Summary

Fixes the "Theme A / Theme B" fault pair from the archived halt investigations. Both bugs surface as the same `RED_GATE_ZERO_TESTS_COLLECTED` symptom but have independent root causes — together they explain the build halt at Phase 1.2 from the user's most recent /build run (vitest reported 14 passing, orchestrator halted with "unknown runner").

Bug A — `extractTestCount` discards stdout for unknown / misclassified runners.

• `detectRunnerFromTestCmd` returns `"unknown"` whenever the resolved `testCmd` is a wrapper script (`sh scripts/test.sh`, `make test`, generic `npm test`) that doesn't contain a recognized runner keyword. The `default` branch in `extractTestCount` then returned a hardcoded zero without attempting any parser on stdout — so even when stdout had a perfectly parseable vitest JSON line or `Tests N passed` summary, the orchestrator treated it as zero.
• Same failure mode when the keyword detector misclassifies: `bun run test` resolves to runner `"bun"`, but if the underlying tool is vitest, the bun-stdout parser sees no `Ran N tests` banner and returns zero. Previously trusted unconditionally.
• Fix: `tryFallbackParsers(output)` walks every parser in priority order (vitest JSON → jest JSON → vitest stdout → jest stdout → pytest → bun → mocha → go) and returns the first non-zero recovery, tagged `source: "stdout-fallback-auto"`. Fires in two places: the `default` branch (unknown runner), and after a named-runner parse that returned zero from stdout (not from a strong JSON signal — JSON's `numTotalTests: 0` is still trusted as authoritative).

Bug B — `RUN_GEMINI_TEST_SPEC` has no commit hygiene gate.

• The test-writer phase only checked `exitCode` and let a sub-agent that returned exit 0 with a summary but zero new commits silently advance to `test_spec_done`. The missing commit surfaced one phase later at `VERIFY_RED` as the misleading `RED_GATE_ZERO_TESTS_COLLECTED` fault — wrong attribution, deferred halt, hard to diagnose. The pattern was already established for `RUN_GEMINI` (primary-impl) via `applyMutableAgentHygiene({ requireNewCommit: true, ... })`. Test-writer was the one role with a "must commit" contract but no gate.
• Fix: mirror the primary-impl hygiene capture (git snapshot + sibling baselines + parent workspace before `runRoleTask`) and wrap the result in `applyMutableAgentHygiene` with `requireNewCommit: true` (unconditional — test-writer only dispatches for `kind === "code"` phases; audit-only doesn't apply). `hygieneFailureResult` returns `exitCode: 1`, which the existing failure branch in `phase-runner.ts` catches and turns into `status: "failed"` with a clear hygiene error — attribution at the right phase, with the right cause.

Bonus — merge resolution completes upstream PR #98's intent. Upstream's subtree-aware `testCmd` inference explicitly assumed the hygiene gate this PR adds ("the test-writer hygiene gate requires a commit per iteration") but didn't implement it. Now the inference's invariant actually holds. The merge also adds a small safety improvement: the subtree inference is gated on `phaseState.status === "test_spec_done"` so a hygiene-failed phase doesn't read `HEAD~1..HEAD` against an unrelated prior commit.

Commits:

• `ab859268` — Fix A: best-effort parser fallback
• `4ce02148` — Fix B: test-writer commit hygiene gate
• `12bd20ca` — Merge `origin/main` (PR fix(build/orchestrator): fault classifier + subtree testCmd + auth preflight (3 bugs from 7-report analysis) #98's subtree-testCmd inference)

Test Coverage

8 new tests added across 2 suites, all passing alongside upstream's 213-line `resolve-test-cmd-inference` suite:

`build/orchestrator/tests/red-gate-runner-introspect.test.ts` (5 new):

• T8 unknown runner + vitest JSON in stdout → fallback recovers 14, source=`stdout-fallback-auto`
• T9 wrapper-script case: runner=`bun` + vitest JSON stdout → fallback recovers
• T10 unknown runner + vitest summary line (no JSON) → fallback recovers via stdout regex
• T11 unknown runner + empty stdout → preserves zero (no spurious recovery)
• T12 vitest JSON with `numTotalTests: 0` → trusts the zero, no fallback override

`build/orchestrator/tests/red-gate-zero-tests-collected.test.ts` (3 new):

• Bug A regression wrapper-script `testCmd: "sh scripts/test.sh"` + vitest JSON stdout → no spurious RED_GATE halt
• Bug B behavioral `applyResult` for `RUN_GEMINI_TEST_SPEC` with exit=1 (hygiene-failed) result → `status: "failed"`, not `test_spec_done`
• Bug B structural cli.ts `RUN_GEMINI_TEST_SPEC` dispatch block contains `applyMutableAgentHygiene` + `label: "test-writer"` + `requireNewCommit: true` (pins the wiring against future refactors)

Targeted suites (4 files): 71/71 pass. Full `bun test` suite: exit=0, zero failures.

Pre-Landing Review

No SQL, no LLM trust boundary violations, no user input handling. Changes are contained to TypeScript orchestrator code following the established `primaryImpl` hygiene pattern. The merge resolution preserves both my hygiene gate and upstream's subtree inference; the safety gate on `phaseState.status === "test_spec_done"` prevents the subtree inference from misattributing tests when the hygiene gate fails.

Test plan

• `bun test build/orchestrator/tests/red-gate-runner-introspect.test.ts build/orchestrator/tests/red-gate-zero-tests-collected.test.ts build/orchestrator/tests/hygiene-sibling-repos.test.ts build/orchestrator/tests/resolve-test-cmd-inference.test.ts` — 71/71 pass
• `bun test` (full free suite) — exit=0, no regressions
• Merge conflict in cli.ts resolved (PR fix(build/orchestrator): fault classifier + subtree testCmd + auth preflight (3 bugs from 7-report analysis) #98 added subtree-aware testCmd inference that depends on this PR's hygiene gate)

🤖 Generated with Claude Code