,
+    /// Slot number of bit 0 of the word `rest` came from.
+    base: u64,
+    rest: u64,
+}
+
+impl Iterator for BitIds<'_> {
+    type Item = u64;
+
+    fn next(&mut self) -> Option<u64> {
+        loop {
+            if self.rest != 0 {
+                let bit = u64::from(self.rest.trailing_zeros());
+                self.rest &= self.rest - 1;
+                return Some((self.base + bit) << SLOT_SHIFT);
+            }
+            let &(chunk, at) = self.chunks.get(self.chunk)?;
+            self.rest = self.words[at][self.word];
+            self.base = (chunk << CHUNK_SLOT_BITS) + self.word as u64 * WORD_BITS;
+            self.word += 1;
+            if self.word == WORDS_PER_CHUNK {
+                self.word = 0;
+                self.chunk += 1;
+            }
+        }
+    }
+}
+
+pub struct IdMapBuilder {
+    bits: BitChunks,
+    /// Set once an id cannot go into `bits`; every later id goes here, and `bits` keeps what it
+    /// had. Unsorted, with duplicates.
+    plain: Option<Vec<u64>>,
+    max_ids: u64,
+}
+
+impl Default for IdMapBuilder {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl IdMapBuilder {
+    pub fn new() -> Self {
+        Self::with_limit(MAX_IDS)
+    }
+
+    /// `max_ids` is a parameter so tests can reach the too-large path without billions of inserts.
+    fn with_limit(max_ids: u64) -> Self {
+        Self { bits: BitChunks::new(), plain: None, max_ids }
+    }
+
+    /// Inserting an id twice is the same as once.
+    pub fn insert(&mut self, id: u64) {
+        match &mut self.plain {
+            Some(plain) => plain.push(id),
+            None => {
+                if id & 7 != 0 || !self.bits.insert(id) {
+                    self.plain = Some(vec![id]);
+                }
+            }
+        }
+    }
+
+    /// Fails with [`HeapError::TooLarge`] when there are more than `u32::MAX - 2` distinct ids,
+    /// and with [`HeapError::Cancelled`] when `cancel` fires (checked at least every `1 << 20`
+    /// steps).
+    pub fn build(self, store: &Store, cancel: &CancellationToken) -> Result<IdMap, HeapError> {
+        match self.plain {
+            None if self.bits.words.is_empty() => sorted_map(store, &[]),
+            None => bit_map(&self.bits, store, self.max_ids, cancel),
+            Some(plain) => merged_map(&self.bits, plain, store, self.max_ids, cancel),
+        }
+    }
+
+    #[cfg(test)]
+    fn allocated_bytes(&self) -> u64 {
+        self.bits.words.len() as u64 * CHUNK_BYTES
+            + self.plain.as_ref().map_or(0, |p| (p.capacity() * size_of::<u64>()) as u64)
+    }
+}
+
+fn check(cancel: &CancellationToken) -> Result<(), HeapError> {
+    if cancel.is_cancelled() { Err(HeapError::Cancelled) } else { Ok(()) }
+}
+
+fn sorted_map(store: &Store, ids: &[u64]) -> Result<IdMap, HeapError> {
+    let mut column = store.column(IDS_SORTED_FILE, ids.len(), 0u64)?;
+    column.copy_from_slice(ids);
+    Ok(IdMap { repr: Repr::Sorted(column) })
+}
+
+fn bit_map(bits: &BitChunks, store: &Store, max_ids: u64, cancel: &CancellationToken) -> Result<IdMap, HeapError> {
+    if bits.distinct > max_ids {
+        return Err(HeapError::TooLarge);
+    }
+    let chunks = bits.sorted_chunks();
+    let span = (bits.max - bits.min + 1) as usize;
+    let mut words = store.column(IDS_BITS_FILE, chunks.len() * WORDS_PER_CHUNK, 0u64)?;
+    let mut wrank = store.column(IDS_WRANK_FILE, chunks.len() * WORDS_PER_CHUNK, 0u32)?;
+    let mut table = store.column(IDS_TABLE_FILE, span, NONE)?;
+    let mut chunk_offsets = store.column(IDS_CHUNKS_FILE, chunks.len(), 0u32)?;
+    let mut before = 0u32;
+    for (ordinal, &(chunk, at)) in chunks.iter().enumerate() {
+        check(cancel)?;
+        let first = ordinal * WORDS_PER_CHUNK;
+        let offset = (chunk - bits.min) as usize;
+        table[offset] = first as u32;
+        chunk_offsets[ordinal] = offset as u32;
+        for (w, &word) in bits.words[at].iter().enumerate() {
+            words[first + w] = word;
+            wrank[first + w] = before;
+            before += word.count_ones();
+        }
+    }
+    Ok(IdMap { repr: Repr::Bits { min_chunk: bits.min, words, wrank, table, chunk_offsets, len: before } })
+}
+
+enum Source<'a> {
+    Bits(BitIds<'a>),
+    Run(std::iter::Copied<std::slice::Iter<'a, u64>>),
+}
+
+impl Iterator for Source<'_> {
+    type Item = u64;
+
+    fn next(&mut self) -> Option<u64> {
+        match self {
+            Source::Bits(it) => it.next(),
+            Source::Run(it) => it.next(),
+        }
+    }
+}
+
+/// Sorts `plain` in runs and merges the runs with the ids already in `bits` (itself a sorted
+/// source), dropping duplicates. A single `sort_unstable` of a very large vector could not be
+/// interrupted.
+fn merged_map(
+    bits: &BitChunks,
+    mut plain: Vec<u64>,
+    store: &Store,
+    max_ids: u64,
+    cancel: &CancellationToken,
+) -> Result<IdMap, HeapError> {
+    for run in plain.chunks_mut(CHECK_EVERY) {
+        check(cancel)?;
+        run.sort_unstable();
+    }
+    let mut sources = vec![Source::Bits(bits.ids())];
+    sources.extend(plain.chunks(CHECK_EVERY).map(|run| Source::Run(run.iter().copied())));
+    let mut heap = BinaryHeap::with_capacity(sources.len());
+    for (i, source) in sources.iter_mut().enumerate() {
+        if let Some(id) = source.next() {
+            heap.push(Reverse((id, i)));
+        }
+    }
+
+    let upper_bound = bits.distinct as usize + plain.len();
+    let mut merged = store.scratch::<u64>(upper_bound, 0)?;
+    let mut count = 0usize;
+    let mut steps = 0usize;
+    while let Some(Reverse((id, i))) = heap.pop() {
+        steps += 1;
+        if steps.is_multiple_of(CHECK_EVERY) {
+            check(cancel)?;
+        }
+        if let Some(next) = sources[i].next() {
+            heap.push(Reverse((next, i)));
+        }
+        if count > 0 && merged[count - 1] == id {
+            continue;
+        }
+        if count as u64 >= max_ids {
+            return Err(HeapError::TooLarge);
+        }
+        merged[count] = id;
+        count += 1;
+    }
+
+    let mut column = store.column(IDS_SORTED_FILE, count, 0u64)?;
+    for (to, from) in column.chunks_mut(CHECK_EVERY).zip(merged[..count].chunks(CHECK_EVERY)) {
+        check(cancel)?;
+        to.copy_from_slice(from);
+    }
+    Ok(IdMap { repr: Repr::Sorted(column) })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn insert_all(builder: &mut IdMapBuilder, ids: &[u64]) {
+        for &id in ids {
+            builder.insert(id);
+        }
+    }
+
+    #[test]
+    fn too_many_ids_is_too_large() {
+        // Dense layout, then the merged layout (one unaligned id); the duplicate does not count.
+        for unaligned in [false, true] {
+            let filled = |limit| {
+                let mut builder = IdMapBuilder::with_limit(limit);
+                insert_all(&mut builder, &[0x1000, 0x1008, 0x1010, 0x1010]);
+                if unaligned {
+                    builder.insert(0x1_0004);
+                }
+                builder
+            };
+            let distinct = if unaligned { 4 } else { 3 };
+            let map = filled(distinct).build(&Store::Ram, &CancellationToken::new()).unwrap();
+            assert_eq!(u64::from(map.len()), distinct);
+            let err = filled(distinct - 1).build(&Store::Ram, &CancellationToken::new()).err().unwrap();
+            assert!(matches!(err, HeapError::TooLarge), "{err}");
+        }
+    }
+
+    #[test]
+    fn cancelled_build_returns_cancelled() {
+        let cancelled = CancellationToken::new();
+        cancelled.cancel();
+        let mut dense = IdMapBuilder::new();
+        insert_all(&mut dense, &[0x1000, 0x1008, 0x20_0000]);
+        assert!(matches!(dense.build(&Store::Ram, &cancelled), Err(HeapError::Cancelled)));
+
+        let mut merged = IdMapBuilder::new();
+        insert_all(&mut merged, &[0x1000, 0x1004, 0x20_0000]);
+        assert!(matches!(merged.build(&Store::Ram, &cancelled), Err(HeapError::Cancelled)));
+
+        let mut ok = IdMapBuilder::new();
+        insert_all(&mut ok, &[0x1000, 0x1004]);
+        assert_eq!(ok.build(&Store::Ram, &CancellationToken::new()).unwrap().len(), 2);
+    }
+
+    #[test]
+    fn sparse_ids_do_not_amplify_builder_memory() {
+        const IDS: u64 = 2_000_000;
+        let mut builder = IdMapBuilder::new();
+        let mut peak = 0;
+        for i in 0..IDS {
+            builder.insert(0x1_0000_0000 + (i << 20));
+            if i % 1024 == 0 {
+                peak = peak.max(builder.allocated_bytes());
+            }
+        }
+        peak = peak.max(builder.allocated_bytes());
+        // Bitmaps up to the budget, plus the plain ids' vector, which may have doubled its capacity.
+        let budget = BITMAP_BASE_BYTES + (BITMAP_BYTES_PER_ID + 2 * size_of::<u64>() as u64) * IDS;
+        assert!(peak <= budget, "peak {peak} bytes, budget {budget}");
+        // One bitmap per id would have taken 16 KiB each.
+        assert!(peak < IDS * CHUNK_BYTES / 100, "peak {peak} bytes");
+        assert!(builder.plain.is_some(), "sparse input must have left the bitmaps");
+
+        let map = builder.build(&Store::Ram, &CancellationToken::new()).unwrap();
+        assert_eq!(map.len() as u64, IDS);
+        assert_eq!(map.rank(0x1_0000_0000 + (1_234_567 << 20)), Some(1_234_567));
+        assert_eq!(map.rank(0x1_0000_0000 + (1_234_567 << 20) + 8), None);
+    }
+
+    #[test]
+    fn dense_input_in_any_order_stays_in_bitmaps() {
+        // 4 GiB of address space is the base allowance; random order must not trip the budget.
+        let mut state = 0x2545_F491_4F6C_DD1Du64;
+        let mut builder = IdMapBuilder::new();
+        for _ in 0..300_000 {
+            state ^= state << 13;
+            state ^= state >> 7;
+            state ^= state << 17;
+            builder.insert((0x7_0000_0000 + (state % (4 << 30))) & !7);
+        }
+        assert!(builder.plain.is_none());
+        assert!(builder.allocated_bytes() <= BITMAP_BASE_BYTES);
+    }
+
+    #[test]
+    fn chunk_hasher_spreads_chunks_that_share_low_zero_bits() {
+        let buckets: std::collections::HashSet<u64> = (0..1024u64)
+            .map(|i| {
+                let mut hasher = ChunkHasher::default();
+                hasher.write_u64(i << 12);
+                hasher.finish() & 1023
+            })
+            .collect();
+        assert!(buckets.len() > 500, "{} buckets for 1024 keys", buckets.len());
+    }
+}
@@ -0,0 +1,109 @@
+//! The result of analyzing a heap dump: per-object columns, the class table and the summary.
+
+mod fields;
+mod path;
+mod queries;
+mod rows;
+
+use std::collections::HashMap;
+use std::path::Path;
+
+use memmap2::Mmap;
+use serde::{Deserialize, Serialize};
+
+use crate::column::Column;
+use crate::error::HeapError;
+use crate::files::SUMMARY_FILE;
+use crate::graph::ClassInfo;
+use crate::hprof::HeapRootKind;
+use crate::ids::IdMap;
+use crate::shallow::SizeSource;
+
+pub(crate) use self::fields::{id_value, instance_fields};
+pub use self::rows::{
+    HeapClassRow, HeapEdge, HeapObjectPage, HeapObjectRef, HeapObjectRow, HeapPathStep, HeapRootPath, MAX_PAGE,
+};
+
+/// Counts over instances and arrays; class objects (class dumps) are left out.
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, specta::Type)]
+#[serde(rename_all = "camelCase")]
+pub struct HeapSummary {
+    /// `reachable_objects + unreachable_objects`.
+    pub objects: u64,
+    pub reachable_objects: u64,
+    pub reachable_bytes: u64,
+    pub unreachable_objects: u64,
+    pub unreachable_bytes: u64,
+    /// Class dumps in the dump.
+    pub classes: u32,
+    pub weak_edges_skipped: u64,
+    pub size_source: SizeSource,
+    pub analysis_ms: u32,
+}
+
+/// One row of the class table, for one type slot.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct ClassStat {
+    /// [`crate::display_name`] of the class (`long[]` for a primitive array slot).
+    pub name: String,
+    /// Reachable instances.
+    pub count: u64,
+    /// Shallow bytes of the reachable instances.
+    pub shallow: u64,
+    /// Bytes the reachable instances retain together: the sum of the retained sizes of those
+    /// with no instance of the same type above them in the dominator tree.
+    pub retained: u64,
+    pub unreachable_count: u64,
+}
+
+/// Columns are indexed by node: objects `0..root`, then the super-root `root`. Columns of the
+/// objects only (`types`, `offsets`) have no entry for the root.
+pub struct HeapIndex {
+    /// The dump, for [`crate::record_at`] at `offsets[v]`.
+    pub hprof: Mmap,
+    pub ids: IdMap,
+    /// Type slot of each object, as in [`crate::HeapGraph::types`].
+    pub types: Column<u32>,
+    pub offsets: Column<u64>,
+    /// Shallow size in units of 8 bytes, rounded up.
+    pub shallow8: Column<u32>,
+    /// Retained bytes; 0 for unreachable objects, every reachable byte for the root.
+    pub retained: Column<u64>,
+    /// Immediate dominator; [`crate::NONE`] for unreachable objects, `root` for the root.
+    pub idom: Column<u32>,
+    /// Children of node `v` in the dominator tree are `kids[kids_start[v]..kids_start[v + 1]]`,
+    /// by retained size descending, then by node.
+    pub kids_start: Column<u32>,
+    pub kids: Column<u32>,
+    /// Parent on a shortest path from the root (BFS over strong edges); [`crate::NONE`] for
+    /// unreachable objects, `root` for the root.
+    pub gc_parent: Column<u32>,
+    /// By type slot: class slots, then the primitive array slots.
+    pub classes: Vec<ClassStat>,
+    /// By class slot, as in [`crate::HeapGraph::classes`].
+    pub class_info: Vec<ClassInfo>,
+    /// Type slot of class objects.
+    pub class_slot: u32,
+    /// Object → kind and thread serial of its first root record.
+    pub roots: HashMap<u32, (HeapRootKind, Option<u32>)>,
+    /// Thread serial → thread object.
+    pub thread_objects: HashMap<u32, u32>,
+    pub id_size: u8,
+    pub summary: HeapSummary,
+}
+
+impl HeapIndex {
+    /// The super-root, whose children in both trees are the rooted objects.
+    pub fn root(&self) -> u32 {
+        self.ids.len()
+    }
+}
+
+pub(crate) fn write_summary(dir: &Path, summary: &HeapSummary) -> Result<(), HeapError> {
+    let json = serde_json::to_vec_pretty(summary).map_err(std::io::Error::from)?;
+    std::fs::write(dir.join(SUMMARY_FILE), json)?;
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests;
@@ -0,0 +1,41 @@
+//! Named field values of an instance record.
+
+use super::HeapIndex;
+use crate::hprof::BasicType;
+
+pub(crate) struct FieldValue<'a> {
+    pub(crate) name: &'a str,
+    pub(crate) ty: BasicType,
+    /// Big-endian, as in the dump.
+    pub(crate) bytes: &'a [u8],
+    /// `java.lang.ref.Reference.referent`, which the graph leaves out.
+    pub(crate) weak: bool,
+}
+
+/// The fields of an instance of class slot `class` whose record holds `data`: the class's own,
+/// then each superclass's, as HotSpot lays them out. Stops where `data` ends.
+pub(crate) fn instance_fields<'a>(index: &'a HeapIndex, class: u32, data: &'a [u8]) -> Vec<FieldValue<'a>> {
+    let mut out = Vec::new();
+    let mut base = 0usize;
+    let mut cursor = Some(class);
+    // Super chains end (their cycles are cut), so this is only a bound on corrupt input.
+    for _ in 0..index.class_info.len() {
+        let Some(info) = cursor.and_then(|c| index.class_info.get(c as usize)) else { break };
+        for (k, (name, ty)) in info.fields.iter().enumerate() {
+            let size = ty.hprof_size(index.id_size);
+            let Some(bytes) = base.checked_add(size).and_then(|end| data.get(base..end)) else { return out };
+            out.push(FieldValue { name, ty: *ty, bytes, weak: info.weak_field == Some(k) });
+            base += size;
+        }
+        cursor = info.super_index;
+    }
+    out
+}
+
+/// A 4- or 8-byte big-endian id.
+pub(crate) fn id_value(bytes: &[u8]) -> Option<u64> {
+    match *bytes {
+        [a, b, c, d] => Some(u64::from(u32::from_be_bytes([a, b, c, d]))),
+        _ => Some(u64::from_be_bytes(bytes.try_into().ok()?)),
+    }
+}
@@ -0,0 +1,105 @@
+//! Paths to GC roots, along the BFS tree of `gc_parent`.
+
+use super::{HeapEdge, HeapIndex, HeapObjectRef, HeapPathStep, HeapRootPath, id_value, instance_fields};
+use crate::column::NONE;
+use crate::hprof::{BasicType, HeapRecord, HeapRootKind};
+use crate::strings::{char_array_text, string_preview};
+
+const THREAD_NAME_CHARS: usize = 100;
+
+/// Steps a long path keeps at each end: one deep in a long linked list would otherwise send
+/// every node over IPC.
+const PATH_END_STEPS: u32 = 500;
+
+impl HeapIndex {
+    /// From `object` up to its GC root, each step with how the next one holds it; past 1000
+    /// steps, only the 500 at each end. `None` for an unreachable object or a ref that names no
+    /// object.
+    pub fn path_to_root(&self, object: HeapObjectRef) -> Option<HeapRootPath> {
+        let root = self.root();
+        let mut v = object.0;
+        if v >= root || self.gc_parent[v as usize] == NONE {
+            return None;
+        }
+        let len = self.path_len(v)?;
+        let end = PATH_END_STEPS;
+        let cut = len > 2 * end;
+        let mut steps = Vec::with_capacity(len.min(2 * end) as usize);
+        for i in 0..len {
+            let parent = self.gc_parent[v as usize];
+            if !cut || i < end || i >= len - end {
+                let via = if parent == root { self.root_edge(v) } else { self.edge(parent, v) };
+                steps.push(HeapPathStep { object: self.object_row(HeapObjectRef(v))?, via });
+            }
+            v = parent;
+        }
+        let omitted = if cut { len - 2 * end } else { 0 };
+        Some(HeapRootPath { steps, omitted, omitted_after: if cut { end } else { 0 } })
+    }
+
+    /// Steps from `v` to the root; `None` when `gc_parent` leads nowhere.
+    fn path_len(&self, mut v: u32) -> Option<u32> {
+        let root = self.root();
+        // A BFS tree has no cycles; the bound only stops a corrupt column from looping.
+        for len in 1..=root {
+            let parent = self.gc_parent[v as usize];
+            if parent == root {
+                return Some(len);
+            }
+            if parent > root {
+                return None;
+            }
+            v = parent;
+        }
+        None
+    }
+
+    fn root_edge(&self, v: u32) -> HeapEdge {
+        let (kind, serial) = self.roots.get(&v).copied().unwrap_or((HeapRootKind::Unknown, None));
+        let thread = serial.and_then(|s| self.thread_objects.get(&s)).and_then(|&t| self.thread_name(t));
+        HeapEdge::Root { root: kind, thread }
+    }
+
+    /// `Thread.name`: a String, or a `char[]` before JDK 9.
+    fn thread_name(&self, thread: u32) -> Option<String> {
+        let Some(HeapRecord::Instance { data, .. }) = self.record(thread) else { return None };
+        let fields = instance_fields(self, *self.types.get(thread as usize)?, data);
+        let name = fields.iter().find(|f| f.name == "name" && f.ty == BasicType::Object)?;
+        let target = self.ids.rank(id_value(name.bytes)?)?;
+        string_preview(self, target, THREAD_NAME_CHARS).or_else(|| char_array_text(self, target, THREAD_NAME_CHARS))
+    }
+
+    /// How `parent` refers to `child`, checked in the order the graph's edges were made. A strong
+    /// edge from `parent` always exists; the empty field name only covers a dump that changed.
