test(github): remove costly navigation test polling

[ymichael]

Human comments

What was wrong

The GitHub removed-file navigation regression test repeatedly ran whole-tree named queryByRole/getByRole accessibility walks after the pull request rendered, then used another polling query after expanding the diff. Instrumentation showed that the negative removed-link lookup alone cost about 42 ms wall time / 100 ms process CPU unloaded and 1.3–1.8 seconds wall time under contention. The packages CI shard amplifies that avoidable work by running 76 package test tasks, with Turbo's default ten concurrent tasks, on 4 vCPUs. That scheduler contention was the reproduction condition; the root cause was the test's repeated accessible-name computation and polling boundaries.

This reproduced the exact failure from run 33001027428, job 98283670219: the exact unchanged case timed out at app.test.tsx:81 under its existing 5-second ceiling. Exact GitHub issue/PR searches for the run ID, job ID, signature, and file/timeout found no open match. BB run-ID/job-ID searches found only this worker, and an exact-substring signature filter found no other active or archived worker.

What changed

The existing UI regression test now:

• flushes the immediate in-memory RPC/effect chain with React's completion boundary instead of elapsed-time polling;
• checks the removed file is not wrapped in a link and the modified file keeps its workspace href;
• scopes the removed-file toggle lookup to that file's header;
• verifies expand/collapse through the toggle's accessible label and checks the rendered diff path synchronously after the click completes.

The behavioral coverage is unchanged: removed files do not navigate live workspaces, modified files remain navigable, removed diffs expand, and collapse state is exposed. There are no product, wire-protocol, CLI, guide, or documentation changes, and no timeout was changed.

How you verified

Verified fresh merge-base before task work: 7a6fe643250a93551f31441ff2c8d39aa09c638f (clean worktree; HEAD, fetched origin/main, and merge-base all exactly matched, independently confirmed by the parent worker).

Measured the exact case through Turbo with the original 5-second ceiling:

• Before, unloaded: 91–98 ms.
• Before, representative 8× local CPU saturation: 1.049 seconds.
• Before, 32×–40× saturation: 4.54–4.55 seconds.
• Before, 48× saturation: exact Test timed out in 5000ms failure, 5.843-second reported body.
• After, unloaded: 29–31 ms.
• After, four representative 8× saturation runs: 28–303 ms, all passing.
• After, the formerly red 48× saturation condition: 1.637 seconds, passing.

Validation commands:

• pnpm exec turbo run test --force --filter=bb-plugin-github — 6 files, 25 tests passed.
• pnpm exec turbo run typecheck --force --filter=bb-plugin-github — passed.
• pnpm exec turbo run build --force --filter=bb-plugin-github — passed.
• pnpm exec oxfmt plugins/github/app.test.tsx --check — passed.
• pnpm exec oxlint plugins/github/app.test.tsx — passed.
• git diff --check — passed.

Follow-up base update: the original independently verified base remains 7a6fe643250a93551f31441ff2c8d39aa09c638f. After PR #2497 resolved the unrelated inherited SDK inventory/version regression, this branch merged current main ead4c605d0fe16066baef12fa544ab14243c9f78 non-destructively. The PR still changes only plugins/github/app.test.tsx relative to main, and the structural fix is preserved unchanged.

AGENT GENERATED: by GPT-5.6-Sol