Stabilize the plan-site dogfood test under parallel load

packages/content/test/plan-site-dogfood-real.test.ts ("builds a browsable plan site from refrakt's real plan/ via entityRoutes + collection") flakes under full-suite parallel load — it failed on most npm test runs across the SPEC-082 work, yet passes reliably in isolation (~5–6s). It is a heavy, real-file pipeline build (reads the whole plan/ tree, runs the full content pipeline), so under vitest's parallel worker pool it appears to hit a timeout or a shared-resource race rather than a genuine assertion failure.

This is pre-existing and unrelated to any one change — it just became visible because every SPEC-082 step ran the full suite.

main View source

Criteria completion

Tracking started Jun 3 — check back for trends.

main current done

claude/rune-contract-hardening done changeset-release/main done main done

Acceptance Criteria

Root-cause the flake: confirm whether it's a timeout (heavy build starved under parallel load) or a shared-resource race (fs / cwd / global state shared with other content tests).
Fix it so the full npm test passes deterministically — e.g. raise this test's timeout, mark its file to run in isolation / non-concurrently (pool/isolate/sequential config or describe-level concurrency off), or remove the shared-state coupling.
No reduction in coverage — the test still builds the real plan site and asserts the same browsable output.
Ten consecutive full npm test runs are green (or a reasonable flake-free bar).

Notes

Observed repeatedly during WORK-321..329: full-suite run shows this one file failing; npx vitest run <file> passes every time. Its isolation passes also served as the regression check that the bag-reading register hooks (WORK-328) and SEO untangle (WORK-329) didn't break plan registration.

Resolution

Completed: 2026-06-03

Branch: claude/rune-contract-hardening

What was done

Root-caused the flake as a starvation timeout, not a hang or a shared-resource race. Captured the failure under full-suite load: Error: Test timed out in 5000ms. The heavy real-file integration build (reads the whole plan/ tree, runs the full content pipeline) takes ~5s even in isolation — right at vitest's 5s default testTimeout — so parallel-worker CPU contention tips it over.
Fix: gave the one heavy it() a 30s timeout (6x headroom over its ~5s work) in packages/content/test/plan-site-dogfood-real.test.ts, with a comment explaining the starvation cause. No global config change.
No coverage change — same build, same assertions; only the timeout differs.

Verification

3 consecutive full vitest run passes (251/251 test files, zero timeouts), where the test previously failed on most full-suite runs.
7 additional content-package runs (18/18 files each) green — 10 green runs total, no flake.

Notes

A genuine hang would still fail well within the 30s bound, so the timeout raise doesn't mask real regressions.