+    fn edge(&self, parent: u32, child: u32) -> HeapEdge {
+        self.find_edge(parent, self.ids.id_at(child)).unwrap_or(HeapEdge::Field { name: String::new() })
+    }
+
+    fn find_edge(&self, parent: u32, child_id: u64) -> Option<HeapEdge> {
+        match self.record(parent)? {
+            HeapRecord::Instance { data, .. } => {
+                let class = self.types[parent as usize];
+                let field = instance_fields(self, class, data)
+                    .into_iter()
+                    .find(|f| !f.weak && f.ty == BasicType::Object && id_value(f.bytes) == Some(child_id));
+                match field {
+                    Some(f) => Some(HeapEdge::Field { name: f.name.to_owned() }),
+                    None => (self.class_info.get(class as usize)?.id == child_id).then_some(HeapEdge::ClassOf),
+                }
+            }
+            HeapRecord::ObjectArray { data, .. } => {
+                let index =
+                    data.chunks_exact(usize::from(self.id_size)).position(|item| id_value(item) == Some(child_id))?;
+                Some(HeapEdge::Element { index: u32::try_from(index).ok()? })
+            }
+            HeapRecord::Class(dump) if dump.super_id == child_id => Some(HeapEdge::SuperClass),
+            HeapRecord::Class(dump) if dump.loader_id == child_id => Some(HeapEdge::ClassLoader),
+            HeapRecord::Class(_) => {
+                let info = self.class_info.get(self.class_of_object(parent)?)?;
+                let (name, _) = info.statics.iter().find(|(_, object)| *object == child_id)?;
+                Some(HeapEdge::Static { name: name.clone() })
+            }
+            HeapRecord::Root { .. } | HeapRecord::PrimitiveArray { .. } => None,
+        }
+    }
+}
@@ -0,0 +1,164 @@
+//! Class table rows, pages of the dominator tree and single object rows.
+
+use std::cmp::Reverse;
+use std::collections::{BinaryHeap, HashMap};
+
+use super::{HeapClassRow, HeapIndex, HeapObjectPage, HeapObjectRef, HeapObjectRow, HeapSummary, MAX_PAGE};
+use crate::column::NONE;
+use crate::hprof::HeapRecord;
+use crate::strings::string_preview;
+
+const PREVIEW_CHARS: usize = 100;
+
+impl HeapIndex {
+    pub fn summary(&self) -> &HeapSummary {
+        &self.summary
+    }
+
+    /// Types that have objects, by retained size descending, then by name.
+    pub fn class_rows(&self) -> Vec<HeapClassRow> {
+        let class_slots = &self.classes[..self.class_info.len().min(self.classes.len())];
+        let mut names: HashMap<&str, u32> = HashMap::with_capacity(class_slots.len());
+        for class in class_slots {
+            *names.entry(class.name.as_str()).or_default() += 1;
+        }
+        let mut rows: Vec<HeapClassRow> = self
+            .classes
+            .iter()
+            .enumerate()
+            .filter(|(_, c)| c.count > 0 || c.unreachable_count > 0)
+            .map(|(slot, c)| HeapClassRow {
+                class_index: slot as u32,
+                name: c.name.clone(),
+                count: c.count,
+                shallow: c.shallow,
+                retained: c.retained,
+                unreachable_count: c.unreachable_count,
+                // Primitive array slots have no class dump and so no second loader.
+                ambiguous: slot < class_slots.len() && names[c.name.as_str()] > 1,
+            })
+            .collect();
+        rows.sort_by(|a, b| b.retained.cmp(&a.retained).then_with(|| a.name.cmp(&b.name)));
+        rows
+    }
+
+    /// A page of objects by retained size descending, then by index:
+    ///
+    /// - no `parent`, no `class_index`: the children of the super-root, class objects left out;
+    /// - `parent`: its children in the dominator tree, of type `class_index` if given;
+    /// - `class_index` only: every reachable object of that type.
+    ///
+    /// `limit` is clamped to [`MAX_PAGE`]. A `parent` or `class_index` that names nothing gives an
+    /// empty page.
+    pub fn dominated(
+        &self,
+        parent: Option<HeapObjectRef>,
+        class_index: Option<u32>,
+        offset: u32,
+        limit: u32,
+    ) -> HeapObjectPage {
+        let limit = limit.min(MAX_PAGE);
+        let root = self.root();
+        if class_index.is_some_and(|c| c as usize >= self.classes.len()) {
+            return HeapObjectPage { rows: Vec::new(), total: 0 };
+        }
+        match (parent, class_index) {
+            (Some(HeapObjectRef(p)), _) if p >= root => HeapObjectPage { rows: Vec::new(), total: 0 },
+            (Some(HeapObjectRef(p)), class) => self.page(
+                self.children(p).filter(|&v| class.is_none_or(|c| self.types[v as usize] == c)),
+                offset,
+                limit,
+            ),
+            (None, None) => {
+                self.page(self.children(root).filter(|&v| self.types[v as usize] != self.class_slot), offset, limit)
+            }
+            (None, Some(class)) => self.largest_of(class, offset, limit),
+        }
+    }
+
+    /// `None` for a ref that names no object, the super-root included.
+    pub fn object_row(&self, object: HeapObjectRef) -> Option<HeapObjectRow> {
+        let v = object.0;
+        if v >= self.root() {
+            return None;
+        }
+        let i = v as usize;
+        let slot = self.types[i];
+        let preview = if slot == self.class_slot {
+            self.class_of_object(v).and_then(|c| self.classes.get(c)).map(|c| c.name.clone())
+        } else {
+            string_preview(self, v, PREVIEW_CHARS)
+        };
+        Some(HeapObjectRow {
+            object,
+            class_name: self.classes.get(slot as usize).map(|c| c.name.clone()).unwrap_or_default(),
+            address: format!("{:#x}", self.ids.id_at(v)),
+            shallow: u64::from(self.shallow8[i]) * 8,
+            retained: self.retained[i],
+            has_children: self.kids_start[i + 1] > self.kids_start[i],
+            preview,
+        })
+    }
+
+    /// The record of object `v`, or `None` when it cannot be read.
+    pub(crate) fn record(&self, v: u32) -> Option<HeapRecord<'_>> {
+        let at = usize::try_from(*self.offsets.get(v as usize)?).ok()?;
+        crate::hprof::record_at(&self.hprof, at, self.id_size).ok()
+    }
+
+    /// The class slot a class object stands for.
+    pub(crate) fn class_of_object(&self, v: u32) -> Option<usize> {
+        if v >= self.root() {
+            return None;
+        }
+        let id = self.ids.id_at(v);
+        self.class_info.iter().position(|c| c.id == id && c.id != 0)
+    }
+
+    /// Children of node `v` in the dominator tree, already in page order.
+    fn children(&self, v: u32) -> impl Iterator<Item = u32> + '_ {
+        let v = v as usize;
+        self.kids[self.kids_start[v] as usize..self.kids_start[v + 1] as usize].iter().copied()
+    }
+
+    fn page(&self, objects: impl Iterator<Item = u32>, offset: u32, limit: u32) -> HeapObjectPage {
+        let mut rows = Vec::new();
+        let mut total = 0u32;
+        for v in objects {
+            if total >= offset && total - offset < limit {
+                rows.extend(self.object_row(HeapObjectRef(v)));
+            }
+            total = total.saturating_add(1);
+        }
+        HeapObjectPage { rows, total }
+    }
+
+    /// Scans every object: the top `offset + limit` stay in a heap whose top is the smallest.
+    fn largest_of(&self, class: u32, offset: u32, limit: u32) -> HeapObjectPage {
+        let count = self.classes[class as usize].count;
+        let total = u32::try_from(count).unwrap_or(u32::MAX);
+        let keep = match offset.checked_add(limit) {
+            Some(keep) if offset < total && limit > 0 => keep.min(total) as usize,
+            _ => return HeapObjectPage { rows: Vec::new(), total },
+        };
+        let mut top: BinaryHeap<Reverse<(u64, Reverse<u32>)>> = BinaryHeap::with_capacity(keep + 1);
+        for v in 0..self.root() {
+            let i = v as usize;
+            if self.types[i] != class || self.idom[i] == NONE {
+                continue;
+            }
+            top.push(Reverse((self.retained[i], Reverse(v))));
+            if top.len() > keep {
+                top.pop();
+            }
+        }
+        // Ascending `Reverse` order is descending retained, then ascending index.
+        let rows = top
+            .into_sorted_vec()
+            .into_iter()
+            .skip(offset as usize)
+            .filter_map(|Reverse((_, Reverse(v)))| self.object_row(HeapObjectRef(v)))
+            .collect();
+        HeapObjectPage { rows, total }
+    }
+}
@@ -0,0 +1,95 @@
+//! What the index queries return over IPC.
+
+use serde::{Deserialize, Serialize};
+
+use crate::hprof::HeapRootKind;
+
+/// Most rows one page of [`crate::HeapIndex::dominated`] returns.
+pub const MAX_PAGE: u32 = 500;
+
+/// An object of one snapshot: its index in the [`crate::HeapIndex`] columns. Any value may
+/// arrive over IPC; queries answer an index that names no object with nothing.
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash, Serialize, Deserialize, specta::Type)]
+#[serde(transparent)]
+pub struct HeapObjectRef(pub u32);
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, specta::Type)]
+#[serde(rename_all = "camelCase")]
+pub struct HeapClassRow {
+    /// Type slot, as `classIndex` in [`crate::HeapIndex::dominated`].
+    pub class_index: u32,
+    pub name: String,
+    pub count: u64,
+    pub shallow: u64,
+    pub retained: u64,
+    pub unreachable_count: u64,
+    /// Another class has the same name (several class loaders): nothing joined by name, such as a
+    /// second snapshot, can tell them apart.
+    pub ambiguous: bool,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, specta::Type)]
+#[serde(rename_all = "camelCase")]
+pub struct HeapObjectRow {
+    pub object: HeapObjectRef,
+    pub class_name: String,
+    /// The object's hprof id in hex, `0x1300`.
+    pub address: String,
+    pub shallow: u64,
+    pub retained: u64,
+    /// Has children in the dominator tree.
+    pub has_children: bool,
+    /// The text of a `java.lang.String`, or the name of the class a class object stands for.
+    pub preview: Option<String>,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, specta::Type)]
+#[serde(rename_all = "camelCase")]
+pub struct HeapObjectPage {
+    pub rows: Vec<HeapObjectRow>,
+    /// Rows of every page together.
+    pub total: u32,
+}
+
+/// How an object is held by the next step of a path to a GC root.
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, specta::Type)]
+#[serde(tag = "kind", rename_all = "camelCase", rename_all_fields = "camelCase")]
+pub enum HeapEdge {
+    /// An instance field; the name is empty when the dump does not show which one.
+    Field {
+        name: String,
+    },
+    Element {
+        index: u32,
+    },
+    Static {
+        name: String,
+    },
+    /// From an instance to its class.
+    ClassOf,
+    SuperClass,
+    ClassLoader,
+    /// `thread` is the name of the thread whose root it is, when the dump has it.
+    Root {
+        root: HeapRootKind,
+        thread: Option<String>,
+    },
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, specta::Type)]
+#[serde(rename_all = "camelCase")]
+pub struct HeapPathStep {
+    pub object: HeapObjectRow,
+    /// How the next step holds this object; on the last step, the root.
+    pub via: HeapEdge,
+}
+
+/// A path from an object up to its GC root. A long one keeps only the steps nearest each end.
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, specta::Type)]
+#[serde(rename_all = "camelCase")]
+pub struct HeapRootPath {
+    pub steps: Vec<HeapPathStep>,
+    /// Steps left out between `steps[omitted_after - 1]` and `steps[omitted_after]`; 0 for none.
+    pub omitted: u32,
+    pub omitted_after: u32,
+}
@@ -0,0 +1,336 @@
+use super::*;
+use crate::hprof::BasicType;
+use crate::test_index::{analyzed, analyzed_bytes, obj, slot};
+use crate::testing::{HprofWriter, MiniHeap, sample, sample_with_string};
+
+const H: u64 = 0x1000;
+const N1: u64 = 0x1100;
+const N3: u64 = 0x1300;
+const L: u64 = 0x1400;
+const W: u64 = 0x1600;
+const N4: u64 = 0x1700;
+
+fn addresses(page: &HeapObjectPage) -> Vec<&str> {
+    page.rows.iter().map(|r| r.address.as_str()).collect()
+}
+
+fn retained(page: &HeapObjectPage) -> Vec<u64> {
+    page.rows.iter().map(|r| r.retained).collect()
+}
+
+fn path(index: &HeapIndex, addr: u64) -> Option<Vec<(String, HeapEdge)>> {
+    let path = index.path_to_root(obj(index, addr))?;
+    assert_eq!((path.omitted, path.omitted_after), (0, 0));
+    Some(path.steps.into_iter().map(|s| (s.object.address, s.via)).collect())
+}
+
+fn field(name: &str) -> HeapEdge {
+    HeapEdge::Field { name: name.to_owned() }
+}
+
+fn root(root: HeapRootKind, thread: Option<&str>) -> HeapEdge {
+    HeapEdge::Root { root, thread: thread.map(str::to_owned) }
+}
+
+#[test]
+fn top_level_dominators_sorted_by_retained() {
+    let index = analyzed(&sample());
+    let page = index.dominated(None, None, 0, 10);
+    assert_eq!(addresses(&page), ["0x1000", "0x1300", "0x1600"]);
+    assert_eq!(retained(&page), [72, 56, 24]);
+    // The class objects, rooted as sticky classes, are left out.
+    assert_eq!(page.total, 3);
+    assert_eq!(addresses(&index.dominated(None, None, 1, 1)), ["0x1300"]);
+}
+
+#[test]
+fn children_of_holder() {
+    let index = analyzed(&sample());
+    let page = index.dominated(Some(obj(&index, H)), None, 0, 10);
+    assert_eq!(addresses(&page), ["0x1100", "0x1200"]);
+    assert_eq!(retained(&page), [24, 24]);
+    assert_eq!(page.total, 2);
+    assert!(page.rows.iter().all(|r| !r.has_children && r.class_name == "Node"));
+    let n3 = index.object_row(obj(&index, N3)).unwrap();
+    assert!(n3.has_children);
+    assert_eq!(index.dominated(Some(n3.object), None, 0, 10).rows[0].class_name, "long[]");
+}
+
+#[test]
+fn largest_of_class() {
+    let index = analyzed(&sample());
+    let node = slot(&index, "Node");
+    let page = index.dominated(None, Some(node), 0, 2);
+    assert_eq!(addresses(&page), ["0x1300", "0x1100"]);
+    assert_eq!(retained(&page), [56, 24]);
+    assert_eq!(page.total, 3);
+    assert_eq!(addresses(&index.dominated(None, Some(node), 2, 2)), ["0x1200"]);
+    assert!(index.dominated(None, Some(node), 3, 2).rows.is_empty());
+    // Within a parent's children.
+    assert_eq!(addresses(&index.dominated(Some(obj(&index, H)), Some(node), 1, 5)), ["0x1200"]);
+    assert_eq!(index.dominated(Some(obj(&index, H)), Some(slot(&index, "Holder")), 0, 5).total, 0);
+}
+
+#[test]
+fn object_row_describes_object() {
+    let index = analyzed(&sample());
+    let row = index.object_row(obj(&index, N3)).unwrap();
+    assert_eq!(
+        row,
+        HeapObjectRow {
+            object: obj(&index, N3),
+            class_name: "Node".to_owned(),
+            address: "0x1300".to_owned(),
+            shallow: 24,
+            retained: 56,
+            has_children: true,
+            preview: None,
+        }
+    );
+    let unreachable = index.object_row(obj(&index, N4)).unwrap();
+    assert_eq!((unreachable.retained, unreachable.has_children), (0, false));
+
+    let holder = &index.class_info[slot(&index, "Holder") as usize];
+    let class_object = obj(&index, holder.id);
+    let row = index.object_row(class_object).unwrap();
+    assert_eq!((row.class_name.as_str(), row.preview.as_deref()), ("java.lang.Class", Some("Holder")));
+}
+
+#[test]
+fn refs_that_name_no_object_answer_nothing() {
+    let index = analyzed(&sample());
+    let node = slot(&index, "Node");
+    for bad in [index.root(), index.root() + 1, u32::MAX] {
+        let bad = HeapObjectRef(bad);
+        assert_eq!(index.object_row(bad), None);
+        assert_eq!(index.path_to_root(bad), None);
+        assert_eq!(index.dominated(Some(bad), None, 0, 10), HeapObjectPage { rows: Vec::new(), total: 0 });
+        assert_eq!(index.dominated(Some(bad), Some(node), 0, 10).total, 0);
+    }
+    for bad in [index.classes.len() as u32, u32::MAX] {
+        assert_eq!(index.dominated(None, Some(bad), 0, 10).total, 0);
+    }
+}
+
+#[test]
+fn path_to_root_names_fields_and_root() {
+    let index = analyzed(&sample());
+    assert_eq!(
+        path(&index, L),
+        Some(vec![("0x1400".to_owned(), field("next")), ("0x1300".to_owned(), root(HeapRootKind::JavaFrame, None))])
+    );
+    assert_eq!(
+        path(&index, N1),
+        Some(vec![("0x1100".to_owned(), field("a")), ("0x1000".to_owned(), root(HeapRootKind::JniGlobal, None))])
+    );
+    assert_eq!(path(&index, W), Some(vec![("0x1600".to_owned(), root(HeapRootKind::JniGlobal, None))]));
+    assert_eq!(path(&index, N4), None);
+}
+
+/// A rooted linked list of `len` Nodes from `0x100000`, each held by the next one's `next`.
+fn chain(len: u64) -> HeapIndex {
+    let mut heap = sample();
+    for i in 0..len {
+        let next = if i == 0 { 0 } else { 0x10_0000 + (i - 1) * 0x20 };
+        heap.object(0x10_0000 + i * 0x20, "Node", &[("next", next)]);
+    }
+    heap.root(HeapRootKind::JniGlobal, 0x10_0000 + (len - 1) * 0x20);
+    analyzed(&heap)
+}
+
+#[test]
+fn path_of_an_object_deep_in_a_linked_list_keeps_both_ends() {
+    let index = chain(3000);
+    let path = index.path_to_root(obj(&index, 0x10_0000)).unwrap();
+    assert_eq!((path.steps.len(), path.omitted, path.omitted_after), (1000, 2000, 500));
+    let address = |i: usize| path.steps[i].object.address.clone();
+    assert_eq!([address(0), address(499)], ["0x100000", "0x103e60"]);
+    // The step after the gap is the one 2000 nodes further up.
+    assert_eq!([address(500), address(999)], ["0x113880", "0x1176e0"]);
+    assert_eq!(path.steps[999].via, root(HeapRootKind::JniGlobal, None));
+    assert!(path.steps[..999].iter().all(|s| s.via == field("next")));
+}
+
+#[test]
+fn path_of_exactly_a_thousand_steps_is_whole() {
+    let index = chain(1000);
+    let path = index.path_to_root(obj(&index, 0x10_0000)).unwrap();
+    assert_eq!((path.steps.len(), path.omitted, path.omitted_after), (1000, 0, 0));
+}
+
+#[test]
+fn path_skips_the_weak_referent() {
+    let mut heap = sample();
+    heap.object(0x1c00, "java.lang.ref.WeakReference", &[("referent", 0x1d00), ("queue", 0x1d00)])
+        .object(0x1d00, "Node", &[])
+        .root(HeapRootKind::JniGlobal, 0x1c00);
+    let index = analyzed(&heap);
+    assert_eq!(path(&index, 0x1d00).unwrap()[0].1, field("queue"));
+}
+
+/// Kid extends Base, loaded by loader `0x300`; Base's static `cache` holds X; K is a rooted Kid;
+/// `0x700` is a rooted `Object[] { null, Y }`. No class is rooted on its own.
+fn class_edges_heap() -> Vec<u8> {
+    let mut w = HprofWriter::new();
+    let names = [(1, "Base"), (2, "Kid"), (3, "Ldr"), (4, "[Ljava/lang/Object;"), (5, "cache")];
+    for (id, name) in names {
+        w.string(id, name);
+    }
+    for (class, name) in [(0x100, 1), (0x200, 2), (0x400, 3), (0x800, 4)] {
+        w.load_class(class, name);
+    }
+    w.class(0x100, 0, &[], &[(5, 0x500)])
+        .class_with_loader(0x200, 0x100, 0x300, &[], &[])
+        .class(0x400, 0, &[], &[])
+        .class(0x800, 0, &[], &[])
+        .instance(0x300, 0x400, &[])
+        .instance(0x500, 0x400, &[])
+        .instance(0x600, 0x200, &[])
+        .obj_array(0x700, 0x800, &[0, 0x900])
+        .instance(0x900, 0x400, &[])
+        .root(HeapRootKind::JniGlobal, 0x600)
+        .root(HeapRootKind::JniGlobal, 0x700);
+    w.finish()
+}
+
+#[test]
+fn path_names_statics_supers_loaders_classes_and_elements() {
+    let index = analyzed_bytes(&class_edges_heap());
+    let k = ("0x600".to_owned(), root(HeapRootKind::JniGlobal, None));
+    assert_eq!(
+        path(&index, 0x500),
+        Some(vec![
+            ("0x500".to_owned(), HeapEdge::Static { name: "cache".to_owned() }),
+            ("0x100".to_owned(), HeapEdge::SuperClass),
+            ("0x200".to_owned(), HeapEdge::ClassOf),
+            k.clone(),
+        ])
+    );
+    assert_eq!(
+        path(&index, 0x300),
+        Some(vec![("0x300".to_owned(), HeapEdge::ClassLoader), ("0x200".to_owned(), HeapEdge::ClassOf), k])
+    );
+    assert_eq!(
+        path(&index, 0x900),
+        Some(vec![
+            ("0x900".to_owned(), HeapEdge::Element { index: 1 }),
+            ("0x700".to_owned(), root(HeapRootKind::JniGlobal, None))
+        ])
+    );
+}
+
+fn with_main_thread(mut heap: MiniHeap, name: impl FnOnce(&mut MiniHeap, u64)) -> MiniHeap {
+    heap.class("java.lang.Thread", Some("java.lang.Object"), &[("name", BasicType::Object)]);
+    name(&mut heap, 0x2000);
+    heap.object(0x2100, "java.lang.Thread", &[("name", 0x2000)]).root_in_thread(HeapRootKind::ThreadObject, 0x2100, 1);
+    heap
+}
+
+#[test]
+fn thread_name_on_java_frame_root() {
+    let heap = with_main_thread(sample_with_string(), |h, at| {
+        h.string(at, "main", true);
+    });
+    let index = analyzed(&heap);
+    let steps = path(&index, L).unwrap();
+    assert_eq!(steps.last().unwrap().1, root(HeapRootKind::JavaFrame, Some("main")));
+    // A JNI global carries no thread.
