fix: keep test-runner retries from laundering failures

[pavelhov]

Test-runner correctness only. No packaging, OAuth, or product-code changes.

Findings

P1 — shared-process batch fail then isolate-pass was a false green

scripts/test-parallel.ts retried a failed --no-isolate batch by splitting it into per-file --isolate runs. Cross-file contamination that only shows up in a shared process could fail the batch, pass in isolation, and stamp CI green.

Contract: a shared-process batch failure is still a failure unless those files pass again in the same shared-process mode. Isolated reruns are diagnostic only and cannot mark the suite green. Isolate-lane files still retry isolated. Default workers and batch size are unchanged; this is not a “always isolate / drop coverage / lower workers” workaround.

Fix: retryQueueForFailures / retryFailuresInSameMode retry each failed item in the mode it failed in (batches stay batches). If a batch is still red after that, diagnosticIsolateRetryItems runs per-file --isolate for signal and prints a contamination warning when every file passed isolated.

P1 — overlap lock only saw --isolate

findCompetingTestRunners used pgrep -f "bun.*test --isolate". The parallel runner’s --no-isolate batches and the GUI suite’s --parallel spawn were invisible, so two full suites could overlap.

Contract: the lock covers the bun test period, not only --isolate. Wrapper scripts (bun scripts/test*.ts) are not treated as competing runners. GUI now waits on the same lock.

Fix: match bun test / bun.exe test (including --no-isolate and --parallel). gui/scripts/test.ts calls waitForExclusiveRun before spawning.

P2 — retries dropped caller flags

GUI retries spawned bun test --isolate <file> and dropped caller argv such as --timeout 1, so a timeout failure could pass on retry. The proxy runner never forwarded extra bun test flags on any spawn.

Contract: retries forward the caller’s flags unchanged, including --timeout.

Fix: partitionBunTestCliArgs splits flags from targets (and does not treat a following test path as a value for boolean flags like --bail). Proxy bunTestArgvForWorkItem and GUI retryTestArgs splice those flags into every retry spawn.

What was verified in the tree

• Planner still batches cheap files and isolates server/spawn/HOME-mutating/e2e files (scripts/test-plan.ts). Unchanged.
• Linux hermetic/systemd fixtures and tests/preload.ts HOME sandbox were not touched.
• Default worker cap remains min(4, CPU); default batch size remains 8.
• Full parallel run: 606 files, 239 isolated, 46 batches of up to 8, 4 workers. No worker-count or coverage change.

How tested

Regression tests that fail on the old behavior:

1. a shared-process batch failure stays a failure if files only pass isolated — batch retry stays --no-isolate; isolate diagnostics passing does not recover the run. A sibling test allows recovery only when the same batch passes in shared-process mode.
2. overlap lock matches bun test including --no-isolate plus a live bun test --no-isolate child that findCompetingTestRunners must see.
3. Proxy and GUI: retry argv keeps --timeout / extra flags.

Local checks on this revision (70b324d):

• bun test tests/test-runner.test.ts — 14 pass
• cd gui && bun test tests/gui-test-runner.test.ts — 5 pass
• bun run typecheck
• cd gui && bun run lint
• cd gui && bun run test — 869 pass, 139 files
• bun run privacy:scan
• bun run test:parallel — exit 0 at default workers (~109s, 606 files)

GitHub ci is the merge gate. Target is development, not main.