…f "exited N" (Bug M)

Closes Bug M — `RUN_GEMINI_FIX` failure handler at phase-runner.ts:825
was the lone holdout still calling `renderRoleStepFailureMessage`
directly. Same legacy holdout that Bug I (PR #107) closed for
`RUN_GEMINI_TEST_SPEC`; the test-fixer dispatcher was missed in that
pass.

Canonical incident: agnt2-prototype-prodl2-f3-f4-soak-and-backup
Phase 3.1 (Extract tier resolution into public-rpc-proxy/tier) at
2026-05-27 03:34 UTC. Codex test-fixer (gpt-5.5) ran for 101 seconds,
called `git status --short` multiple times, emitted "Files changed:"
in its output, exited 0 with no commit. applyMutableAgentHygiene's
requireNewCommit gate wrapped the success as exitCode=1 with
hygieneFailure=true and body "test fixer did not create a new
commit". The pre-fix attribution surfaced as

  `✗ Phase 3.1 ... failed: test-fixer exited 1`

— the operator had to grep the hygiene log for the actual reason. Same
shape Bug I closed for test-spec writer; same fix applies.

Fix: swap `renderRoleStepFailureMessage("test-fixer", result)` →
`geminiExitError("test-fixer", result)` at phase-runner.ts:825. The
helper extracts the first non-header line of the hygiene body when
`result.hygieneFailure === true` and falls through to the legacy
`exited N; see <log>` shape when no hygiene marker is present.

Post-fix, four distinct dispatcher families in phase-runner.ts use
geminiExitError: primary-impl, test-spec writer (post-Bug-I),
test-fixer (this PR), and dual-impl/judge. Pinned via T-M4
(`geminiExitError` call count ≥ 4) so a future regression that swaps
any of them back to renderRoleStepFailureMessage shows up at review
time.

Tests (T-M1..T-M5 in
build/orchestrator/__tests__/bug-m-test-fixer-hygiene-attribution.test.ts):

- T-M1: hygieneFailure result → next.error names the hygiene reason
  ("did not create a new commit") and includes the agent log path
  for forensics. Legacy "test-fixer exited N" form MUST NOT appear.
- T-M2: vanilla nonzero exit (no hygiene marker) → legacy "exited N"
  shape preserved. Important: when codex genuinely crashed with
  exit 2, "exited 2" is the right signal — only hygiene-converted
  failures need the new attribution.
- T-M3: static-grep guard that RUN_GEMINI_FIX uses
  geminiExitError("test-fixer", result) literally.
- T-M4: static-grep guard that phase-runner.ts has at least 4
  geminiExitError callers (count pin so a future regression that
  swaps any of them back shows up at review time).
- T-M5: negative guard — `renderRoleStepFailureMessage("test-fixer",`
  must NOT appear anywhere in phase-runner.ts.

5/5 pass.

Architectural payoff:

| Bug | Hygiene-aware attribution for | Resolved by |
|---|---|---|
| I | test-spec writer (RUN_GEMINI_TEST_SPEC) | PR #107 |
| M | test-fixer (RUN_GEMINI_FIX) | THIS PR |

After Bug M lands, all role-step dispatchers that route through
applyMutableAgentHygiene also route through geminiExitError. The
"agent did nothing" symptom now carries a one-line reason naming the
actual gate it tripped, for every role.

Note on related architectural issue: the test-fixer hygiene gate firing
on no-commit-but-output-claimed-completion is a separate Codex quality
issue (the agent reported "Files changed:" but didn't actually commit
anything). The orchestrator fix here surfaces the right error
attribution; the agent-quality fix (re-run tests after no-commit and
detect "tests now green" as a success signal) is a deeper change
deferred to a follow-up. Bug M only addresses the operator-facing
attribution.

Plan reference: ~/.claude/plans/fixing-plan-bugs-k-through-n-post-pr-108.md

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>