+    assert_eq!(path(&index, H).unwrap()[0].1, root(HeapRootKind::JniGlobal, None));
+}
+
+#[test]
+fn jdk8_thread_name_is_a_char_array() {
+    let heap = with_main_thread(sample(), |h, at| {
+        let utf16: Vec<u8> = "main".encode_utf16().flat_map(u16::to_be_bytes).collect();
+        h.prim_bytes(at, BasicType::Char, &utf16);
+    });
+    let index = analyzed(&heap);
+    assert_eq!(path(&index, N3).unwrap()[0].1, root(HeapRootKind::JavaFrame, Some("main")));
+}
+
+#[test]
+fn page_limit_is_clamped() {
+    let mut heap = MiniHeap::new();
+    heap.class("Node", Some("java.lang.Object"), &[("next", BasicType::Object)]);
+    for k in 0..600u64 {
+        let addr = 0x10_0000 + k * 0x100;
+        heap.object(addr, "Node", &[]).root(HeapRootKind::JniGlobal, addr);
+    }
+    let index = analyzed(&heap);
+    let node = slot(&index, "Node");
+    let page = index.dominated(None, Some(node), 0, 10_000);
+    assert_eq!((page.rows.len(), page.total), (MAX_PAGE as usize, 600));
+    let top = index.dominated(None, None, 0, 10_000);
+    assert_eq!((top.rows.len(), top.total), (MAX_PAGE as usize, 600));
+    assert_eq!(index.dominated(None, None, 550, 10_000).rows.len(), 50);
+    for (offset, limit) in [(u32::MAX, u32::MAX), (u32::MAX - 10, 100), (599, u32::MAX)] {
+        let page = index.dominated(None, Some(node), offset, limit);
+        assert_eq!((page.rows.len(), page.total), (usize::from(offset == 599), 600), "{offset} {limit}");
+        assert!(index.dominated(None, None, offset, limit).rows.len() <= 1);
+    }
+}
+
+#[test]
+fn class_rows_skip_empty_slots() {
+    let index = analyzed(&sample());
+    let rows = index.class_rows();
+    let node = rows.iter().find(|r| r.name == "Node").unwrap();
+    assert_eq!(
+        (node.class_index, node.count, node.shallow, node.retained, node.unreachable_count, node.ambiguous),
+        (slot(&index, "Node"), 3, 72, 104, 1, false)
+    );
+    assert!(rows.iter().all(|r| r.count + r.unreachable_count > 0));
+    assert!(!rows.iter().any(|r| r.name == "boolean[]"));
+    assert!(rows.windows(2).all(|p| p[0].retained >= p[1].retained));
+    assert_eq!(index.summary().objects, 8);
+}
+
+#[test]
+fn class_rows_mark_names_of_several_classes() {
+    let mut heap = sample();
+    // `object` takes the class defined last under a name, so each Foo gets one instance.
+    heap.class("Foo", Some("java.lang.Object"), &[])
+        .object(0x3000, "Foo", &[])
+        .class("Foo", Some("java.lang.Object"), &[])
+        .object(0x3100, "Foo", &[])
+        .object(0x3200, "Foo", &[]);
+    let index = analyzed(&heap);
+    let rows = index.class_rows();
+    let foos: Vec<_> = rows.iter().filter(|r| r.name == "Foo").collect();
+    assert_eq!(foos.len(), 2);
+    assert!(foos.iter().all(|r| r.ambiguous));
+    let mut counts: Vec<u64> = foos.iter().map(|r| r.unreachable_count).collect();
+    counts.sort_unstable();
+    assert_eq!(counts, [1, 2]);
+    assert!(rows.iter().filter(|r| r.name != "Foo").all(|r| !r.ambiguous));
+}
+
+#[test]
+fn ipc_shapes() {
+    assert_eq!(json(&HeapObjectRef(7)), "7");
+    assert_eq!(json(&HeapEdge::ClassOf), r#"{"kind":"classOf"}"#);
+    assert_eq!(json(&HeapEdge::Element { index: 2 }), r#"{"kind":"element","index":2}"#);
+    assert_eq!(
+        json(&root(HeapRootKind::JavaFrame, Some("main"))),
+        r#"{"kind":"root","root":"javaFrame","thread":"main"}"#
+    );
+    let row = HeapClassRow {
+        class_index: 1,
+        name: "A".to_owned(),
+        count: 2,
+        shallow: 3,
+        retained: 4,
+        unreachable_count: 5,
+        ambiguous: false,
+    };
+    assert_eq!(
+        json(&row),
+        r#"{"classIndex":1,"name":"A","count":2,"shallow":3,"retained":4,"unreachableCount":5,"ambiguous":false}"#
+    );
+    let back: HeapObjectRef = serde_json::from_str("4294967295").unwrap();
+    assert_eq!(back, HeapObjectRef(u32::MAX));
+}
+
+fn json(value: &impl serde::Serialize) -> String {
+    serde_json::to_string(value).unwrap()
+}
@@ -0,0 +1,44 @@
+//! HotSpot `.hprof` heap-dump reading and analysis for the debugger's JVM Memory view.
+
+mod analysis;
+mod column;
+mod compare;
+mod disk;
+mod dominators;
+mod error;
+mod files;
+mod graph;
+mod hprof;
+mod ids;
+mod index;
+mod names;
+mod progress;
+mod reader;
+mod registry;
+mod shallow;
+mod strings;
+#[cfg(test)]
+mod test_index;
+/// Fixture builders shared by this crate's tests and by dependants' tests and benches; not API.
+#[doc(hidden)]
+pub mod testing;
+
+pub use analysis::{AnalyzeOptions, analyze};
+pub use column::{Column, NONE, Pod, Store};
+pub use compare::{HeapClassDiff, compare};
+pub use disk::{DISK_MARGIN, capture_need, create_snapshot_dir, free_bytes, index_need, remove_stale, snapshots_root};
+pub use dominators::{Dominators, dominators};
+pub use error::HeapError;
+pub use graph::{ClassInfo, HeapGraph, PRIM_SLOTS, build_graph};
+pub use hprof::{
+    BasicType, ClassDump, HeapRecord, HeapRootKind, HprofHeader, Record, parse_header, record_at, scan, scan_segment,
+};
+pub use ids::{IdMap, IdMapBuilder};
+pub use index::{
+    ClassStat, HeapClassRow, HeapEdge, HeapIndex, HeapObjectPage, HeapObjectRef, HeapObjectRow, HeapPathStep,
+    HeapRootPath, HeapSummary, MAX_PAGE,
+};
+pub use names::display_name;
+pub use progress::HeapPhase;
+pub use registry::{HeapOrigin, HeapRegistry, HeapSnapshotId, HeapSnapshotInfo, OpenSnapshot};
+pub use shallow::{HistogramEntry, ShallowModel, SizeSource, VmLayout};
@@ -0,0 +1,66 @@
+//! Class names in the form the JVM's class histogram and the UI use.
+
+/// The Java source form of an hprof class name (`java/util/HashMap$Node`, `[Ljava/lang/String;`),
+/// as `ide_jvm_attach::display_class_name` writes it, plus `+0x` → `/0x` for hidden classes,
+/// which the histogram spells with a slash. Names that already use dots pass through unchanged.
+pub fn display_name(hprof_name: &str) -> String {
+    let dotted = hprof_name.replace('/', ".");
+    let dims = dotted.bytes().take_while(|&b| b == b'[').count();
+    let shown = if dims == 0 { dotted } else { array_name(&dotted, dims) };
+    // After the slash replacement, or the new slash would become a dot.
+    shown.replace("+0x", "/0x")
+}
+
+/// `[[I` → `int[][]`; a descriptor that is not one stays as it is.
+fn array_name(dotted: &str, dims: usize) -> String {
+    let element = &dotted[dims..];
+    let base = match element {
+        "Z" => "boolean",
+        "B" => "byte",
+        "C" => "char",
+        "S" => "short",
+        "I" => "int",
+        "J" => "long",
+        "F" => "float",
+        "D" => "double",
+        _ => match element.strip_prefix('L').and_then(|e| e.strip_suffix(';')) {
+            Some(class) => class,
+            None => return dotted.to_owned(),
+        },
+    };
+    format!("{base}{}", "[]".repeat(dims))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn display_names() {
+        for (hprof, shown) in [
+            ("[Z", "boolean[]"),
+            ("[J", "long[]"),
+            ("[[D", "double[][]"),
+            ("[Ljava/lang/String;", "java.lang.String[]"),
+            ("java/util/HashMap$Node", "java.util.HashMap$Node"),
+            ("Gen$$Lambda+0x0000007800000a00", "Gen$$Lambda/0x0000007800000a00"),
+        ] {
+            assert_eq!(display_name(hprof), shown, "{hprof}");
+        }
+    }
+
+    #[test]
+    fn dotted_names_and_hidden_class_arrays() {
+        for (hprof, shown) in [
+            ("java.lang.String", "java.lang.String"),
+            ("[Ljava.lang.Object;", "java.lang.Object[]"),
+            ("[[Lpkg.Gen$$Lambda+0x0000007800000a00;", "pkg.Gen$$Lambda/0x0000007800000a00[][]"),
+            ("[Ljava.lang.String", "[Ljava.lang.String"),
+            ("[Q", "[Q"),
+            ("0x7f00", "0x7f00"),
+            ("", ""),
+        ] {
+            assert_eq!(display_name(hprof), shown, "{hprof}");
+        }
+    }
+}
@@ -0,0 +1,103 @@
+//! Phases of a heap analysis and the progress/cancellation checks every long pass makes.
+
+use ide_core::CancellationToken;
+use serde::{Deserialize, Serialize};
+
+use crate::error::HeapError;
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize, specta::Type)]
+#[serde(rename_all = "camelCase")]
+pub enum HeapPhase {
+    Reading,
+    Linking,
+    Dominators,
+    Retained,
+    Paths,
+    Saving,
+}
+
+/// Longest stretch of work between two progress reports (and looks at the cancellation token).
+pub(crate) const CHECK_EVERY: u64 = 1 << 20;
+
+/// The caller's progress callback plus the token, looked at right after every report so a callback
+/// that cancels stops the work before the next step.
+pub(crate) struct Progress<'a> {
+    report: &'a dyn Fn(HeapPhase, u64, u64) -> Result<(), HeapError>,
+    cancel: &'a CancellationToken,
+}
+
+impl<'a> Progress<'a> {
+    pub(crate) fn new(
+        report: &'a dyn Fn(HeapPhase, u64, u64) -> Result<(), HeapError>,
+        cancel: &'a CancellationToken,
+    ) -> Self {
+        Self { report, cancel }
+    }
+
+    pub(crate) fn report(&self, phase: HeapPhase, done: u64, total: u64) -> Result<(), HeapError> {
+        (self.report)(phase, done, total)?;
+        self.check()
+    }
+
+    pub(crate) fn token(&self) -> &'a CancellationToken {
+        self.cancel
+    }
+
+    /// Looks at the token without reporting.
+    pub(crate) fn check(&self) -> Result<(), HeapError> {
+        if self.cancel.is_cancelled() { Err(HeapError::Cancelled) } else { Ok(()) }
+    }
+
+    /// Looks at the token at step 0 and reports every [`CHECK_EVERY`] steps after it. Step 0 does
+    /// not report: a loop starts right after its phase's own report.
+    pub(crate) fn step(&self, step: u64, phase: HeapPhase, done: u64, total: u64) -> Result<(), HeapError> {
+        if step == 0 {
+            self.check()
+        } else if step.is_multiple_of(CHECK_EVERY) {
+            self.report(phase, done, total)
+        } else {
+            Ok(())
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::cell::RefCell;
+
+    use super::*;
+
+    #[test]
+    fn steps_check_at_zero_and_report_every_check_interval() {
+        let calls = RefCell::new(Vec::new());
+        let report = |phase, done, total| {
+            calls.borrow_mut().push((phase, done, total));
+            Ok(())
+        };
+        let cancel = CancellationToken::new();
+        let progress = Progress::new(&report, &cancel);
+        for step in [0, 1, CHECK_EVERY - 1, CHECK_EVERY, CHECK_EVERY + 1, 3 * CHECK_EVERY] {
+            progress.step(step, HeapPhase::Linking, step, 9).unwrap();
+        }
+        assert_eq!(*calls.borrow(), [(HeapPhase::Linking, CHECK_EVERY, 9), (HeapPhase::Linking, 3 * CHECK_EVERY, 9)]);
+
+        cancel.cancel();
+        assert!(matches!(progress.step(0, HeapPhase::Linking, 0, 9), Err(HeapError::Cancelled)));
+        assert!(matches!(progress.report(HeapPhase::Saving, 1, 1), Err(HeapError::Cancelled)));
+        assert_eq!(calls.borrow().last(), Some(&(HeapPhase::Saving, 1, 1)), "reports before it looks at the token");
+    }
+
+    #[test]
+    fn report_errors_pass_through() {
+        let report = |_, _, _| Err(HeapError::NotEnoughDiskSpace { needed: 2, available: 1 });
+        let cancel = CancellationToken::new();
+        let progress = Progress::new(&report, &cancel);
+        assert!(matches!(progress.report(HeapPhase::Linking, 0, 1), Err(HeapError::NotEnoughDiskSpace { .. })));
+    }
+
+    #[test]
+    fn phases_serialize_in_camel_case() {
+        assert_eq!(serde_json::to_string(&HeapPhase::Dominators).unwrap(), "\"dominators\"");
+        assert_eq!(serde_json::from_str::<HeapPhase>("\"saving\"").unwrap(), HeapPhase::Saving);
+    }
+}
@@ -0,0 +1,92 @@
+use crate::error::HeapError;
+
+/// Big-endian cursor over a dump. Every read is bounds-checked; running off `buf` yields
+/// `Truncated` at `buf.len()`, so callers bound a record by slicing `buf` to its end.
+pub(crate) struct Reader<'a> {
+    buf: &'a [u8],
+    pos: usize,
+    id_size: u8,
+}
+
+impl<'a> Reader<'a> {
+    pub(crate) fn new(buf: &'a [u8], pos: usize, id_size: u8) -> Self {
+        Self { buf, pos, id_size }
+    }
+
+    pub(crate) fn pos(&self) -> usize {
+        self.pos
+    }
+
+    pub(crate) fn id_size(&self) -> u8 {
+        self.id_size
+    }
+
+    pub(crate) fn buf(&self) -> &'a [u8] {
+        self.buf
+    }
+
+    pub(crate) fn bytes(&mut self, n: usize) -> Result<&'a [u8], HeapError> {
+        let end = self.pos.checked_add(n).filter(|&end| end <= self.buf.len());
+        let Some(end) = end else {
+            return Err(HeapError::Truncated { at: self.buf.len() as u64 });
+        };
+        let slice = &self.buf[self.pos..end];
+        self.pos = end;
+        Ok(slice)
+    }
+
+    pub(crate) fn skip(&mut self, n: usize) -> Result<(), HeapError> {
+        self.bytes(n).map(|_| ())
+    }
+
+    fn array<const N: usize>(&mut self) -> Result<[u8; N], HeapError> {
+        let mut out = [0u8; N];
+        out.copy_from_slice(self.bytes(N)?);
+        Ok(out)
+    }
+
+    pub(crate) fn u1(&mut self) -> Result<u8, HeapError> {
+        Ok(self.array::<1>()?[0])
+    }
+
+    pub(crate) fn u2(&mut self) -> Result<u16, HeapError> {
+        Ok(u16::from_be_bytes(self.array()?))
+    }
+
+    pub(crate) fn u4(&mut self) -> Result<u32, HeapError> {
+        Ok(u32::from_be_bytes(self.array()?))
+    }
+
+    pub(crate) fn u8(&mut self) -> Result<u64, HeapError> {
+        Ok(u64::from_be_bytes(self.array()?))
+    }
+
+    /// `id_size` is 4 or 8; the public entry points reject anything else before building a reader.
+    pub(crate) fn id(&mut self) -> Result<u64, HeapError> {
+        if self.id_size == 4 { self.u4().map(u64::from) } else { self.u8() }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn reads_big_endian_and_ids_of_both_sizes() {
+        let buf = [1, 0, 2, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 4, 0, 0, 0, 5];
+        let mut r = Reader::new(&buf, 0, 8);
+        assert_eq!((r.u1().unwrap(), r.u2().unwrap(), r.u4().unwrap(), r.u8().unwrap()), (1, 2, 3, 4));
+        assert_eq!(r.pos(), 15);
+        assert_eq!(Reader::new(&buf[15..], 0, 4).id().unwrap(), 5);
+    }
+
+    #[test]
+    fn running_off_the_end_reports_the_buffer_end() {
+        let buf = [0u8; 6];
+        let mut r = Reader::new(&buf, 4, 8);
+        assert!(matches!(r.u4(), Err(HeapError::Truncated { at: 6 })));
+        assert_eq!(r.pos(), 4, "a failed read does not advance");
+        assert!(matches!(r.skip(usize::MAX), Err(HeapError::Truncated { at: 6 })));
+        assert!(matches!(Reader::new(&buf, 99, 8).u1(), Err(HeapError::Truncated { at: 6 })));
+    }
+}
@@ -0,0 +1,163 @@
+//! The heap snapshots that are open in this process.
+
+use std::collections::BTreeMap;
+use std::path::{Path, PathBuf};
+use std::sync::{Arc, Mutex, MutexGuard, PoisonError};
+
+use ide_core::process::process_alive;
+use ide_core::{DebugSessionId, IpcPath};
+use serde::{Deserialize, Serialize};
+
+use crate::disk::remove_stale;
+use crate::index::{HeapIndex, HeapSummary};
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash, Serialize, Deserialize, specta::Type)]
+#[serde(transparent)]
+pub struct HeapSnapshotId(pub u32);
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, specta::Type)]
+#[serde(tag = "kind", rename_all = "camelCase", rename_all_fields = "camelCase")]
+pub enum HeapOrigin {
+    /// Dumped from a live debug session.
+    Captured { session: DebugSessionId, pid: u32, java_version: String, taken_while_suspended: bool, dump_ms: u32 },
+    /// An `.hprof` the user opened; it is never deleted.
+    File { path: IpcPath },
+}
+
+impl HeapOrigin {
+    /// Text identifying the snapshot in the UI: the process id of a capture, the file name of a
+    /// file. The capture time is `created_ms`, which the UI formats in local time.
+    pub fn label(&self) -> String {
+        match self {
+            HeapOrigin::Captured { pid, .. } => pid.to_string(),
+            HeapOrigin::File { path } => path.file_name().unwrap_or(path.as_os_str()).to_string_lossy().into_owned(),
+        }
+    }
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, specta::Type)]
+#[serde(rename_all = "camelCase")]
+pub struct HeapSnapshotInfo {
+    pub id: HeapSnapshotId,
+    pub label: String,
+    pub origin: HeapOrigin,
+    pub summary: HeapSummary,
+    /// Unix time in milliseconds when the snapshot was opened.
+    pub created_ms: u64,
+}
+
+/// An analyzed snapshot together with the directory that holds its files. Dropping it unmaps the
+/// index and then removes the directory, dump included for a capture; a dump opened from a file
+/// is outside the directory and stays.
+pub struct OpenSnapshot {
+    pub info: HeapSnapshotInfo,
+    /// `None` only while dropping, so the mappings are gone before the directory is deleted
+    /// (Windows cannot delete a mapped file).
+    index: Option<HeapIndex>,
+    /// Holds the index files and the dump of a capture.
+    pub dir: PathBuf,
+    pub dump: PathBuf,
+    /// Set by [`HeapRegistry::insert`] once `dir` is known to be one of the registry's own.
+    remove_dir: bool,
+}
+
+impl OpenSnapshot {
+    /// The directory is not removed on drop until the snapshot is in a [`HeapRegistry`] whose
+    /// root holds it.
+    pub fn new(info: HeapSnapshotInfo, index: HeapIndex, dir: PathBuf, dump: PathBuf) -> Self {
+        Self { info, index: Some(index), dir, dump, remove_dir: false }
+    }
+
+    pub fn index(&self) -> &HeapIndex {
+        self.index.as_ref().expect("the index is only taken while dropping")
+    }
+}
+
+impl Drop for OpenSnapshot {
+    fn drop(&mut self) {
+        drop(self.index.take());
+        if !self.remove_dir {
+            return;
+        }
+        // The user's own dump that happens to sit inside `dir` (an opened file is analyzed from a
+        // private copy, which goes with the directory).
+        if matches!(&self.info.origin, HeapOrigin::File { path } if **path == self.dump)
+            && self.dump.starts_with(&self.dir)
+        {
+            return;
+        }
+        // A link is not followed: only a real directory is deleted.
