UI benchmarks and Java evals now arm their observation before triggering an action and wait for that action's own postcondition (current file, text, caret, completion/navigation outcome, rename scope, fresh diagnostics publication) instead of accepting editor or result state left by an earlier step. Budgets and goldens are unchanged.
Co-authored-by: Mr0Bread <[email protected]>
pnpm eval drives the production workbench in Chromium against a release openidea-server and real jdtls, judging each scenario (navigation, usages, implementations, gutter markers, lenses, diagnostics, completion, docs, rename) on time to feedback, first and complete result, and checking results against evals/java/golden/. Supporting changes: data-* hooks in the usages popup, hints and rows; a shared in-page Tauri shim for smoke tests and evals; a shared perf bundle builder; language-server versions in .github/tool-versions.env shared with Language smoke; the eval agent skill.