test(PL002): make plugin-validation timeout coverage deterministic

[Jamie-BitFlight]

Problem

Valid plugin fixtures intermittently become PL002 errors because the external Claude plugin validation surface reports Claude plugin validation timed out after 3 seconds. The same failure has blocked unrelated exact-SHA gates.

Reproduced observable

At Todo 12 SHA caa111a3540b0a48076a682c95e505e13b8f00d9:

uv run pytest
ERROR [PL002] (plugin-validation): Claude plugin validation timed out after 3 seconds
8 failed, 1610 passed, 10 skipped
EXIT=1

Rerunning the six non-benchmark failures sequentially with coverage and repository addopts disabled still produced five PL002 timeout failures:

uv run pytest -o addopts= --no-cov <six exact node ids> -q
5 failed, 1 passed
EXIT=1

Todo 17 SHA 8df691fea721d794beddd434537f853308af44de independently recorded the same PL002 timeout class: 4 failed, 1607 passed, 10 skipped, exit 1.

Green baselines also exist: Todo 12 SHA df0e9c7... passed 1617 passed, 10 skipped in 96.54s; Todo 17 SHA 89d4149... passed 1611 passed, 10 skipped in 189.92s; Todo 9 SHA 3dca6c6... reports a full-suite exit 0. This establishes variability but does not establish causality.

Required outcome

Make PL002 timeout behavior deterministic and keep valid-plugin fixtures from failing solely because an external validation process crosses an undocumented test-sensitive wall-clock boundary. Preserve a real integration check for the vendor command; do not turn every fixture test into an external timing test.

Implementation-ready scope

1. Add a failing-first test at the subprocess boundary that distinguishes a successful-but-slow vendor result from a real timeout/failure. Control elapsed behavior without sleeping or depending on host load.
2. Move ordinary rule/CLI fixture coverage onto a deterministic process seam or retained result fixture so unrelated tests do not depend on the live vendor command finishing inside three seconds.
3. Keep one explicitly marked integration scenario that invokes the real vendor validator and reports unavailable/slow external state distinctly from a product PL002 finding.
4. Source any production timeout from a documented contract or explicit configuration. Do not merely increase the literal until the current machine passes.
5. Keep PL002 mapping for genuine vendor validation failures unchanged.

Acceptance

• The new negative control fails on the current three-second-dependent behavior and passes after the repair.
• Valid plugin fixtures are green in both -n 0 and repository-default xdist runs without relying on wall-clock luck.
• A controlled genuine timeout still produces exactly one PL002 result with the expected message and exit behavior.
• A controlled non-timeout vendor failure retains its existing PL002 classification.
• The explicitly marked real-vendor integration scenario has a clear skip/unavailable contract and cannot turn unrelated unit/fixture tests red.
• uv run prek run --all-files and three consecutive uv run pytest runs pass on a quiet host; retain exact exits and durations.

Related, not duplicate

• perf(tests): suite spends ~two thirds of its wall clock on subprocess startup #148 tracks broad suite subprocess/import performance and worker-count measurement. This issue owns deterministic PL002 timeout semantics and fixture isolation.

ULTRAQA classification

• flaky-test: CONFIRMED variable suite outcome across retained red and green runs; cause unverified.
• hung-command: NOT OBSERVED. Every cited command terminated with an explicit exit; the observable is a bounded three-second child-process timeout.
• stale-state: REJECTED. Evidence is keyed to exact SHAs and the live issue search was refreshed before filing.
• misleading-success: CONTROLLED. Only receipts containing command summary plus explicit exit are treated as passes; a green focused cache/action test is not substituted for full-suite status.

Manual QA evidence

80-pl002-backlog-manual-qa.txt extracts the exact timeout observable, isolated rerun result, and last-success baselines from retained terminal receipts. It reports MANUAL_QA=PASS.

Cleanup

This investigation starts no server, subprocess under test, or temporary fixture. Final process/state receipt is retained alongside the report.