+        if std::fs::symlink_metadata(&self.dir).is_ok_and(|meta| meta.file_type().is_dir())
+            && let Err(err) = std::fs::remove_dir_all(&self.dir)
+        {
+            tracing::warn!(dir = %self.dir.display(), %err, "could not remove a heap snapshot directory");
+        }
+    }
+}
+
+struct State {
+    next_id: u32,
+    open: BTreeMap<u32, Arc<OpenSnapshot>>,
+}
+
+/// The open snapshots, shared between threads. Ids are never reused within a process.
+pub struct HeapRegistry {
+    root: PathBuf,
+    state: Mutex<State>,
+}
+
+impl HeapRegistry {
+    /// Registry for snapshot directories under `root`; first removes the directories there that
+    /// belong to processes that no longer exist.
+    pub fn new(root: PathBuf) -> Self {
+        remove_stale(&root, process_alive);
+        Self { root, state: Mutex::new(State { next_id: 1, open: BTreeMap::new() }) }
+    }
+
+    pub fn root(&self) -> &Path {
+        &self.root
+    }
+
+    fn state(&self) -> MutexGuard<'_, State> {
+        self.state.lock().unwrap_or_else(PoisonError::into_inner)
+    }
+
+    /// Registers the snapshot that `make` builds for a fresh id. `make` runs outside the lock.
+    pub fn insert(&self, make: impl FnOnce(HeapSnapshotId) -> OpenSnapshot) -> HeapSnapshotInfo {
+        let id = {
+            let mut state = self.state();
+            let id = state.next_id;
+            state.next_id = id.saturating_add(1);
+            HeapSnapshotId(id)
+        };
+        let mut snapshot = make(id);
+        snapshot.info.id = id;
+        snapshot.remove_dir = snapshot.dir.file_name().is_some() && snapshot.dir.parent() == Some(self.root.as_path());
+        if !snapshot.remove_dir {
+            tracing::warn!(dir = %snapshot.dir.display(), "heap snapshot directory is outside the registry root; keeping it");
+        }
+        let info = snapshot.info.clone();
+        self.state().open.insert(id.0, Arc::new(snapshot));
+        info
+    }
+
+    pub fn get(&self, id: HeapSnapshotId) -> Option<Arc<OpenSnapshot>> {
+        self.state().open.get(&id.0).cloned()
+    }
+
+    /// Forgets the snapshot; its directory goes once the last query using it lets go. `false` for
+    /// an id that is not open.
+    pub fn close(&self, id: HeapSnapshotId) -> bool {
+        // Bound first so the lock is released before the snapshot may be dropped: that deletes files.
+        let removed = self.state().open.remove(&id.0);
+        removed.is_some()
+    }
+
+    /// In the order the snapshots were opened.
+    pub fn list(&self) -> Vec<HeapSnapshotInfo> {
+        self.state().open.values().map(|snapshot| snapshot.info.clone()).collect()
+    }
+}
+
+#[cfg(test)]
+mod tests;
@@ -0,0 +1,265 @@
+use std::sync::Arc;
+
+use ide_core::{CancellationToken, DebugSessionId};
+
+use super::*;
+use crate::analysis::{AnalyzeOptions, analyze};
+use crate::disk::create_snapshot_dir;
+use crate::error::HeapError;
+use crate::progress::HeapPhase;
+use crate::testing::sample;
+
+fn no_progress(_: HeapPhase, _: u64, _: u64) -> Result<(), HeapError> {
+    Ok(())
+}
+
+fn captured() -> HeapOrigin {
+    HeapOrigin::Captured {
+        session: DebugSessionId(7),
+        pid: 4242,
+        java_version: "21.0.2".into(),
+        taken_while_suspended: true,
+        dump_ms: 35,
+    }
+}
+
+fn external(path: &Path) -> HeapOrigin {
+    HeapOrigin::File { path: IpcPath(path.to_path_buf()) }
+}
+
+/// A snapshot whose index files are mapped from `dir`, like a real one: `dump` is analyzed in
+/// place and `dir` receives the columns.
+fn open_in(dir: &Path, dump: &Path, origin: HeapOrigin) -> impl FnOnce(HeapSnapshotId) -> OpenSnapshot {
+    let dir = dir.to_path_buf();
+    let dump = dump.to_path_buf();
+    move |id| {
+        std::fs::write(&dump, sample().bytes()).unwrap();
+        let index = analyze(&dump, &dir, AnalyzeOptions::default(), &no_progress, &CancellationToken::new()).unwrap();
+        let info = HeapSnapshotInfo {
+            id,
+            label: origin.label(),
+            origin,
+            summary: index.summary.clone(),
+            created_ms: 1_700_000_000_000,
+        };
+        OpenSnapshot::new(info, index, dir, dump)
+    }
+}
+
+/// Inserts a captured snapshot with its own directory under the registry's root.
+fn insert_captured(registry: &HeapRegistry) -> (HeapSnapshotInfo, PathBuf) {
+    let dir = create_snapshot_dir(registry.root()).unwrap();
+    let dump = dir.join("dump.hprof");
+    let info = registry.insert(open_in(&dir, &dump, captured()));
+    (info, dir)
+}
+
+fn names(dir: &Path) -> Vec<String> {
+    let mut names: Vec<_> =
+        std::fs::read_dir(dir).unwrap().map(|e| e.unwrap().file_name().into_string().unwrap()).collect();
+    names.sort();
+    names
+}
+
+#[test]
+fn closing_snapshot_removes_dir() {
+    let root = tempfile::tempdir().unwrap();
+    let registry = HeapRegistry::new(root.path().to_path_buf());
+    let (info, dir) = insert_captured(&registry);
+    assert!(dir.join("dump.hprof").exists() && dir.join("summary.json").exists());
+    assert!(registry.close(info.id));
+    assert!(!dir.exists());
+    assert!(!registry.close(info.id));
+    assert!(registry.get(info.id).is_none());
+}
+
+#[test]
+fn external_dump_is_never_deleted() {
+    let root = tempfile::tempdir().unwrap();
+    let outside = tempfile::tempdir().unwrap();
+    let dump = outside.path().join("app.hprof");
+    let registry = HeapRegistry::new(root.path().to_path_buf());
+    let dir = create_snapshot_dir(registry.root()).unwrap();
+    let info = registry.insert(open_in(&dir, &dump, external(&dump)));
+    assert_eq!(info.label, "app.hprof");
+    assert!(dir.join("summary.json").exists());
+    assert!(registry.close(info.id));
+    assert!(!dir.exists());
+    assert!(dump.exists());
+}
+
+#[test]
+fn a_dump_inside_the_dir_of_a_file_snapshot_keeps_the_dir() {
+    let root = tempfile::tempdir().unwrap();
+    let registry = HeapRegistry::new(root.path().to_path_buf());
+    let dir = create_snapshot_dir(registry.root()).unwrap();
+    let dump = dir.join("opened.hprof");
+    let info = registry.insert(open_in(&dir, &dump, external(&dump)));
+    assert!(registry.close(info.id));
+    assert!(dump.exists());
+}
+
+#[test]
+fn a_dir_outside_the_root_is_never_deleted() {
+    let root = tempfile::tempdir().unwrap();
+    let elsewhere = tempfile::tempdir().unwrap();
+    let registry = HeapRegistry::new(root.path().to_path_buf());
+    let dump = elsewhere.path().join("dump.hprof");
+    let info = registry.insert(open_in(elsewhere.path(), &dump, captured()));
+    assert!(registry.close(info.id));
+    assert!(dump.exists());
+    // Neither the root itself nor a path that climbs out of it counts as a snapshot dir.
+    let sneaky = root.path().join("..").join(elsewhere.path().file_name().unwrap());
+    let info = registry.insert(open_in(&sneaky, &dump, captured()));
+    assert!(registry.close(info.id));
+    assert!(dump.exists());
+}
+
+#[cfg(unix)]
+#[test]
+fn a_symlinked_dir_is_not_followed() {
+    let root = tempfile::tempdir().unwrap();
+    let outside = tempfile::tempdir().unwrap();
+    let dump = outside.path().join("dump.hprof");
+    let link = root.path().join("link");
+    std::os::unix::fs::symlink(outside.path(), &link).unwrap();
+    let registry = HeapRegistry::new(root.path().to_path_buf());
+    let info = registry.insert(open_in(&link, &dump, captured()));
+    assert!(registry.close(info.id));
+    assert!(dump.exists());
+}
+
+#[test]
+fn ids_are_never_reused() {
+    let root = tempfile::tempdir().unwrap();
+    let registry = HeapRegistry::new(root.path().to_path_buf());
+    let (first, _) = insert_captured(&registry);
+    assert!(registry.close(first.id));
+    let (second, _) = insert_captured(&registry);
+    let (third, _) = insert_captured(&registry);
+    assert!(first.id.0 < second.id.0 && second.id.0 < third.id.0);
+}
+
+#[test]
+fn the_registry_decides_the_id_in_the_info() {
+    let root = tempfile::tempdir().unwrap();
+    let registry = HeapRegistry::new(root.path().to_path_buf());
+    let dir = create_snapshot_dir(registry.root()).unwrap();
+    let dump = dir.join("dump.hprof");
+    let make = open_in(&dir, &dump, captured());
+    let info = registry.insert(|id| {
+        let mut snapshot = make(id);
+        snapshot.info.id = HeapSnapshotId(u32::MAX);
+        snapshot
+    });
+    assert_ne!(info.id, HeapSnapshotId(u32::MAX));
+    assert_eq!(registry.get(info.id).unwrap().info.id, info.id);
+}
+
+#[test]
+fn list_is_in_opening_order_and_get_shares_the_snapshot() {
+    let root = tempfile::tempdir().unwrap();
+    let registry = HeapRegistry::new(root.path().to_path_buf());
+    let (a, _) = insert_captured(&registry);
+    let (b, _) = insert_captured(&registry);
+    let (c, _) = insert_captured(&registry);
+    assert!(registry.close(b.id));
+    assert_eq!(registry.list(), [a.clone(), c.clone()]);
+    let one = registry.get(a.id).unwrap();
+    let two = registry.get(a.id).unwrap();
+    assert!(Arc::ptr_eq(&one, &two));
+    assert_eq!(one.info, a);
+    assert_eq!(one.index().summary, a.summary);
+}
+
+#[test]
+fn a_snapshot_in_use_keeps_its_dir_until_the_last_holder_lets_go() {
+    let root = tempfile::tempdir().unwrap();
+    let registry = HeapRegistry::new(root.path().to_path_buf());
+    let (info, dir) = insert_captured(&registry);
+    let held = registry.get(info.id).unwrap();
+    assert!(registry.close(info.id));
+    assert!(registry.get(info.id).is_none() && registry.list().is_empty());
+    assert!(dir.exists());
+    drop(held);
+    assert!(!dir.exists());
+}
+
+#[test]
+fn dropping_the_registry_removes_the_open_snapshots() {
+    let root = tempfile::tempdir().unwrap();
+    let registry = HeapRegistry::new(root.path().to_path_buf());
+    let (_, dir) = insert_captured(&registry);
+    drop(registry);
+    assert!(!dir.exists());
+    assert!(names(root.path()).is_empty());
+}
+
+#[cfg(unix)]
+#[test]
+fn new_clears_dirs_whose_owner_died() {
+    let root = tempfile::tempdir().unwrap();
+    let mut child = std::process::Command::new("true").spawn().unwrap();
+    let dead = child.id();
+    child.wait().unwrap();
+    let dead_dir = root.path().join("dead");
+    std::fs::create_dir(&dead_dir).unwrap();
+    std::fs::write(dead_dir.join("owner"), dead.to_string()).unwrap();
+    let live_dir = create_snapshot_dir(root.path()).unwrap();
+    let registry = HeapRegistry::new(root.path().to_path_buf());
+    assert!(!dead_dir.exists());
+    assert!(live_dir.exists());
+    assert_eq!(registry.root(), root.path());
+}
+
+#[test]
+fn concurrent_inserts_get_distinct_ids() {
+    let root = tempfile::tempdir().unwrap();
+    let registry = Arc::new(HeapRegistry::new(root.path().to_path_buf()));
+    let handles: Vec<_> = (0..6)
+        .map(|_| {
+            let registry = registry.clone();
+            std::thread::spawn(move || (0..3).map(|_| insert_captured(&registry).0.id).collect::<Vec<_>>())
+        })
+        .collect();
+    let mut ids: Vec<_> = handles.into_iter().flat_map(|h| h.join().unwrap()).collect();
+    ids.sort_by_key(|id| id.0);
+    ids.dedup();
+    assert_eq!(ids.len(), 18);
+    assert_eq!(registry.list().len(), 18);
+    for id in ids {
+        assert!(registry.close(id));
+    }
+    assert!(names(root.path()).is_empty());
+}
+
+#[test]
+fn labels_name_the_pid_or_the_file() {
+    assert_eq!(captured().label(), "4242");
+    assert_eq!(external(Path::new("/tmp/dumps/app.hprof")).label(), "app.hprof");
+    assert_eq!(external(Path::new("/")).label(), "/");
+}
+
+#[test]
+fn origin_and_info_serialize_as_the_ui_expects() {
+    let value = serde_json::to_value(captured()).unwrap();
+    assert_eq!(
+        value,
+        serde_json::json!({
+            "kind": "captured", "session": 7, "pid": 4242, "javaVersion": "21.0.2",
+            "takenWhileSuspended": true, "dumpMs": 35
+        })
+    );
+    assert_eq!(serde_json::to_value(external(Path::new("/a/b.hprof"))).unwrap()["path"], "/a/b.hprof");
+    assert_eq!(serde_json::to_value(HeapSnapshotId(3)).unwrap(), serde_json::json!(3));
+    let back: HeapOrigin = serde_json::from_value(value).unwrap();
+    assert_eq!(back, captured());
+}
+
+// `HeapRegistry::get` hands out `Arc<OpenSnapshot>` to query threads.
+const _: fn() = || {
+    fn check<T: Send + Sync>() {}
+    check::<HeapIndex>();
+    check::<OpenSnapshot>();
+    check::<HeapRegistry>();
+};
@@ -0,0 +1,265 @@
+//! Shallow sizes: the bytes each object takes in the JVM heap, from the object layout rules, the
+//! JVM's own class histogram, or the distances between object addresses in the dump.
+
+mod layout;
+mod mirrors;
+
+use std::collections::HashMap;
+
+use ide_core::CancellationToken;
+use serde::{Deserialize, Serialize};
+
+pub use self::layout::VmLayout;
+use self::mirrors::{Mirrors, injected_by_table};
+use crate::error::HeapError;
+use crate::graph::{ClassInfo, HeapGraph};
+use crate::hprof::BasicType;
+use crate::names::display_name;
+use crate::progress::CHECK_EVERY;
+
+/// How far past the layout size an address gap may be and still count as an object size.
+const GAP_SLACK: u64 = 64;
+/// Share of the instance types that can have a gap (two or more instances), in tenths, that must
+/// take their size from address gaps for the sizes to be reported as coming from them.
+const GAP_TRUST_TENTHS: usize = 9;
+
+/// Where the sizes of instances came from.
+///
+/// Class objects (mirrors) are approximate in every case: HotSpot leaves the fields it injects
+/// into `java.lang.Class` out of the dump, so their layout size is short by a fixed amount per
+/// mirror. With a histogram that amount is fitted to its `java.lang.Class` row; otherwise (or when
+/// the fit is rejected) it comes from a table measured per JDK and reference size, within a few
+/// bytes per mirror, or is 32 when the JDK is unknown.
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize, specta::Type)]
+#[serde(rename_all = "camelCase")]
+pub enum SizeSource {
+    /// The JVM's class histogram from the capture.
+    Jvm,
+    /// The smallest distance between neighbouring objects of a type, in a dump the IDE did not
+    /// capture.
+    AddressGaps,
+    Layout,
+}
+
+/// A row of the JVM's class histogram.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct HistogramEntry {
+    /// Display form: `java.lang.String`, `int[]`, `Gen$$Lambda/0x…`.
+    pub name: String,
+    pub instances: u64,
+    pub bytes: u64,
+}
+
+enum TypeSize {
+    Fixed(u64),
+    ObjectArray,
+    /// Class objects are sized one by one; this is for an instance of `java.lang.Class` that is
+    /// not a class object.
+    ClassObject(u64),
+}
+
+/// Shallow size of every object of a [`HeapGraph`]. Instances of a type share one size; arrays
+/// follow the layout; class objects add their statics to the size of `java.lang.Class`.
+pub struct ShallowModel {
+    pub source: SizeSource,
+    layout: VmLayout,
+    /// Per type slot of a class; primitive array slots follow and have no entry.
+    types: Vec<TypeSize>,
+    /// `(node, size)` of each class object, by node.
+    class_objects: Vec<(u32, u64)>,
+    /// Per class slot: another class has the same display name.
+    ambiguous: Vec<bool>,
+}
+
+impl ShallowModel {
+    /// `histogram` is the JVM's class histogram of the same stop as the dump, when the IDE captured
+    /// it. Without it, sizes come from address gaps where they agree with the layout.
+    ///
+    /// A name shared by several classes (several class loaders) never takes a histogram size,
+    /// since the name is the only link between the dump and the histogram; see
+    /// [`Self::is_ambiguous`]. `java.lang.Class` never does either, because mirror sizes vary with
+    /// their statics; the histogram only tells how many bytes HotSpot adds to each mirror.
+    pub fn new(
+        graph: &HeapGraph,
+        layout: VmLayout,
+        histogram: Option<&[HistogramEntry]>,
+        cancel: &CancellationToken,
+    ) -> Result<Self, HeapError> {
+        if cancel.is_cancelled() {
+            return Err(HeapError::Cancelled);
+        }
+        let classes = &graph.classes;
+        let field_bytes = field_bytes(classes, layout);
+        let by_layout: Vec<u64> = field_bytes.iter().map(|&bytes| layout.instance_size(bytes)).collect();
+        let names: Vec<String> = classes.iter().map(|c| display_name(&c.name)).collect();
+        let ambiguous = ambiguous_names(&names);
+
+        let class_slot = graph.class_slot as usize;
+        let mirrors = Mirrors::collect(graph, by_layout[class_slot], layout);
+        let mut fitted = None;
+        if let Some(rows) = histogram
+            && !ambiguous[class_slot]
+        {
+            let mut class_rows = rows.iter().filter(|row| row.name == names[class_slot]);
+            if let (Some(row), None) = (class_rows.next(), class_rows.next()) {
+                fitted = mirrors.fit_injected(graph, layout, by_layout[class_slot], row, cancel)?;
+            }
+        }
+        let injected = fitted.unwrap_or_else(|| injected_by_table(layout));
+
+        let mut types: Vec<TypeSize> = classes
+            .iter()
+            .enumerate()
+            .map(|(slot, class)| {
+                if class.name.starts_with('[') {
+                    TypeSize::ObjectArray
+                } else if slot == class_slot {
+                    TypeSize::ClassObject(layout.align(by_layout[slot] + injected))
+                } else {
+                    TypeSize::Fixed(by_layout[slot])
+                }
+            })
+            .collect();
+        let source = match histogram {
+            Some(rows) => {
+                apply_histogram(&mut types, &names, &ambiguous, rows);
+                SizeSource::Jvm
+            }
+            None => apply_address_gaps(&mut types, graph, &by_layout, cancel)?,
+        };
+        Ok(Self { source, layout, types, class_objects: mirrors.sizes(layout, injected), ambiguous })
+    }
+
+    /// Shallow size of `object` (a node of the graph); 0 for the super-root.
+    pub fn size_of(&self, graph: &HeapGraph, object: u32) -> u64 {
+        let Some(&slot) = graph.types.get(object as usize) else { return 0 };
+        match self.types.get(slot as usize) {
+            Some(&TypeSize::Fixed(size)) => size,
+            Some(TypeSize::ObjectArray) => self.array_size(BasicType::Object, graph.lens[object as usize]),
+            Some(&TypeSize::ClassObject(fallback)) => self
+                .class_objects
+                .binary_search_by_key(&object, |&(node, _)| node)
+                .map_or(fallback, |at| self.class_objects[at].1),
+            None => match graph.prim_type(slot) {
+                Some(ty) => self.array_size(ty, graph.lens[object as usize]),
+                None => 0,
+            },
+        }
+    }
+
+    /// Whether another class has the same display name as the class in `slot`. Sizes of such
+    /// classes come from the layout, and anything joined to them by name (histogram rows, a
+    /// second snapshot) cannot tell them apart.
+    pub fn is_ambiguous(&self, slot: u32) -> bool {
+        self.ambiguous.get(slot as usize).copied().unwrap_or(false)
+    }
+
+    fn array_size(&self, ty: BasicType, len: u32) -> u64 {
+        self.layout.array_size(ty, u64::from(len))
+    }
+}
+
+/// Bytes of instance fields per class, its superclasses included. Super chains always end
+/// (`ClassInfo::super_index` has its cycles cut).
