Deferred low-priority polish from review-severity-binary (PRs #168–#173)

[brent-hoover]

Tracking issue for the Low / advisory findings deferred during the review-severity-binary stack (PRs #168–#173). All correctness and feature-breaking issues from those PRs were fixed and verified across multiple review rounds; the items below are non-blocking polish surfaced by roborev and the GitHub review bot, consciously deferred so the stack could land. None affect correctness or CI.

Source attribution in each item (roborev job numbers, bot reviews). Knock these off individually or in a small cleanup PR — split this issue if you'd prefer separate tracking.

Observability

• SAAdjudication analytics event lacks a cycle field. — DONE in feat(orchestrator): SA adjudication of persistent review findings #173 (commit adds cycle to the event + emit + test). Its sibling ReviewFindingPersisted carries cycle: int, which anchors the event to the ticket timeline; without it you can't tell from the log whether an adjudication fired early (first blocked round) or late (after the cap), making rate calibration harder. Add cycle: int to the event and pass it from _run_sa_adjudication. (jig/analytics/events.py; roborev 590 / feat(orchestrator): SA adjudication of persistent review findings #173 bot. Note: feat(orchestrator): SA adjudication of persistent review findings #173 is the stack tip and unmerged — this could fold into feat(orchestrator): SA adjudication of persistent review findings #173 directly instead.)

Stuck-watchdog refinement

• An unrelated pending-proposed ticket can mask a genuine deadlock. _check_stuck's has_pending_proposed treats any non-review-notable PROPOSED ticket as "alive", so a front-door issue with no dependency relationship to a stuck cycle (x → y → x) suppresses project_stuck until that unrelated issue is approved/rejected. Tighten so a proposed ticket only counts as alive when it's actually a dependency of a stuck candidate (approving it could unblock something). (jig/orchestrator.py _check_stuck; roborev 586.)

Eval-runner robustness

• analysis_dir discarded on the build-failure COMPLETED_WITH_FAILURES path. If analysis_complete arrived before the build ran, analysis_out_dir is set but the early-return manifest path returns analysis_dir=None, dropping the analysis artifact from the result. Carry it through. (jig/eval/runner.py; roborev 587.)
• Tracer exit-nonzero return site untested with tickets_failed > 0. The manifest.tracer is None site is covered; the tracer-nonzero-exit site (same fail_outcome variable) isn't exercised on the COMPLETED_WITH_FAILURES path. Patch _check_tracer_outcome to return TRACER_FAIL on a _FailWS-style run and assert COMPLETED_WITH_FAILURES. (tests/test_eval_runner.py; feat(eval): classify ticket-failure and stuck outcomes distinctly #170 bot, minor.)

Cleanup nits

• Stale docstring in _run_review_federation. Still reads "notable comments … RESOLVED but deferred via the Coordinator (DEFERRED queue)" — the pre-binary-severity wording. Actual behavior is "filed as proposed issues." (jig/orchestrator.py ~line 781; roborev 543/585.)
• Unused coord locals in disposition tests. tests/test_reviewers_disposition.py still unpacks coord from _build_stores in several TestCriticalDisposition/TestImportantDisposition cases though apply_severity_disposition no longer takes a coordinator. Harmless today (CI lints only jig/, not tests/) but ruff F841 fires if tests/ is ever linted. (roborev 585.)
• Private cross-module import from jig.worktree import _run_git. _worktree_head reaches into a private symbol; a thin public async def get_head_commit(path: Path) -> str in jig.worktree would be cleaner and avoid silent degradation (the except Exception → return None fallback) on a future rename. (jig/orchestrator.py; feat(review): multi-pass first review, delta-scoped re-review #171 bot, low-priority refactor.)
• Misleading test comment. test_sweep_fails_late_ticket_with_failed_dep comments that it verifies blame attribution to the original failed dependency, but only asserts a ticket_failed event exists — the root_failure_id attribution is actually pinned by the separate test_sweep_blame_attributes_to_original_dep. Reword the comment. (tests/test_orchestrator_failure_cascade.py; roborev 586.)

---

🤖 Generated with Claude Code