+fn field_bytes(classes: &[ClassInfo], layout: VmLayout) -> Vec<u64> {
+    let mut total: Vec<Option<u64>> = vec![None; classes.len()];
+    let mut path = Vec::new();
+    for start in 0..classes.len() {
+        let mut cursor = Some(start);
+        let mut below = 0;
+        while let Some(c) = cursor {
+            if let Some(known) = total[c] {
+                below = known;
+                break;
+            }
+            path.push(c);
+            cursor = classes[c].super_index.map(|s| s as usize);
+        }
+        while let Some(c) = path.pop() {
+            let own: u64 = classes[c].fields.iter().map(|(_, ty)| ty.heap_size(layout.ref_size)).sum();
+            below += own;
+            total[c] = Some(below);
+        }
+    }
+    total.into_iter().map(Option::unwrap_or_default).collect()
+}
+
+fn ambiguous_names(names: &[String]) -> Vec<bool> {
+    let mut counts: HashMap<&str, u32> = HashMap::with_capacity(names.len());
+    for name in names {
+        *counts.entry(name).or_default() += 1;
+    }
+    names.iter().map(|name| counts[name.as_str()] > 1).collect()
+}
+
+/// `bytes / instances` for each plain class with a histogram row of its name. A name that is
+/// ambiguous in the dump or repeated in the histogram, or a row that is not a whole number of
+/// objects, keeps its layout size.
+fn apply_histogram(types: &mut [TypeSize], names: &[String], ambiguous: &[bool], rows: &[HistogramEntry]) {
+    let mut per_object: HashMap<&str, Option<u64>> = HashMap::with_capacity(rows.len());
+    for row in rows {
+        let size = row.bytes.checked_div(row.instances).filter(|&s| s > 0 && row.bytes % row.instances == 0);
+        per_object.entry(&row.name).and_modify(|seen| *seen = None).or_insert(size);
+    }
+    for (slot, ty) in types.iter_mut().enumerate() {
+        if let TypeSize::Fixed(size) = ty
+            && !ambiguous[slot]
+            && let Some(&Some(from_histogram)) = per_object.get(names[slot].as_str())
+        {
+            *size = from_histogram;
+        }
+    }
+}
+
+/// Takes the smallest distance from an instance to the next object in address order as the size
+/// of its type, for types with two or more instances where that distance is no smaller than the
+/// layout size and at most [`GAP_SLACK`] bigger. The result says [`SizeSource::AddressGaps`] when
+/// nine in ten of the types with two or more instances did; types with one instance have no gap
+/// to measure, so they neither count for nor against.
+fn apply_address_gaps(
+    types: &mut [TypeSize],
+    graph: &HeapGraph,
+    by_layout: &[u64],
+    cancel: &CancellationToken,
+) -> Result<SizeSource, HeapError> {
+    let slots = types.len();
+    let mut instances = vec![0u32; slots];
+    let mut smallest_gap = vec![u64::MAX; slots];
+    let mut ids = graph.ids.ids();
+    let mut current = ids.next();
+    let mut object = 0usize;
+    while let Some(id) = current {
+        if (object as u64).is_multiple_of(CHECK_EVERY) && cancel.is_cancelled() {
+            return Err(HeapError::Cancelled);
+        }
+        let next = ids.next();
+        let slot = graph.types[object] as usize;
+        if matches!(types.get(slot), Some(TypeSize::Fixed(_))) {
+            instances[slot] = instances[slot].saturating_add(1);
+            if let Some(next) = next.filter(|&n| n > id) {
+                smallest_gap[slot] = smallest_gap[slot].min(next - id);
+            }
+        }
+        current = next;
+        object += 1;
+    }
+
+    let (mut measurable, mut by_gap) = (0usize, 0usize);
+    for (slot, ty) in types.iter_mut().enumerate() {
+        let TypeSize::Fixed(size) = ty else { continue };
+        if instances[slot] < 2 {
+            continue;
+        }
+        measurable += 1;
+        let gap = smallest_gap[slot];
+        if (by_layout[slot]..=by_layout[slot] + GAP_SLACK).contains(&gap) {
+            *size = gap;
+            by_gap += 1;
+        }
+    }
+    let trusted = measurable > 0 && by_gap * 10 >= measurable * GAP_TRUST_TENTHS;
+    Ok(if trusted { SizeSource::AddressGaps } else { SizeSource::Layout })
+}
+
+#[cfg(test)]
+mod tests;
@@ -0,0 +1,113 @@
+//! The object layout rules of the JVM that wrote a dump.
+
+use std::collections::HashMap;
+
+use crate::hprof::BasicType;
+
+/// Object layout of the JVM that wrote the dump.
+#[derive(Clone, Copy, Debug, PartialEq)]
+pub struct VmLayout {
+    pub header: u8,
+    /// Bytes of a reference: 4 with compressed oops.
+    pub ref_size: u8,
+    /// Offset of the length field of an array.
+    pub length_offset: u8,
+    /// JDK 21 and older put array elements at the next 8-byte boundary after the length, later
+    /// JDKs at the next boundary of the element size.
+    pub legacy_array_base: bool,
+    pub alignment: u16,
+    /// The JDK the flags came from; `None` when nothing is known about the JVM.
+    pub java_major: Option<u32>,
+}
+
+impl VmLayout {
+    pub const DEFAULT: VmLayout = VmLayout {
+        header: 12,
+        ref_size: 4,
+        length_offset: 12,
+        legacy_array_base: false,
+        alignment: 8,
+        java_major: None,
+    };
+
+    /// `flags` is the output of `VM.flags`, which lists the flags set on this run. Compressed oops
+    /// are off by default and only switched on by ergonomics, so `-XX:+UseCompressedOops` is
+    /// printed exactly when references take 4 bytes; a heap too big for them prints no such flag.
+    /// Before JDK 15 compressed class pointers follow the same rule; with `java_major: None` (JDK
+    /// unknown) they follow the JDK 15+ rule and the array base that of the newest JDKs. Flags
+    /// without any `-XX:` option (an external dump) give [`Self::DEFAULT`] with the array base of
+    /// `java_major` and an unknown JDK, since the reference size is a guess.
+    /// An `ObjectAlignmentInBytes` that is not a power of two in `8..=256` is ignored.
+    pub fn from_vm_flags(flags: &str, java_major: Option<u32>) -> Self {
+        let legacy_array_base = java_major.map_or(Self::DEFAULT.legacy_array_base, |major| major <= 21);
+        let flags = parse_flags(flags);
+        if flags.is_empty() {
+            return VmLayout { legacy_array_base, ..Self::DEFAULT };
+        }
+        let on = |name: &str| flags.get(name) == Some(&"true");
+        let compressed_classes = if java_major.is_none_or(|major| major >= 15) {
+            flags.get("UseCompressedClassPointers") != Some(&"false")
+        } else {
+            on("UseCompressedClassPointers")
+        };
+        let (header, length_offset) = if on("UseCompactObjectHeaders") {
+            (8, 8)
+        } else if compressed_classes {
+            (12, 12)
+        } else {
+            (16, 16)
+        };
+        let alignment = flags
+            .get("ObjectAlignmentInBytes")
+            .and_then(|v| v.parse::<u16>().ok())
+            .filter(|a| (8..=256).contains(a) && a.is_power_of_two())
+            .unwrap_or(Self::DEFAULT.alignment);
+        VmLayout {
+            header,
+            ref_size: if on("UseCompressedOops") { 4 } else { 8 },
+            length_offset,
+            legacy_array_base,
+            alignment,
+            java_major,
+        }
+    }
+
+    /// Size of an instance whose fields (its class and every superclass) take `field_bytes`.
+    pub fn instance_size(&self, field_bytes: u64) -> u64 {
+        let smallest = if self.header == 8 { 8 } else { 16 };
+        self.align(u64::from(self.header) + field_bytes).max(smallest)
+    }
+
+    pub fn array_size(&self, ty: BasicType, len: u64) -> u64 {
+        let element = ty.heap_size(self.ref_size);
+        let base_alignment = if self.legacy_array_base { 8 } else { element };
+        let base = align_to(u64::from(self.length_offset) + 4, base_alignment);
+        self.align(base + element * len)
+    }
+
+    /// `value` rounded up to the object alignment.
+    pub(super) fn align(&self, value: u64) -> u64 {
+        align_to(value, u64::from(self.alignment))
+    }
+}
+
+fn align_to(value: u64, to: u64) -> u64 {
+    value.div_ceil(to.max(1)) * to.max(1)
+}
+
+/// `-XX:+Name` / `-XX:-Name` / `-XX:Name=value` tokens by name; booleans become `"true"` / `"false"`.
+/// A flag given twice keeps its last value, as in the JVM.
+fn parse_flags(flags: &str) -> HashMap<&str, &str> {
+    let mut parsed = HashMap::new();
+    for token in flags.split_whitespace() {
+        let Some(body) = token.strip_prefix("-XX:") else { continue };
+        if let Some(name) = body.strip_prefix('+') {
+            parsed.insert(name, "true");
+        } else if let Some(name) = body.strip_prefix('-') {
+            parsed.insert(name, "false");
+        } else if let Some((name, value)) = body.split_once('=') {
+            parsed.insert(name, value);
+        }
+    }
+    parsed
+}
@@ -0,0 +1,119 @@
+//! Sizes of class objects (mirrors): the `java.lang.Class` instance plus the class's statics, plus
+//! the fields HotSpot injects into every mirror and leaves out of the class dump.
+
+use ide_core::CancellationToken;
+
+use super::{HistogramEntry, VmLayout};
+use crate::error::HeapError;
+use crate::graph::HeapGraph;
+use crate::progress::CHECK_EVERY;
+
+/// Largest injected byte count (per mirror, a multiple of 4) the histogram fit tries.
+const MAX_INJECTED: u64 = 128;
+/// How far, in percent of the histogram's total, the best fit may be off before the row is taken
+/// for something else than the mirrors of this dump.
+const MAX_FIT_ERROR_PERCENT: u128 = 5;
+
+/// Bytes HotSpot injects into every mirror (`klass`, `array_klass`, `oop_size`,
+/// `static_oop_field_count`, and oops such as `protection_domain`) without listing them in the class
+/// dump, as measured by comparing the dump with `GC.class_histogram` on JDK 8, 11, 21, 23 and 25,
+/// with and without compressed oops. Each X is the best fit in steps of 4; the real value is
+/// within a few bytes (statics padding). `None` for a reference size the table does not cover.
+pub(super) fn injected_mirror_bytes(java_major: u32, ref_size: u8) -> Option<u64> {
+    match (ref_size, java_major) {
+        // 8, 11, 21 and 23 measure 36; 25 is the first JDK that shows 28.
+        (4, ..=24) => Some(36),
+        (4, 25..) => Some(28),
+        // 8, 11 and 21 measure 44; 23 is the first to show 40 and 25 the first to show 32.
+        (8, ..=22) => Some(44),
+        (8, 23..=24) => Some(40),
+        (8, 25..) => Some(32),
+        _ => None,
+    }
+}
+
+/// Used when the JVM is unknown (a dump without `VM.flags`): the midpoint of the measured values
+/// with 4-byte references, 36 and 28.
+const UNKNOWN_JVM_INJECTED: u64 = 32;
+
+/// The injected bytes per mirror the layout implies, without looking at a histogram.
+pub(super) fn injected_by_table(layout: VmLayout) -> u64 {
+    layout.java_major.and_then(|major| injected_mirror_bytes(major, layout.ref_size)).unwrap_or(UNKNOWN_JVM_INJECTED)
+}
+
+pub(super) struct Mirrors {
+    /// `(node, size before alignment)` of every class dump that is an object of the graph, by node.
+    dumped: Vec<(u32, u64)>,
+}
+
+impl Mirrors {
+    /// `instance` is the layout size of `java.lang.Class`, the base of every mirror.
+    pub(super) fn collect(graph: &HeapGraph, instance: u64, layout: VmLayout) -> Self {
+        let mut dumped: Vec<(u32, u64)> = graph
+            .classes
+            .iter()
+            .filter(|class| class.id != 0)
+            .filter_map(|class| {
+                let node = graph.ids.rank(class.id)?;
+                let references = class.statics.len() as u64 * u64::from(layout.ref_size);
+                Some((node, instance + class.static_bytes + references))
+            })
+            .collect();
+        dumped.sort_unstable_by_key(|&(node, _)| node);
+        Self { dumped }
+    }
+
+    /// `(node, size)` of each class object when HotSpot adds `injected` bytes to every mirror.
+    pub(super) fn sizes(&self, layout: VmLayout, injected: u64) -> Vec<(u32, u64)> {
+        self.dumped.iter().map(|&(node, bytes)| (node, layout.align(bytes + injected))).collect()
+    }
+
+    /// The injected bytes per mirror (a multiple of 4, at most 128) that make the mirrors of this
+    /// dump add up to `class_row`'s share of the histogram, or `None` when no value gets within 5%:
+    /// a histogram of other mirrors than these (a different JVM stop) must not skew the sizes.
+    ///
+    /// The fit compares averages, since the histogram also counts mirrors the dump has no class
+    /// dump for (primitive types, classes only archived by CDS). Every mirror still gets the same
+    /// extra bytes, never the average size, because statics make mirror sizes differ.
+    pub(super) fn fit_injected(
+        &self,
+        graph: &HeapGraph,
+        layout: VmLayout,
+        instance: u64,
+        class_row: &HistogramEntry,
+        cancel: &CancellationToken,
+    ) -> Result<Option<u64>, HeapError> {
+        if class_row.instances == 0 {
+            return Ok(None);
+        }
+        // Mirrors without a class dump (primitive types) have no statics.
+        let class_typed = count_slot(graph, graph.class_slot, cancel)?;
+        let bare = class_typed.saturating_sub(self.dumped.len() as u64);
+        let mirrors = self.dumped.len() as u64 + bare;
+        if mirrors == 0 {
+            return Ok(None);
+        }
+        let target = u128::from(class_row.bytes) * u128::from(mirrors) / u128::from(class_row.instances);
+        let total = |injected: u64| -> u128 {
+            let dumped: u128 = self.dumped.iter().map(|&(_, bytes)| u128::from(layout.align(bytes + injected))).sum();
+            dumped + u128::from(bare) * u128::from(layout.align(instance + injected))
+        };
+        let (best, error) = (0..=MAX_INJECTED)
+            .step_by(4)
+            .map(|injected| (injected, total(injected).abs_diff(target)))
+            .min_by_key(|&(injected, error)| (error, injected))
+            .unwrap_or((0, u128::MAX));
+        Ok((error * 100 <= target * MAX_FIT_ERROR_PERCENT).then_some(best))
+    }
+}
+
+fn count_slot(graph: &HeapGraph, slot: u32, cancel: &CancellationToken) -> Result<u64, HeapError> {
+    let mut count = 0;
+    for chunk in graph.types.chunks(CHECK_EVERY as usize) {
+        if cancel.is_cancelled() {
+            return Err(HeapError::Cancelled);
+        }
+        count += chunk.iter().filter(|&&ty| ty == slot).count() as u64;
+    }
+    Ok(count)
+}
@@ -0,0 +1,593 @@
+use super::*;
+use crate::column::Store;
+use crate::graph::build_graph;
+use crate::progress::HeapPhase;
+use crate::testing::{HprofWriter, MiniHeap, sample};
+
+const ROOT: Option<&str> = Some("java.lang.Object");
+
+fn no_progress(_: HeapPhase, _: u64, _: u64) -> Result<(), HeapError> {
+    Ok(())
+}
+
+fn graph_of(heap: &MiniHeap) -> HeapGraph {
+    build_graph(&heap.bytes(), &Store::Ram, &no_progress, &CancellationToken::new()).unwrap()
+}
+
+fn model(g: &HeapGraph, layout: VmLayout, histogram: Option<&[HistogramEntry]>) -> ShallowModel {
+    ShallowModel::new(g, layout, histogram, &CancellationToken::new()).unwrap()
+}
+
+fn size_at(g: &HeapGraph, m: &ShallowModel, addr: u64) -> u64 {
+    m.size_of(g, g.ids.rank(addr).unwrap_or_else(|| panic!("{addr:#x} is not in the graph")))
+}
+
+fn slot_at(g: &HeapGraph, addr: u64) -> u32 {
+    g.types[g.ids.rank(addr).unwrap() as usize]
+}
+
+fn entry(name: &str, instances: u64, bytes: u64) -> HistogramEntry {
+    HistogramEntry { name: name.to_owned(), instances, bytes }
+}
+
+fn flags(flags: &str, java_major: u32) -> VmLayout {
+    VmLayout::from_vm_flags(flags, Some(java_major))
+}
+
+const SHAPES: [&str; 9] = ["A", "B", "C", "D", "E", "F", "G", "H", "Node"];
+
+/// One instance of each shape, 0x100 apart so no address gap qualifies.
+fn shapes() -> MiniHeap {
+    use BasicType::{Byte, Int, Long, Object, Short};
+    let mut heap = MiniHeap::new();
+    heap.class("A", ROOT, &[("f", Byte)])
+        .class("B", ROOT, &[("f", Long)])
+        .class("C", ROOT, &[("i", Int), ("l", Long), ("b", Byte)])
+        .class("D", Some("C"), &[("s", Short), ("o", Object)])
+        .class("E", ROOT, &[("a", Object), ("b", Object), ("c", Object)])
+        .class("F", Some("A"), &[("g", Byte)])
+        .class("G", ROOT, &[])
+        .class("H", Some("B"), &[("i", Int)])
+        .class("Node", ROOT, &[("l", Long), ("a", Object), ("b", Object)]);
+    for (i, name) in SHAPES.iter().enumerate() {
+        heap.object(0x1000 + i as u64 * 0x100, name, &[]);
+    }
+    heap
+}
+
+fn shape_sizes(layout: VmLayout) -> Vec<u64> {
+    let g = graph_of(&shapes());
+    let m = model(&g, layout, None);
+    (0..SHAPES.len() as u64).map(|i| size_at(&g, &m, 0x1000 + i * 0x100)).collect()
+}
+
+// `jcmd <pid> VM.flags` of `-Xmx1g` and `-Xmx40g` runs (the flags that matter here).
+const J8_11_1G: &str = "-XX:MaxHeapSize=1073741824 -XX:+UseCompressedClassPointers -XX:+UseCompressedOops";
+const J8_11_40G: &str = "-XX:MaxHeapSize=42949672960";
+const J21_25_1G: &str = "-XX:MaxHeapSize=1073741824 -XX:SoftMaxHeapSize=1073741824 -XX:+UseCompressedOops";
+const J21_25_40G: &str = "-XX:MaxHeapSize=42949672960 -XX:SoftMaxHeapSize=42949672960";
+/// `-Xmx40g -XX:+UseCompressedOops`: the JVM warns and turns them off.
+const J21_40G_OFF: &str = "-XX:MaxHeapSize=42949672960 -XX:SoftMaxHeapSize=42949672960 -XX:-UseCompressedOops";
+const J25_COMPACT: &str = "-XX:MaxHeapSize=1073741824 -XX:+UseCompressedOops -XX:+UseCompactObjectHeaders";
+
+#[test]
+fn layout_from_flags() {
+    // The layout of a JVM of `major` with default flags otherwise.
+    let jvm = |major: u32, layout: VmLayout| VmLayout { java_major: Some(major), ..layout };
+    let legacy = VmLayout { legacy_array_base: true, ..VmLayout::DEFAULT };
+    assert_eq!(flags(J8_11_1G, 8), jvm(8, legacy));
+    assert_eq!(flags(J8_11_1G, 11), jvm(11, legacy));
+    assert_eq!(flags(J21_25_1G, 21), jvm(21, legacy));
+    assert_eq!(flags(J21_25_1G, 25), jvm(25, VmLayout::DEFAULT));
+    assert_eq!(flags(J25_COMPACT, 25), jvm(25, VmLayout { header: 8, length_offset: 8, ..VmLayout::DEFAULT }));
+    assert_eq!(flags(&format!("{J21_25_1G} -XX:+UseG1GC"), 21), jvm(21, legacy));
+    assert_eq!(
+        flags(&format!("{J21_25_1G} -XX:ObjectAlignmentInBytes=16"), 21),
+        jvm(21, VmLayout { alignment: 16, ..legacy })
+    );
+}
+
+#[test]
+fn heaps_too_big_for_compressed_oops_print_no_flag() {
+    let jvm = |major: u32, layout: VmLayout| VmLayout { java_major: Some(major), ..layout };
+    let wide_legacy =
+        VmLayout { header: 16, ref_size: 8, length_offset: 16, legacy_array_base: true, ..VmLayout::DEFAULT };
+    assert_eq!(flags(J8_11_40G, 8), jvm(8, wide_legacy));
+    assert_eq!(flags(J8_11_40G, 11), jvm(11, wide_legacy));
+    // From JDK 15 compressed class pointers stay on without compressed oops.
+    let wide = VmLayout { ref_size: 8, ..VmLayout::DEFAULT };
+    assert_eq!(flags(J21_25_40G, 21), jvm(21, VmLayout { legacy_array_base: true, ..wide }));
+    assert_eq!(flags(J21_25_40G, 25), jvm(25, wide));
+    assert_eq!(flags(J21_40G_OFF, 21), flags(J21_25_40G, 21));
+}
+
+#[test]
+fn flags_of_an_unknown_jdk_follow_the_jdk_15_class_pointer_rule() {
+    let unknown = |f: &str| VmLayout::from_vm_flags(f, None);
+    assert_eq!(unknown(J21_25_1G), VmLayout::DEFAULT);
+    assert_eq!(unknown(J8_11_40G), VmLayout { ref_size: 8, ..VmLayout::DEFAULT });
+    let no_ccp = unknown("-XX:+UseCompressedOops -XX:-UseCompressedClassPointers");
+    assert_eq!((no_ccp.header, no_ccp.length_offset, no_ccp.java_major), (16, 16, None));
+    assert_eq!(unknown(J25_COMPACT).header, 8);
+    // The mirror table has no row for an unknown JDK.
+    let (g, addrs) = mirror_heap(&MIRROR_STATICS);
+    assert_eq!(mirror_sizes(&g, &addrs, unknown(J21_25_1G), None), expected_mirrors(&MIRROR_STATICS, 32));
+}
+
+#[test]
+fn layout_flag_edge_cases() {
+    let legacy = VmLayout { legacy_array_base: true, ..VmLayout::DEFAULT };
+    // No VM.flags (an external dump): the default layout, of an unknown JVM.
+    assert_eq!(flags("", 25), VmLayout::DEFAULT);
+    assert_eq!(flags("", 21), legacy);
+    assert_eq!(flags("-Xmx1g --add-opens=x -XX:Foo", 25), VmLayout::DEFAULT);
+    // Before JDK 15 compressed class pointers are on only when the flags say so.
+    assert_eq!(flags("-XX:+UseCompressedOops", 11).header, 16);
+    assert_eq!(flags("-XX:+UseCompressedOops -XX:+UseCompressedClassPointers", 11).header, 12);
+    assert_eq!(flags("-XX:+UseCompressedOops -XX:-UseCompressedClassPointers", 11).header, 16);
+    assert_eq!(flags("-XX:+UseCompressedOops", 15).header, 12);
+    let no_ccp = flags("-XX:+UseCompressedOops -XX:-UseCompressedClassPointers", 21);
+    assert_eq!((no_ccp.header, no_ccp.length_offset, no_ccp.ref_size), (16, 16, 4));
+    // Compact headers win over the class pointer flag; the last mention of a flag wins.
+    assert_eq!(flags("-XX:-UseCompressedClassPointers -XX:+UseCompactObjectHeaders", 25).header, 8);
+    assert_eq!(flags("-XX:-UseCompressedOops -XX:+UseCompressedOops", 21).ref_size, 4);
+    assert_eq!(flags("-XX:+UseCompressedOops -XX:-UseCompressedOops", 21).ref_size, 8);
+}
+
+#[test]
+fn object_alignment_accepts_powers_of_two_from_8_to_256() {
+    for good in [8u16, 16, 32, 64, 128, 256] {
+        assert_eq!(flags(&format!("-XX:ObjectAlignmentInBytes={good}"), 25).alignment, good);
+    }
+    for bad in ["0", "4", "7", "12", "255", "512", "abc", "", "999999999999"] {
+        assert_eq!(flags(&format!("-XX:ObjectAlignmentInBytes={bad}"), 25).alignment, 8, "{bad:?}");
+    }
+    let wide = flags("-XX:+UseCompressedOops -XX:ObjectAlignmentInBytes=256", 21);
+    assert_eq!(wide.instance_size(1), 256);
+    assert_eq!(wide.instance_size(300), 512);
+    assert_eq!(wide.array_size(BasicType::Byte, 1), 256);
+}
+
+#[test]
+fn instance_sizes_match_the_jvm() {
+    let at = |sizes: &[u64], name: &str| sizes[SHAPES.iter().position(|n| *n == name).unwrap()];
+
+    let coops = shape_sizes(flags(J21_25_1G, 21));
+    for (name, size) in
+        [("A", 16), ("B", 24), ("C", 32), ("D", 32), ("E", 24), ("F", 16), ("G", 16), ("H", 24), ("Node", 32)]
+    {
+        assert_eq!(at(&coops, name), size, "{name} with compressed oops");
+    }
+    let wide = shape_sizes(flags(J21_25_40G, 21));
+    for (name, size) in [("D", 40), ("E", 40), ("Node", 40)] {
+        assert_eq!(at(&wide, name), size, "{name} without compressed oops");
+    }
+    let compact = shape_sizes(flags(J25_COMPACT, 25));
+    for (name, size) in [("G", 8), ("B", 16), ("C", 24), ("D", 32), ("Node", 24)] {
+        assert_eq!(at(&compact, name), size, "{name} with compact headers");
+    }
+}
+
+#[test]
+fn instance_size_bounds() {
+    assert_eq!(VmLayout::DEFAULT.instance_size(0), 16);
+    assert_eq!(flags("-XX:+UseCompactObjectHeaders", 25).instance_size(0), 8);
+    assert_eq!(flags("-XX:ObjectAlignmentInBytes=16", 21).instance_size(1), 16);
+    assert_eq!(flags("-XX:ObjectAlignmentInBytes=16", 21).instance_size(8), 32);
+}
+
+#[test]
+fn array_sizes() {
+    let d = VmLayout::DEFAULT;
+    assert_eq!(d.array_size(BasicType::Long, 2), 32);
+    assert_eq!(d.array_size(BasicType::Byte, 5), 24);
+    assert_eq!(d.array_size(BasicType::Object, 3), 32);
+    let compact = flags("-XX:+UseCompactObjectHeaders", 25);
+    assert_eq!(compact.array_size(BasicType::Byte, 5), 24);
+    assert_eq!(compact.array_size(BasicType::Int, 1), 16);
+    assert_eq!(compact.array_size(BasicType::Long, 1), 24);
+    assert_eq!(flags("-XX:-UseCompressedClassPointers", 21).array_size(BasicType::Byte, 1), 32);
+    assert_eq!(flags("-XX:-UseCompressedClassPointers", 25).array_size(BasicType::Byte, 1), 24);
+    assert_eq!(flags("-XX:-UseCompressedClassPointers", 25).array_size(BasicType::Long, 1), 32);
+    // Empty arrays and the largest length fit.
+    assert_eq!(d.array_size(BasicType::Int, 0), 16);
+    assert_eq!(d.array_size(BasicType::Long, u64::from(u32::MAX)), 16 + 8 * u64::from(u32::MAX));
+}
+
+#[test]
+fn arrays_in_the_graph_follow_the_layout() {
+    let g = graph_of(&sample());
+    let m = model(&g, VmLayout::DEFAULT, None);
+    assert_eq!(size_at(&g, &m, 0x1400), 32, "long[2]");
+    assert_eq!(size_at(&g, &m, 0x1500), 32, "Object[3]");
+    assert_eq!(size_at(&g, &m, 0x1100), 24, "Node");
+    let wide = model(&g, flags("-XX:-UseCompressedOops", 21), None);
+    assert_eq!(size_at(&g, &wide, 0x1500), 40, "Object[3] with 8-byte references");
+    assert_eq!(m.size_of(&g, g.root), 0, "the super-root has no size");
+    assert_eq!(m.source, SizeSource::Layout);
+}
+
+#[test]
+fn histogram_wins_for_instances() {
+    use BasicType::Int;
+    let fields: Vec<_> = (0..23).map(|i| (format!("f{i}"), Int)).collect();
+    let fields: Vec<(&str, BasicType)> = fields.iter().map(|(n, t)| (n.as_str(), *t)).collect();
+    let mut heap = MiniHeap::new();
+    heap.class("java.lang.Thread", ROOT, &fields)
+        .class("Plain", ROOT, &[("f", Int)])
+        .object(0x1000, "java.lang.Thread", &[])
+        .object(0x1100, "Plain", &[]);
+    let g = graph_of(&heap);
+
+    let without = model(&g, VmLayout::DEFAULT, None);
+    assert_eq!(size_at(&g, &without, 0x1000), 104);
+
+    let histogram = [entry("java.lang.Thread", 13, 1560), entry("Unrelated", 2, 64)];
+    let with = model(&g, VmLayout::DEFAULT, Some(&histogram));
+    assert_eq!(with.source, SizeSource::Jvm);
+    assert_eq!(size_at(&g, &with, 0x1000), 120);
+    assert_eq!(size_at(&g, &with, 0x1100), 16, "no row: layout");
+}
+
+#[test]
+fn histogram_rows_that_do_not_divide_or_are_empty_fall_back_to_layout() {
+    let mut heap = MiniHeap::new();
+    heap.class("Plain", ROOT, &[("f", BasicType::Int)]).object(0x1000, "Plain", &[]);
+    let g = graph_of(&heap);
+    for row in [entry("Plain", 3, 100), entry("Plain", 0, 0), entry("Plain", 0, 48), entry("Plain", 2, 0)] {
+        let m = model(&g, VmLayout::DEFAULT, Some(std::slice::from_ref(&row)));
+        assert_eq!(m.source, SizeSource::Jvm);
+        assert_eq!(size_at(&g, &m, 0x1000), 16, "{row:?}");
+    }
+    // The same name twice in the histogram cannot be told apart either.
+    let twice = [entry("Plain", 1, 48), entry("Plain", 1, 64)];
+    assert_eq!(size_at(&g, &model(&g, VmLayout::DEFAULT, Some(&twice)), 0x1000), 16);
+}
+
+#[test]
+fn histogram_matches_arrays_by_display_name_but_never_sizes_them() {
+    let g = graph_of(&sample());
+    let histogram = [entry("java.lang.Object[]", 1, 4096), entry("long[]", 1, 4096), entry("Node", 4, 160)];
+    let m = model(&g, VmLayout::DEFAULT, Some(&histogram));
+    assert_eq!(size_at(&g, &m, 0x1500), 32);
+    assert_eq!(size_at(&g, &m, 0x1400), 32);
+    assert_eq!(size_at(&g, &m, 0x1100), 40, "Node from the histogram");
+}
+
+#[test]
+fn a_name_shared_by_several_classes_never_takes_a_histogram_size() {
+    use BasicType::{Int, Long};
+    let mut heap = MiniHeap::new();
+    heap.class("dup.Foo", ROOT, &[("a", Int)])
+        .object(0x1000, "dup.Foo", &[])
+        .class("dup.Foo", ROOT, &[("a", Long), ("b", Long)])
+        .object(0x1100, "dup.Foo", &[])
+        .class("Solo", ROOT, &[("a", Int)])
+        .object(0x1200, "Solo", &[]);
+    let g = graph_of(&heap);
+    let histogram = [entry("dup.Foo", 2, 80), entry("Solo", 1, 48)];
+    let m = model(&g, VmLayout::DEFAULT, Some(&histogram));
+
+    assert_eq!(size_at(&g, &m, 0x1000), 16, "first dup.Foo uses its layout");
+    assert_eq!(size_at(&g, &m, 0x1100), 32, "second dup.Foo uses its layout");
+    assert_eq!(size_at(&g, &m, 0x1200), 48, "Solo still takes its histogram size");
+    let (first, second, solo) = (slot_at(&g, 0x1000), slot_at(&g, 0x1100), slot_at(&g, 0x1200));
+    assert_ne!(first, second);
+    assert!(m.is_ambiguous(first) && m.is_ambiguous(second));
+    assert!(!m.is_ambiguous(solo));
+    let ambiguous: Vec<_> = (0..g.classes.len() as u32).filter(|&s| m.is_ambiguous(s)).collect();
+    assert_eq!(ambiguous.len(), 2, "no other class is flagged");
+    assert!(!m.is_ambiguous(u32::MAX), "slots past the classes are never ambiguous");
+}
+
+#[test]
+fn ambiguity_covers_array_classes_and_ignores_the_source() {
+    let mut heap = MiniHeap::new();
+    heap.class("[Lx.Y;", ROOT, &[]).class("[Lx/Y;", ROOT, &[]).class("Solo", ROOT, &[]);
+    let g = graph_of(&heap);
+    let m = model(&g, VmLayout::DEFAULT, None);
+    let flagged: Vec<_> =
+        g.classes.iter().enumerate().filter(|(i, _)| m.is_ambiguous(*i as u32)).map(|(_, c)| c.name.as_str()).collect();
+    assert_eq!(flagged, ["[Lx.Y;", "[Lx.Y;"]);
+}
+
+/// A dump with three classes: `java.lang.Class` (three reference fields), `Statics` (two reference
+/// statics, one null) and `Plain`; `at` is where each class object lives.
+fn class_dump() -> Vec<u8> {
+    use BasicType::Object;
+    let mut w = HprofWriter::new();
+    w.string(1, "java/lang/Class").load_class(0x100, 1).class(
+        0x100,
+        0,
+        &[(10, Object), (11, Object), (12, Object)],
+        &[],
+    );
+    w.string(2, "Statics").load_class(0x200, 2).class(0x200, 0, &[], &[(13, 0x300), (13, 0)]);
+    w.string(3, "Plain").load_class(0x300, 3).class(0x300, 0, &[], &[]);
+    w.finish()
+}
+
+#[test]
+fn class_objects_use_the_layout_even_when_the_histogram_names_them() {
+    let g = build_graph(&class_dump(), &Store::Ram, &no_progress, &CancellationToken::new()).unwrap();
+    let histogram = [entry("java.lang.Class", 3, 3000)];
+    let by_histogram = model(&g, VmLayout::DEFAULT, Some(&histogram));
+    let by_layout = model(&g, VmLayout::DEFAULT, None);
+    assert_eq!(by_histogram.source, SizeSource::Jvm);
+    for m in [&by_histogram, &by_layout] {
+        // java.lang.Class has 12 bytes of fields: 24; Statics adds two references; every mirror
+        // gets the 32 bytes assumed for an unknown JVM (the histogram row does not fit).
+        assert_eq!(size_at(&g, m, 0x100), 56, "java.lang.Class itself");
+        assert_eq!(size_at(&g, m, 0x200), 64, "two reference statics");
+        assert_eq!(size_at(&g, m, 0x300), 56, "no statics");
+    }
+    // 8-byte references: 40 for java.lang.Class, 16 for the statics, 44 injected (JDK 21), aligned.
+    let wide = model(&g, flags("-XX:-UseCompressedOops", 21), None);
+    assert_eq!(size_at(&g, &wide, 0x200), 104);
+}
+
+#[test]
+fn primitive_statics_add_to_the_class_object() {
+    let mut g = build_graph(&class_dump(), &Store::Ram, &no_progress, &CancellationToken::new()).unwrap();
+    let statics = g.classes.iter().position(|c| c.name == "Statics").unwrap();
+    g.classes[statics].static_bytes = 12;
+    let m = model(&g, VmLayout::DEFAULT, None);
+    assert_eq!(size_at(&g, &m, 0x200), 80, "24 + 12 + 2 × 4 + 32 injected = 76, aligned");
+}
+
+#[test]
+fn a_dump_without_java_lang_class_sizes_class_objects_as_empty_mirrors() {
+    let mut w = HprofWriter::new();
+    w.string(1, "Statics").load_class(0x200, 1).class(0x200, 0, &[], &[(2, 0), (2, 0), (2, 0)]);
+    let g = build_graph(&w.finish(), &Store::Ram, &no_progress, &CancellationToken::new()).unwrap();
+    let m = model(&g, VmLayout::DEFAULT, None);
+    assert_eq!(size_at(&g, &m, 0x200), 64, "16 + 3 × 4 + 32 injected = 60, aligned");
+}
+
+fn one_class_two_instances(stride: u64) -> (HeapGraph, ShallowModel) {
+    let mut heap = MiniHeap::new();
+    heap.class("A", ROOT, &[("f", BasicType::Byte)]).object(0x1000, "A", &[]).object(0x1000 + stride, "A", &[]);
+    let g = graph_of(&heap);
+    let m = model(&g, VmLayout::DEFAULT, None);
+    (g, m)
+}
+
+#[test]
+fn address_gaps_for_external_dumps() {
+    let mut heap = MiniHeap::new();
+    heap.class("A", ROOT, &[("f", BasicType::Byte)])
+        .class("Wide", ROOT, &[("f", BasicType::Long)])
+        .class("Module", ROOT, &[("f", BasicType::Long)])
+        .object(0x1000, "A", &[])
+        .object(0x1018, "A", &[])
+        .object(0x1030, "A", &[])
+        .object(0x1048, "Module", &[])
+        .object(0x2000, "Wide", &[])
+        .object(0x3000, "Wide", &[]);
+    let g = graph_of(&heap);
+    let m = model(&g, VmLayout::DEFAULT, None);
+    assert_eq!(size_at(&g, &m, 0x1000), 24, "layout says 16, the smallest gap between As is 24");
+    assert_eq!(size_at(&g, &m, 0x1030), 24, "every A takes the type's size");
+    assert_eq!(size_at(&g, &m, 0x1048), 24, "a single instance has no gap to trust: layout");
+    assert_eq!(size_at(&g, &m, 0x2000), 24, "a gap beyond layout + 64 is not an object size: layout");
+    assert_eq!(m.source, SizeSource::Layout, "one of three types agrees");
+}
+
+#[test]
+fn address_gaps_are_taken_within_layout_and_layout_plus_64() {
+    for (stride, expected) in [(8, 16), (16, 16), (24, 24), (80, 80), (88, 16)] {
+        let (g, m) = one_class_two_instances(stride);
+        assert_eq!(size_at(&g, &m, 0x1000), expected, "stride {stride}");
+        assert_eq!(
+            m.source,
+            if expected == stride && stride >= 16 { SizeSource::AddressGaps } else { SizeSource::Layout }
+        );
+    }
+}
+
+/// `used` types with two instances 16 bytes apart (the layout size), `far` types with two
+/// instances 0x1000 apart (a gap that is no object size) and `single` types with one instance.
+fn types_by_gap_use(used: usize, far: usize, single: usize) -> SizeSource {
+    let mut heap = MiniHeap::new();
+    for i in 0..used + far + single {
+        let name = format!("T{i}");
+        heap.class(&name, ROOT, &[("f", BasicType::Byte)]);
+        let base = 0x10_0000 * (i as u64 + 1);
+        heap.object(base, &name, &[]);
+        if i < used {
+            heap.object(base + 16, &name, &[]);
+        } else if i < used + far {
+            heap.object(base + 0x1000, &name, &[]);
+        }
+    }
+    let g = graph_of(&heap);
+    model(&g, VmLayout::DEFAULT, None).source
+}
+
+#[test]
+fn address_gap_source_needs_ninety_percent_of_the_types_that_have_a_gap() {
+    assert_eq!(types_by_gap_use(10, 0, 0), SizeSource::AddressGaps);
+    assert_eq!(types_by_gap_use(9, 1, 0), SizeSource::AddressGaps);
+    assert_eq!(types_by_gap_use(8, 2, 0), SizeSource::Layout);
+    // Types with one instance have no gap: they count neither for nor against.
+    assert_eq!(types_by_gap_use(9, 1, 5), SizeSource::AddressGaps);
+    assert_eq!(types_by_gap_use(8, 2, 5), SizeSource::Layout);
+    assert_eq!(types_by_gap_use(10, 0, 400), SizeSource::AddressGaps);
+    assert_eq!(types_by_gap_use(0, 0, 3), SizeSource::Layout, "no type has a gap");
+    assert_eq!(types_by_gap_use(0, 2, 0), SizeSource::Layout);
+    assert_eq!(types_by_gap_use(0, 0, 0), SizeSource::Layout, "nothing to size");
+}
+
+/// A dump of `java.lang.Class` (three reference fields, so 24 bytes of layout) and one class per
+/// entry of `statics`: `(reference statics, primitive static bytes)`. Class `i` is at `0x1000 + 0x100 × i`.
+fn mirror_heap(statics: &[(usize, u64)]) -> (HeapGraph, Vec<u64>) {
+    use BasicType::Object;
+    let mut w = HprofWriter::new();
+    w.string(1, "java/lang/Class").load_class(0x100, 1).class(
+        0x100,
+        0,
+        &[(10, Object), (11, Object), (12, Object)],
+        &[],
+    );
+    let mut addrs = Vec::new();
+    for (i, &(refs, _)) in statics.iter().enumerate() {
+        let (addr, name_id) = (0x1000 + 0x100 * i as u64, 100 + i as u64);
+        w.string(name_id, &format!("C{i}")).load_class(addr, name_id).class(addr, 0, &[], &vec![(13, 0); refs]);
+        addrs.push(addr);
+    }
+    let mut g = build_graph(&w.finish(), &Store::Ram, &no_progress, &CancellationToken::new()).unwrap();
+    for (i, &(_, bytes)) in statics.iter().enumerate() {
+        let class = g.classes.iter().position(|c| c.name == format!("C{i}")).unwrap();
+        g.classes[class].static_bytes = bytes;
+    }
+    (g, addrs)
+}
+
+/// What HotSpot would make of those classes if it added `injected` bytes to every mirror.
+fn expected_mirrors(statics: &[(usize, u64)], injected: u64) -> Vec<u64> {
+    statics.iter().map(|&(refs, bytes)| (24 + bytes + 4 * refs as u64 + injected).next_multiple_of(8)).collect()
+}
+
+const MIRROR_STATICS: [(usize, u64); 8] = [(0, 0), (3, 0), (1, 12), (6, 8), (2, 0), (9, 24), (0, 4), (4, 1)];
+
+#[test]
+fn class_objects_get_the_injected_bytes_the_histogram_implies() {
+    let (g, addrs) = mirror_heap(&MIRROR_STATICS);
+    // java.lang.Class itself and the eight classes.
+    let mut sizes = expected_mirrors(&MIRROR_STATICS, 36);
+    let java_lang_class = 24 + 36;
+    let total = sizes.iter().sum::<u64>() + (java_lang_class as u64).next_multiple_of(8);
+    sizes.insert(0, (java_lang_class as u64).next_multiple_of(8));
+    let histogram = [entry("java.lang.Class", 9, total)];
+    let m = model(&g, VmLayout::DEFAULT, Some(&histogram));
+    assert_eq!(m.source, SizeSource::Jvm);
+    assert_eq!(size_at(&g, &m, 0x100), sizes[0]);
+    for (i, &addr) in addrs.iter().enumerate() {
+        assert_eq!(size_at(&g, &m, addr), sizes[i + 1], "class {i}");
+    }
+    let spread: Vec<_> = addrs.iter().map(|&a| size_at(&g, &m, a)).collect();
+    assert!(spread.iter().any(|&s| s != spread[0]), "statics still make the mirrors differ: {spread:?}");
+}
+
+#[test]
+fn injected_bytes_are_found_for_every_multiple_of_four() {
+    let (g, addrs) = mirror_heap(&MIRROR_STATICS);
+    for injected in (0..=128).step_by(4) {
+        let total =
+            expected_mirrors(&MIRROR_STATICS, injected).iter().sum::<u64>() + (24 + injected).next_multiple_of(8);
+        let histogram = [entry("java.lang.Class", 9, total)];
+        let m = model(&g, VmLayout::DEFAULT, Some(&histogram));
+        let got: Vec<_> = addrs.iter().map(|&a| size_at(&g, &m, a)).collect();
+        assert_eq!(got, expected_mirrors(&MIRROR_STATICS, injected), "injected {injected}");
+    }
+}
+
+#[test]
+fn the_fit_tolerates_mirrors_the_dump_has_no_class_for() {
+    let (g, addrs) = mirror_heap(&MIRROR_STATICS);
+    // The histogram counts the same average mirror 90 times more often (CDS-archived classes).
+    let total = expected_mirrors(&MIRROR_STATICS, 28).iter().sum::<u64>() + (24u64 + 28).next_multiple_of(8);
+    let histogram = [entry("java.lang.Class", 9 * 91, total * 91)];
+    let m = model(&g, VmLayout::DEFAULT, Some(&histogram));
+    let got: Vec<_> = addrs.iter().map(|&a| size_at(&g, &m, a)).collect();
+    assert_eq!(got, expected_mirrors(&MIRROR_STATICS, 28));
+}
+
+fn mirror_sizes(g: &HeapGraph, addrs: &[u64], layout: VmLayout, histogram: Option<&[HistogramEntry]>) -> Vec<u64> {
+    let m = model(g, layout, histogram);
+    addrs.iter().map(|&a| size_at(g, &m, a)).collect()
+}
+
+#[test]
+fn a_histogram_row_that_cannot_describe_these_mirrors_falls_back_to_the_table() {
+    let (g, addrs) = mirror_heap(&MIRROR_STATICS);
+    let rows = [
+        entry("java.lang.Class", 9, 9 * 5000),
+        entry("java.lang.Class", 9, 9),
+        entry("java.lang.Class", 0, 0),
+        entry("Something.Else", 9, 9 * 100),
+    ];
+    for row in &rows {
+        let one = std::slice::from_ref(row);
+        // Unknown JVM: the documented 32; JDK 11 with 4-byte references: 36.
+        assert_eq!(
+            mirror_sizes(&g, &addrs, VmLayout::DEFAULT, Some(one)),
+            expected_mirrors(&MIRROR_STATICS, 32),
+            "{row:?}"
+        );
+        assert_eq!(
+            mirror_sizes(&g, &addrs, flags(J8_11_1G, 11), Some(one)),
+            expected_mirrors(&MIRROR_STATICS, 36),
+            "{row:?}"
+        );
+    }
+    let twice = [entry("java.lang.Class", 9, 9 * 60), entry("java.lang.Class", 9, 9 * 60)];
+    assert_eq!(mirror_sizes(&g, &addrs, flags(J8_11_1G, 11), Some(&twice)), expected_mirrors(&MIRROR_STATICS, 36));
+}
+
+#[test]
+fn an_accepted_fit_beats_the_table() {
+    let (g, addrs) = mirror_heap(&MIRROR_STATICS);
+    let total = expected_mirrors(&MIRROR_STATICS, 20).iter().sum::<u64>() + 48;
+    let histogram = [entry("java.lang.Class", 9, total)];
+    // JDK 11's table says 36; the histogram says 20.
+    assert_eq!(mirror_sizes(&g, &addrs, flags(J8_11_1G, 11), Some(&histogram)), expected_mirrors(&MIRROR_STATICS, 20));
+}
+
+#[test]
+fn injected_mirror_bytes_table() {
+    // (JDK, 4-byte references, 8-byte references): the measured cells, then the JDKs between.
+    for (major, narrow, wide) in
+        [(8, 36, 44), (11, 36, 44), (21, 36, 44), (22, 36, 44), (23, 36, 40), (24, 36, 40), (25, 28, 32), (26, 28, 32)]
+    {
+        assert_eq!(mirrors::injected_mirror_bytes(major, 4), Some(narrow), "JDK {major}, compressed oops");
+        assert_eq!(mirrors::injected_mirror_bytes(major, 8), Some(wide), "JDK {major}, 8-byte references");
+    }
+    assert_eq!(mirrors::injected_mirror_bytes(21, 2), None);
+}
+
+#[test]
+fn without_a_histogram_class_objects_get_the_injected_bytes_of_their_jvm() {
+    let (g, addrs) = mirror_heap(&MIRROR_STATICS);
+    // The class dump of the fixture has 4-byte references; the layout under test decides the table
+    // cell, so use layouts whose reference size matches it.
+    for (layout, injected) in [
+        (flags(J8_11_1G, 8), 36),
+        (flags(J8_11_1G, 11), 36),
+        (flags(J21_25_1G, 21), 36),
+        (flags(J21_25_1G, 23), 36),
+        (flags(J21_25_1G, 25), 28),
+        (flags(J25_COMPACT, 25), 28),
+    ] {
+        assert_eq!(mirror_sizes(&g, &addrs, layout, None), expected_mirrors(&MIRROR_STATICS, injected), "{layout:?}");
+    }
+}
+
+#[test]
+fn eight_byte_references_use_their_own_cells() {
+    let g = build_graph(&class_dump(), &Store::Ram, &no_progress, &CancellationToken::new()).unwrap();
+    // java.lang.Class = 12 + 3 × 8 = 36 → 40; Plain has no statics: 40 + injected, aligned.
+    for (major, injected) in [(8, 44), (21, 44), (23, 40), (25, 32)] {
+        let m = model(&g, flags(J21_25_40G, major), None);
+        assert_eq!(size_at(&g, &m, 0x300), (40u64 + injected).next_multiple_of(8), "JDK {major}");
+    }
+}
+
+#[test]
+fn an_unknown_jvm_uses_the_midpoint_of_the_compressed_oop_cells() {
+    let (g, addrs) = mirror_heap(&MIRROR_STATICS);
+    assert_eq!(mirror_sizes(&g, &addrs, VmLayout::DEFAULT, None), expected_mirrors(&MIRROR_STATICS, 32));
+    // Flags without any -XX: option say nothing about the JVM either.
+    assert_eq!(mirror_sizes(&g, &addrs, flags("", 21), None), expected_mirrors(&MIRROR_STATICS, 32));
+    assert_eq!(VmLayout::DEFAULT.java_major, None);
+}
+
+#[test]
+fn cancelled_token_stops_the_model() {
+    let g = graph_of(&sample());
+    let cancel = CancellationToken::new();
+    cancel.cancel();
+    assert!(matches!(ShallowModel::new(&g, VmLayout::DEFAULT, None, &cancel), Err(HeapError::Cancelled)));
+    let histogram = [entry("Node", 1, 24)];
+    assert!(matches!(ShallowModel::new(&g, VmLayout::DEFAULT, Some(&histogram), &cancel), Err(HeapError::Cancelled)));
+}
@@ -0,0 +1,66 @@
+//! Text of `java.lang.String` objects.
+
+use crate::hprof::{BasicType, HeapRecord};
+use crate::index::{HeapIndex, id_value, instance_fields};
+
+const STRING: &str = "java.lang.String";
+const LATIN1: u8 = 0;
+const UTF16: u8 = 1;
+
+/// The text of String `object`, cut to `max_chars` chars plus `…`. `None` for anything that is
+/// not a String, or one whose `value` / `coder` the dump does not hold as expected.
+///
+/// JDK 9+ keeps a `byte[]` and a `coder`; UTF-16 bytes are in the JVM's native order, which
+/// hprof copies as is, read here as little-endian (x86-64, AArch64; a big-endian JVM's text would
+/// come out garbled). JDK 8 keeps a `char[]`, whose elements hprof writes big-endian.
+pub(crate) fn string_preview(index: &HeapIndex, object: u32, max_chars: usize) -> Option<String> {
+    let slot = *index.types.get(object as usize)?;
+    if index.class_info.get(slot as usize)?.name != STRING {
+        return None;
+    }
+    let Some(HeapRecord::Instance { data, .. }) = index.record(object) else { return None };
+    let fields = instance_fields(index, slot, data);
+    let value = fields.iter().find(|f| f.name == "value" && f.ty == BasicType::Object)?;
+    let array = index.ids.rank(id_value(value.bytes).filter(|&id| id != 0)?)?;
+    let coder = fields.iter().find(|f| f.name == "coder" && f.ty == BasicType::Byte).and_then(|f| f.bytes.first());
+    let Some(HeapRecord::PrimitiveArray { ty, data, .. }) = index.record(array) else { return None };
+    match (ty, coder) {
+        (BasicType::Byte, Some(&LATIN1)) => Some(cut(data.iter().map(|&b| char::from(b)), max_chars)),
+        (BasicType::Byte, Some(&UTF16)) => utf16(data, u16::from_le_bytes, max_chars),
+        (BasicType::Char, _) => utf16(data, u16::from_be_bytes, max_chars),
+        _ => None,
+    }
+}
+
+/// The text of `char[]` `object`, as [`string_preview`] cuts it.
+pub(crate) fn char_array_text(index: &HeapIndex, object: u32, max_chars: usize) -> Option<String> {
+    match index.record(object)? {
+        HeapRecord::PrimitiveArray { ty: BasicType::Char, data, .. } => utf16(data, u16::from_be_bytes, max_chars),
+        _ => None,
+    }
+}
+
+/// An odd byte count is `None`; a lone surrogate becomes U+FFFD.
+fn utf16(bytes: &[u8], unit: fn([u8; 2]) -> u16, max_chars: usize) -> Option<String> {
+    if !bytes.len().is_multiple_of(2) {
+        return None;
+    }
+    let units = bytes.as_chunks::<2>().0.iter().map(|&pair| unit(pair));
+    Some(cut(char::decode_utf16(units).map(|c| c.unwrap_or(char::REPLACEMENT_CHARACTER)), max_chars))
+}
+
+/// Chars are whole code points, so a cut never splits a surrogate pair.
+fn cut(chars: impl Iterator<Item = char>, max_chars: usize) -> String {
+    let mut out = String::new();
+    for (i, c) in chars.enumerate() {
+        if i == max_chars {
+            out.push('…');
+            break;
+        }
+        out.push(c);
+    }
+    out
+}
+
+#[cfg(test)]
+mod tests;
@@ -0,0 +1,113 @@
+use super::*;
+use crate::hprof::{BasicType, HeapRootKind};
+use crate::test_index::{analyzed, obj};
+use crate::testing::{MiniHeap, sample_with_string};
+
+const AT: u64 = 0x1000;
+const ARRAY: u64 = 0x2000;
+const JDK9: [(&str, BasicType); 2] = [("value", BasicType::Object), ("coder", BasicType::Byte)];
+
+fn preview(heap: &MiniHeap, addr: u64) -> Option<String> {
+    let index = analyzed(heap);
+    string_preview(&index, obj(&index, addr).0, 100)
+}
+
+fn text(text: &str, latin1: bool) -> Option<String> {
+    let mut heap = MiniHeap::new();
+    heap.string(AT, text, latin1).root(HeapRootKind::JniGlobal, AT);
+    preview(&heap, AT)
+}
+
+/// A `java.lang.String` class with `fields`, and one String at [`AT`] with `values`.
+fn custom(fields: &[(&str, BasicType)], values: &[(&str, u64)], array: impl FnOnce(&mut MiniHeap)) -> Option<String> {
+    let mut heap = MiniHeap::new();
+    heap.class("java.lang.String", Some("java.lang.Object"), fields);
+    array(&mut heap);
+    heap.object(AT, "java.lang.String", values);
+    preview(&heap, AT)
+}
+
+#[test]
+fn string_preview_latin1_and_utf16() {
+    assert_eq!(preview(&sample_with_string(), 0x1800).as_deref(), Some("v42"));
+    assert_eq!(text("hé", false).as_deref(), Some("hé"));
+    assert_eq!(text("ÿ é", true).as_deref(), Some("ÿ é"));
+    assert_eq!(text("", true).as_deref(), Some(""));
+
+    let long = "a".repeat(150);
+    assert_eq!(text(&long, true), Some(format!("{}…", "a".repeat(100))));
+    assert_eq!(text(&long, false), Some(format!("{}…", "a".repeat(100))));
+    let exactly = "b".repeat(100);
+    assert_eq!(text(&exactly, true), Some(exactly));
+}
+
+#[test]
+fn truncation_keeps_surrogate_pairs_whole() {
+    let s = format!("{}😀{}", "a".repeat(99), "b".repeat(50));
+    assert_eq!(text(&s, false), Some(format!("{}😀…", "a".repeat(99))));
+    // A lone surrogate becomes the replacement character.
+    let lone: Vec<u8> = [0x68u16, 0xD800, 0x69].iter().flat_map(|u| u.to_le_bytes()).collect();
+    let got = custom(&JDK9, &[("value", ARRAY), ("coder", 1)], |h| {
+        h.prim_bytes(ARRAY, BasicType::Byte, &lone);
+    });
+    assert_eq!(got.as_deref(), Some("h\u{FFFD}i"));
+}
+
+#[test]
+fn jdk8_strings_hold_big_endian_chars() {
+    let chars: Vec<u8> = "hé".encode_utf16().flat_map(u16::to_be_bytes).collect();
+    let fields = [("value", BasicType::Object), ("hash", BasicType::Int)];
+    let got = custom(&fields, &[("value", ARRAY)], |h| {
+        h.prim_bytes(ARRAY, BasicType::Char, &chars);
+    });
+    assert_eq!(got.as_deref(), Some("hé"));
+}
+
+#[test]
+fn malformed_strings_have_no_preview() {
+    let odd = custom(&JDK9, &[("value", ARRAY), ("coder", 1)], |h| {
+        h.prim_bytes(ARRAY, BasicType::Byte, &[0x68, 0, 0x69]);
+    });
+    assert_eq!(odd, None);
+    let bad_coder = custom(&JDK9, &[("value", ARRAY), ("coder", 2)], |h| {
+        h.prim_bytes(ARRAY, BasicType::Byte, b"hi");
+    });
+    assert_eq!(bad_coder, None);
+    let no_value = custom(&[("coder", BasicType::Byte)], &[], |_| {});
+    assert_eq!(no_value, None);
+    let no_coder = custom(&[("value", BasicType::Object)], &[("value", ARRAY)], |h| {
+        h.prim_bytes(ARRAY, BasicType::Byte, b"hi");
+    });
+    assert_eq!(no_coder, None);
+    let null_value = custom(&JDK9, &[], |_| {});
+    assert_eq!(null_value, None);
+    let int_array = custom(&JDK9, &[("value", ARRAY)], |h| {
+        h.prim(ARRAY, BasicType::Int, 2);
+    });
+    assert_eq!(int_array, None);
+    let not_an_array = custom(&JDK9, &[("value", ARRAY)], |h| {
+        h.object(ARRAY, "java.lang.Object", &[]);
+    });
+    assert_eq!(not_an_array, None);
+    // `value` stored as a long instead of a reference.
+    let wrong_type = custom(&[("value", BasicType::Long), ("coder", BasicType::Byte)], &[("value", ARRAY)], |h| {
+        h.prim_bytes(ARRAY, BasicType::Byte, b"hi");
+    });
+    assert_eq!(wrong_type, None);
+}
+
+#[test]
+fn only_strings_have_a_preview() {
+    let mut heap = MiniHeap::new();
+    heap.class("Text", Some("java.lang.Object"), &JDK9).prim_bytes(ARRAY, BasicType::Byte, b"hi").object(
+        AT,
+        "Text",
+        &[("value", ARRAY)],
+    );
+    assert_eq!(preview(&heap, AT), None);
+    assert_eq!(preview(&heap, ARRAY), None);
+    let index = analyzed(&heap);
+    for bad in [index.root(), u32::MAX] {
+        assert_eq!(string_preview(&index, bad, 100), None);
+    }
+}
@@ -0,0 +1,35 @@
+//! Analyzed fixtures for the query tests.
+
+use ide_core::CancellationToken;
+
+use crate::analysis::{AnalyzeOptions, analyze};
+use crate::error::HeapError;
+use crate::index::{HeapIndex, HeapObjectRef};
+use crate::progress::HeapPhase;
+use crate::testing::MiniHeap;
+
+fn no_progress(_: HeapPhase, _: u64, _: u64) -> Result<(), HeapError> {
+    Ok(())
+}
+
+/// `heap` analyzed in memory, with layout sizes.
+pub(crate) fn analyzed(heap: &MiniHeap) -> HeapIndex {
+    analyzed_bytes(&heap.bytes())
+}
+
+pub(crate) fn analyzed_bytes(hprof: &[u8]) -> HeapIndex {
+    let dir = tempfile::tempdir().unwrap();
+    let dump = dir.path().join("dump.hprof");
+    std::fs::write(&dump, hprof).unwrap();
+    let opts = AnalyzeOptions { in_memory: true, ..AnalyzeOptions::default() };
+    analyze(&dump, dir.path(), opts, &no_progress, &CancellationToken::new()).unwrap()
+}
+
+pub(crate) fn obj(index: &HeapIndex, addr: u64) -> HeapObjectRef {
+    HeapObjectRef(index.ids.rank(addr).unwrap_or_else(|| panic!("{addr:#x} is not in the dump")))
+}
+
+/// The slot of the class named `name`.
+pub(crate) fn slot(index: &HeapIndex, name: &str) -> u32 {
+    index.classes.iter().position(|c| c.name == name).unwrap_or_else(|| panic!("no class {name}")) as u32
+}
@@ -0,0 +1,430 @@
+//! Builders for `.hprof` fixtures. They panic on misuse (unknown class name, text that does not
+//! fit the requested encoding): they are test helpers, not input handling.
+
+use std::collections::HashMap;
+use std::path::Path;
+
+use crate::hprof::{BasicType, HeapRootKind};
+
+const TAG_UTF8: u8 = 0x01;
+const TAG_LOAD_CLASS: u8 = 0x02;
+const TAG_HEAP_SEGMENT: u8 = 0x1C;
+const TAG_HEAP_DUMP_END: u8 = 0x2C;
+const TIMESTAMP_MS: u64 = 1_700_000_000_000;
+
+fn put_id(out: &mut Vec<u8>, id_size: u8, id: u64) {
+    if id_size == 4 {
+        out.extend_from_slice(&(id as u32).to_be_bytes());
+    } else {
+        out.extend_from_slice(&id.to_be_bytes());
+    }
+}
+
+/// Writes records one by one; heap sub-records are collected into a single heap-dump segment that
+/// `finish` places after the UTF8 / LOAD CLASS records, followed by HEAP DUMP END.
+pub struct HprofWriter {
+    id_size: u8,
+    top: Vec<u8>,
+    heap: Vec<u8>,
+    class_serial: u32,
+}
+
+impl Default for HprofWriter {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl HprofWriter {
+    pub fn new() -> Self {
+        Self::with_id_size(8)
+    }
+
+    /// `id_size` is 4 or 8; ids are truncated to the width.
+    pub fn with_id_size(id_size: u8) -> Self {
+        assert!(id_size == 4 || id_size == 8, "hprof ids are 4 or 8 bytes");
+        let mut top = b"JAVA PROFILE 1.0.2\0".to_vec();
+        top.extend_from_slice(&u32::from(id_size).to_be_bytes());
+        top.extend_from_slice(&TIMESTAMP_MS.to_be_bytes());
+        Self { id_size, top, heap: Vec::new(), class_serial: 0 }
+    }
+
+    fn top_record(&mut self, tag: u8, body: &[u8]) {
+        self.top.push(tag);
+        self.top.extend_from_slice(&0u32.to_be_bytes());
+        self.top.extend_from_slice(&(body.len() as u32).to_be_bytes());
+        self.top.extend_from_slice(body);
+    }
+
+    fn heap_id(&mut self, id: u64) {
+        put_id(&mut self.heap, self.id_size, id);
+    }
+
+    fn heap_u4(&mut self, v: u32) {
+        self.heap.extend_from_slice(&v.to_be_bytes());
+    }
+
+    pub fn string(&mut self, id: u64, s: &str) -> &mut Self {
+        let mut body = Vec::new();
+        put_id(&mut body, self.id_size, id);
+        body.extend_from_slice(s.as_bytes());
+        self.top_record(TAG_UTF8, &body);
+        self
+    }
+
+    pub fn load_class(&mut self, class_id: u64, name_id: u64) -> &mut Self {
+        self.class_serial += 1;
+        let mut body = self.class_serial.to_be_bytes().to_vec();
+        put_id(&mut body, self.id_size, class_id);
+        body.extend_from_slice(&0u32.to_be_bytes());
+        put_id(&mut body, self.id_size, name_id);
+        self.top_record(TAG_LOAD_CLASS, &body);
+        self
+    }
+
+    /// `statics` are `(name id, object id)` reference statics.
+    pub fn class(&mut self, id: u64, super_id: u64, fields: &[(u64, BasicType)], statics: &[(u64, u64)]) -> &mut Self {
+        self.class_with_loader(id, super_id, 0, fields, statics)
+    }
+
+    pub fn class_with_loader(
+        &mut self,
+        id: u64,
+        super_id: u64,
+        loader_id: u64,
+        fields: &[(u64, BasicType)],
+        statics: &[(u64, u64)],
+    ) -> &mut Self {
+        self.heap.push(0x20);
+        self.heap_id(id);
+        self.heap_u4(0);
+        self.heap_id(super_id);
+        self.heap_id(loader_id);
+        for _ in 0..4 {
+            self.heap_id(0); // signers, protection domain, 2 reserved
+        }
+        let instance_size: usize = fields.iter().map(|(_, ty)| ty.hprof_size(self.id_size)).sum();
+        self.heap_u4(instance_size as u32);
+        self.heap.extend_from_slice(&0u16.to_be_bytes()); // constant pool
+        self.heap.extend_from_slice(&(statics.len() as u16).to_be_bytes());
+        for &(name, object) in statics {
+            self.heap_id(name);
+            self.heap.push(BasicType::Object as u8);
+            self.heap_id(object);
+        }
+        self.heap.extend_from_slice(&(fields.len() as u16).to_be_bytes());
+        for &(name, ty) in fields {
+            self.heap_id(name);
+            self.heap.push(ty as u8);
+        }
+        self
+    }
+
+    pub fn instance(&mut self, id: u64, class_id: u64, data: &[u8]) -> &mut Self {
+        self.heap.push(0x21);
+        self.heap_id(id);
+        self.heap_u4(0);
+        self.heap_id(class_id);
+        self.heap_u4(data.len() as u32);
+        self.heap.extend_from_slice(data);
+        self
+    }
+
+    pub fn obj_array(&mut self, id: u64, class_id: u64, items: &[u64]) -> &mut Self {
+        self.heap.push(0x22);
+        self.heap_id(id);
+        self.heap_u4(0);
+        self.heap_u4(items.len() as u32);
+        self.heap_id(class_id);
+        for &item in items {
+            self.heap_id(item);
+        }
+        self
+    }
+
+    /// An array of `len` zeroed elements.
+    pub fn prim_array(&mut self, id: u64, ty: BasicType, len: u32) -> &mut Self {
+        let bytes = vec![0u8; len as usize * ty.hprof_size(self.id_size)];
+        self.prim_bytes(id, ty, &bytes)
+    }
+
+    /// An array whose elements are `bytes` (big-endian, as stored in the dump).
+    pub fn prim_bytes(&mut self, id: u64, ty: BasicType, bytes: &[u8]) -> &mut Self {
+        let size = ty.hprof_size(self.id_size);
+        assert!(ty != BasicType::Object && bytes.len().is_multiple_of(size), "{ty:?} array needs whole elements");
+        self.heap.push(0x23);
+        self.heap_id(id);
+        self.heap_u4(0);
+        self.heap_u4((bytes.len() / size) as u32);
+        self.heap.push(ty as u8);
+        self.heap.extend_from_slice(bytes);
+        self
+    }
+
+    /// Kinds that carry a thread serial get serial 1; see [`Self::root_in_thread`].
+    pub fn root(&mut self, kind: HeapRootKind, id: u64) -> &mut Self {
+        self.root_in_thread(kind, id, 1)
+    }
+
+    /// `serial` is ignored for kinds that do not carry one.
+    pub fn root_in_thread(&mut self, kind: HeapRootKind, id: u64, serial: u32) -> &mut Self {
+        let (tag, has_serial, trailing) = match kind {
+            HeapRootKind::Unknown => (0xFF, false, Trailing::None),
+            HeapRootKind::JniGlobal => (0x01, false, Trailing::Id),
+            HeapRootKind::JniLocal => (0x02, true, Trailing::U4),
+            HeapRootKind::JavaFrame => (0x03, true, Trailing::U4),
+            HeapRootKind::NativeStack => (0x04, true, Trailing::None),
+            HeapRootKind::StickyClass => (0x05, false, Trailing::None),
+            HeapRootKind::ThreadBlock => (0x06, true, Trailing::None),
+            HeapRootKind::MonitorUsed => (0x07, false, Trailing::None),
+            HeapRootKind::ThreadObject => (0x08, true, Trailing::U4),
+        };
+        self.heap.push(tag);
+        self.heap_id(id);
+        if has_serial {
+            self.heap_u4(serial);
+        }
+        match trailing {
+            Trailing::None => {}
+            Trailing::U4 => self.heap_u4(0),
+            Trailing::Id => self.heap_id(0),
+        }
+        self
+    }
+
+    pub fn finish(&self) -> Vec<u8> {
+        let mut out = self.top.clone();
+        out.push(TAG_HEAP_SEGMENT);
+        out.extend_from_slice(&0u32.to_be_bytes());
+        let heap_len = u32::try_from(self.heap.len()).expect("fixture heap segment exceeds 4 GiB");
+        out.extend_from_slice(&heap_len.to_be_bytes());
+        out.extend_from_slice(&self.heap);
+        out.push(TAG_HEAP_DUMP_END);
+        out.extend_from_slice(&0u32.to_be_bytes());
+        out.extend_from_slice(&0u32.to_be_bytes());
+        out
+    }
+}
+
+enum Trailing {
+    None,
+    U4,
+    Id,
+}
+
+struct MiniClass {
+    id: u64,
+    super_name: Option<String>,
+    fields: Vec<(String, BasicType)>,
+}
+
+/// Ids for classes, names and string backing arrays come from here, above any address a test
+/// would pick by hand.
+const GENERATED_ID_BASE: u64 = 1 << 32;
+
+/// Heap fixtures by class and field name on top of [`HprofWriter`] (id size 8).
+///
+/// Class names are written in the HotSpot form (`.` becomes `/`); look-ups use the name as given.
+pub struct MiniHeap {
+    w: HprofWriter,
+    next_id: u64,
+    classes: HashMap<String, MiniClass>,
+}
+
+impl Default for MiniHeap {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl MiniHeap {
+    /// Starts with `java.lang.Object` and `java.lang.Class` already defined.
+    pub fn new() -> Self {
+        let mut heap = Self { w: HprofWriter::new(), next_id: GENERATED_ID_BASE, classes: HashMap::new() };
+        heap.class("java.lang.Object", None, &[]);
+        heap.class("java.lang.Class", Some("java.lang.Object"), &[]);
+        heap
+    }
+
+    fn fresh_id(&mut self) -> u64 {
+        self.next_id += 1;
+        self.next_id
+    }
+
+    fn class_id(&self, name: &str) -> u64 {
+        self.classes.get(name).unwrap_or_else(|| panic!("MiniHeap: class {name} is not defined")).id
+    }
+
+    /// `super_name: None` writes super id 0.
+    pub fn class(&mut self, name: &str, super_name: Option<&str>, fields: &[(&str, BasicType)]) -> &mut Self {
+        let super_id = super_name.map_or(0, |sup| self.class_id(sup));
+        let id = self.fresh_id();
+        let name_id = self.fresh_id();
+        let mut field_ids = Vec::with_capacity(fields.len());
+        for &(field, ty) in fields {
+            let field_id = self.fresh_id();
+            self.w.string(field_id, field);
+            field_ids.push((field_id, ty));
+        }
+        self.w.string(name_id, &name.replace('.', "/"));
+        self.w.load_class(id, name_id);
+        self.w.class(id, super_id, &field_ids, &[]);
+        self.w.root(HeapRootKind::StickyClass, id);
+        let fields = fields.iter().map(|&(field, ty)| (field.to_owned(), ty)).collect();
+        self.classes.insert(name.to_owned(), MiniClass { id, super_name: super_name.map(str::to_owned), fields });
+        self
+    }
+
+    /// Fields are laid out class first, then superclass, as HotSpot does. A field missing from
+    /// `values` is 0; narrower types keep the low bytes of the value.
+    pub fn object(&mut self, addr: u64, class: &str, values: &[(&str, u64)]) -> &mut Self {
+        let class_id = self.class_id(class);
+        let mut data = Vec::new();
+        let mut cursor = Some(class);
+        while let Some(name) = cursor {
+            let c = &self.classes[name];
+            for (field, ty) in &c.fields {
+                let value = values.iter().find(|(n, _)| *n == field.as_str()).map_or(0, |&(_, v)| v);
+                let width = ty.hprof_size(8);
+                data.extend_from_slice(&value.to_be_bytes()[8 - width..]);
+            }
+            cursor = c.super_name.as_deref();
+        }
+        self.w.instance(addr, class_id, &data);
+        self
+    }
+
+    pub fn array_of(&mut self, addr: u64, class: &str, items: &[u64]) -> &mut Self {
+        let class_id = self.class_id(class);
+        self.w.obj_array(addr, class_id, items);
+        self
+    }
+
+    /// A primitive array of `len` zeroed elements.
+    pub fn prim(&mut self, addr: u64, ty: BasicType, len: u32) -> &mut Self {
+        self.w.prim_array(addr, ty, len);
+        self
+    }
+
+    /// A primitive array holding `bytes` (big-endian elements).
+    pub fn prim_bytes(&mut self, addr: u64, ty: BasicType, bytes: &[u8]) -> &mut Self {
+        self.w.prim_bytes(addr, ty, bytes);
+        self
+    }
+
+    /// A `java.lang.String` (JDK 9+ layout: `value` byte[] + `coder`) at `addr`. `latin1` stores
+    /// one byte per char (coder 0, every char must be ≤ U+00FF); otherwise UTF-16LE (coder 1).
+    pub fn string(&mut self, addr: u64, text: &str, latin1: bool) -> &mut Self {
+        if !self.classes.contains_key("java.lang.String") {
+            let fields = [("value", BasicType::Object), ("coder", BasicType::Byte)];
+            self.class("java.lang.String", Some("java.lang.Object"), &fields);
+        }
+        let bytes: Vec<u8> = if latin1 {
+            text.chars()
+                .map(|c| u8::try_from(u32::from(c)).unwrap_or_else(|_| panic!("{c:?} is not Latin-1")))
+                .collect()
+        } else {
+            text.encode_utf16().flat_map(u16::to_le_bytes).collect()
+        };
+        let array = self.fresh_id();
+        self.w.prim_bytes(array, BasicType::Byte, &bytes);
+        self.object(addr, "java.lang.String", &[("value", array), ("coder", u64::from(!latin1))])
+    }
+
+    /// Kinds that carry a thread serial get serial 1.
+    pub fn root(&mut self, kind: HeapRootKind, addr: u64) -> &mut Self {
+        self.w.root(kind, addr);
+        self
+    }
+
+    pub fn root_in_thread(&mut self, kind: HeapRootKind, addr: u64, serial: u32) -> &mut Self {
+        self.w.root_in_thread(kind, addr, serial);
+        self
+    }
+
+    pub fn bytes(&self) -> Vec<u8> {
+        self.w.finish()
+    }
+
+    pub fn write_to(&self, path: &Path) {
+        std::fs::write(path, self.bytes()).unwrap_or_else(|e| panic!("MiniHeap: writing {}: {e}", path.display()));
+    }
+}
+
+/// The heap most analysis tests share:
+///
+/// - `Holder { a: Object, b: Object }`, `Node { next: Object, v: long }`,
+///   `java.lang.ref.Reference { referent: Object, queue: Object }`,
+///   `java.lang.ref.WeakReference extends Reference`, `[Ljava.lang.Object;`
+/// - H `0x1000` Holder { a: N1, b: N2 }; N1 `0x1100` Node { next: N3 }; N2 `0x1200` Node { next: N3 };
+///   N3 `0x1300` Node { next: L }; L `0x1400` long[2]; R `0x1500` Object[] { N1, null, null };
+///   W `0x1600` WeakReference { referent: N4 }; N4 `0x1700` Node {}
+/// - roots: JNI global H, Java frame N3 (thread serial 1), JNI global W
+///
+/// R and N4 are unreachable (N4 only through the weak referent).
+pub fn sample() -> MiniHeap {
+    let mut heap = MiniHeap::new();
+    heap.class("Holder", Some("java.lang.Object"), &[("a", BasicType::Object), ("b", BasicType::Object)])
+        .class("Node", Some("java.lang.Object"), &[("next", BasicType::Object), ("v", BasicType::Long)])
+        .class(
+            "java.lang.ref.Reference",
+            Some("java.lang.Object"),
+            &[("referent", BasicType::Object), ("queue", BasicType::Object)],
+        )
+        .class("java.lang.ref.WeakReference", Some("java.lang.ref.Reference"), &[])
+        .class("[Ljava.lang.Object;", Some("java.lang.Object"), &[])
+        .object(0x1000, "Holder", &[("a", 0x1100), ("b", 0x1200)])
+        .object(0x1100, "Node", &[("next", 0x1300)])
+        .object(0x1200, "Node", &[("next", 0x1300)])
+        .object(0x1300, "Node", &[("next", 0x1400)])
+        .prim(0x1400, BasicType::Long, 2)
+        .array_of(0x1500, "[Ljava.lang.Object;", &[0x1100, 0, 0])
+        .object(0x1600, "java.lang.ref.WeakReference", &[("referent", 0x1700)])
+        .object(0x1700, "Node", &[])
+        .root(HeapRootKind::JniGlobal, 0x1000)
+        .root(HeapRootKind::JavaFrame, 0x1300)
+        .root(HeapRootKind::JniGlobal, 0x1600);
+    heap
+}
+
+/// [`sample`] plus S `0x1800` String `"v42"` (Latin-1) held as `a` of a second Holder `0x1900`,
+/// rooted as a JNI global. That Holder retains 72 bytes, as much as H.
+pub fn sample_with_string() -> MiniHeap {
+    let mut heap = sample();
+    heap.string(0x1800, "v42", true).object(0x1900, "Holder", &[("a", 0x1800)]).root(HeapRootKind::JniGlobal, 0x1900);
+    heap
+}
+
+/// [`sample`] plus two Nodes `0x1a00` and `0x1b00`, each rooted as a JNI global.
+pub fn sample_with_two_more_nodes() -> MiniHeap {
+    let mut heap = sample();
+    heap.object(0x1a00, "Node", &[])
+        .object(0x1b00, "Node", &[])
+        .root(HeapRootKind::JniGlobal, 0x1a00)
+        .root(HeapRootKind::JniGlobal, 0x1b00);
+    heap
+}
+
+/// Seeded xorshift64 numbers for reproducible random test data (the workspace has no `rand`).
+pub struct XorShift(u64);
+
+impl XorShift {
+    pub fn new(seed: u64) -> Self {
+        // Zero is xorshift's one fixed point.
+        Self(seed.max(1))
+    }
+
+    pub fn next_u64(&mut self) -> u64 {
+        let mut x = self.0;
+        x ^= x << 13;
+        x ^= x >> 7;
+        x ^= x << 17;
+        self.0 = x;
+        x
+    }
+
+    /// A number in `0..bound`; `bound` must not be 0.
+    pub fn below(&mut self, bound: u32) -> u32 {
+        (self.next_u64() % u64::from(bound)) as u32
+    }
+}
@@ -27,6 +27,8 @@ ide-search.workspace = true
 ide-lsp.workspace = true
 ide-vcs.workspace = true
 ide-dap.workspace = true
+ide-heap.workspace = true
+ide-jvm-attach.workspace = true
 ide-build.workspace = true
 ide-analysis.workspace = true
 ide-coverage.workspace = true
@@ -49,6 +51,7 @@ ide-notebook.workspace = true
 ide-extension.workspace = true
 ide-plugin.workspace = true
 ide-tools = { path = "../ide-tools" }
+ide-scaffold.workspace = true
 ide-docker = { path = "../ide-docker" }
 semver = "1"
 serde.workspace = true
@@ -22,7 +22,7 @@
 mod templates;

 use ide_extension::{Manifest, validate_plugin_id};
-use ide_host::plugins::{PluginHost, plugin_icon, read_market_url, resolve_market_href};
+use ide_host::plugins::{PluginHost, market_href_reachable_from, plugin_icon, read_market_url, resolve_market_href};
 use ide_plugin::{Limits, PluginEngine};
 use std::path::{Path, PathBuf};
 use std::process::{Command, ExitCode};
@@ -591,7 +591,7 @@ fn cmd_index(args: &[String]) -> Result<(), String> {
    let base = a.value("download-base").map(|b| b.trim_end_matches('/'));
    // Revocations and other publishers of an existing index are kept.
    let mut index = match std::fs::read_to_string(&out) {
-        Ok(text) => current_format(Index::parse(&text)?, &out.to_string_lossy())?,
+        Ok(text) => parse_index(&text, &out.to_string_lossy())?,
        Err(_) => Index { format: package::INDEX_FORMAT, ..Index::default() },
    };
    let publisher = match (a.value("publisher"), a.value("key")) {
@@ -653,6 +653,11 @@ fn cmd_index(args: &[String]) -> Result<(), String> {
                engine: package::engine_requirement(&m.engine, built_for.as_ref()),
                categories: m.categories(),
                icon: plugin_icon(&ide_extension::Package::new(m.clone(), scratch.clone()))?,
+                homepage: m.homepage.clone().filter(|h| !h.trim().is_empty()),
+                notes: m.change_notes.clone().filter(|n| !n.trim().is_empty()),
+                downloads: None,
+                featured: false,
+                suggest: Vec::new(),
            })
        })();
        let entry = entry.map_err(|e| format!("{}: {e}", path.display()))?;
@@ -717,14 +722,22 @@ fn register_key(
    Ok(publishers)
 }

-/// `index`, unless a newer `openidea-plugin` wrote it (rewriting it would drop what this one can't read).
-fn current_format(index: Index, file: &str) -> Result<Index, String> {
+/// The index in `text` (read from `file`), unless a newer `openidea-plugin` wrote it (rewriting it
+/// would drop what this one can't read). Says which entries it does not keep: no IDE offers them.
+fn parse_index(text: &str, file: &str) -> Result<Index, String> {
+    let (index, dropped) = Index::parse_dropping(text)?;
    if index.is_newer() {
        return Err(format!(
            "{file}: index format {} is newer than this tool writes; update openidea-plugin",
            index.format
        ));
    }
+    for e in &dropped {
+        eprintln!(
+            "warning: {file}: dropped {:?} {:?}: its id or version is not one a manifest may carry",
+            e.id, e.version
+        );
+    }
    Ok(index)
 }

@@ -735,14 +748,15 @@ fn cmd_merge(args: &[String]) -> Result<(), String> {
    };
    let keep = a.value("keep").ok_or("--keep is required")?;
    let out = a.value("out").unwrap_or(base);
-    let mut index =
-        current_format(Index::parse(&std::fs::read_to_string(base).map_err(|e| format!("{base}: {e}"))?)?, base)?;
+    let mut index = parse_index(&std::fs::read_to_string(base).map_err(|e| format!("{base}: {e}"))?, base)?;
    let bytes = read_market_url(other, 16 << 20)?;
-    let theirs = Index::parse(&String::from_utf8(bytes).map_err(|_| format!("{other}: not UTF-8"))?)?;
-    let mut theirs = current_format(theirs, other)?;
+    let mut theirs = parse_index(&String::from_utf8(bytes).map_err(|_| format!("{other}: not UTF-8"))?, other)?;
    // Their downloads are relative to their index, not to ours.
    for e in &mut theirs.plugins {
        e.download = resolve_market_href(other, &e.download);
+        if !market_href_reachable_from(other, &e.download) {
+            return Err(format!("{other}: the download of {} {} is not on the network", e.id, e.version));
+        }
    }
    let count = theirs.plugins.len();
    index.include(theirs, &[keep]).map_err(|e| format!("{other}: {e}"))?;
@@ -942,6 +956,47 @@ mod tests {
        assert!(err.contains("unknown publisher"), "{err}");
    }

+    /// An index read over the network is not merged when it names a download on this machine:
+    /// the merged index is a local one, which may.
+    #[test]
+    fn an_index_read_over_the_network_is_not_merged_with_local_downloads() {
+        use std::io::{BufRead, BufReader, Write};
+        let tmp = tempfile::tempdir().unwrap();
+        let base = tmp.path().join("index.json");
+        write_index(&Index { format: package::INDEX_FORMAT, ..Index::default() }, &base).unwrap();
+        let local = tmp.path().join("alice.dawn-1.0.0.oiplugin").to_string_lossy().into_owned();
+        let theirs = serde_json::json!({
+            "format": package::INDEX_FORMAT,
+            "publishers": [{ "id": "alice", "name": "Alice", "keys": [] }],
+            "plugins": [{
+                "id": "alice.dawn", "name": "Dawn", "version": "1.0.0", "publisher": "alice",
+                "download": local, "digest": "d", "engine": "^1.0",
+            }],
+        })
+        .to_string();
+        let listener = std::net::TcpListener::bind("127.0.0.1:0").unwrap();
+        let other = format!("http://127.0.0.1:{}/index.json", listener.local_addr().unwrap().port());
+        std::thread::spawn(move || {
+            for mut stream in listener.incoming().flatten() {
+                let mut reader = BufReader::new(stream.try_clone().unwrap());
+                let mut line = String::new();
+                while reader.read_line(&mut line).unwrap_or(0) > 0 && line != "\r\n" {
+                    line.clear();
+                }
+                let _ = write!(
+                    stream,
+                    "HTTP/1.1 200 OK\r\nContent-Length: {}\r\nConnection: close\r\n\r\n{theirs}",
+                    theirs.len()
+                );
+            }
+        });
+
+        let err = cmd_merge(&strings(&[base.to_str().unwrap(), &other, "--keep", "openidea"])).unwrap_err();
+        assert!(err.contains("is not on the network"), "{err}");
+        let merged = Index::parse(&std::fs::read_to_string(&base).unwrap()).unwrap();
+        assert!(merged.plugins.is_empty(), "{:?}", merged.plugins);
+    }
+
    #[test]
    fn the_index_carries_plugin_icons() {
        let tmp = tempfile::tempdir().unwrap();