fix(ci): fork first-principles sweep passes the computed verdict

[chenmingwei23]

Problem / Motivation

The four ported fork reviewer lanes (design/gpt/opus/ux) complete their stranded check-runs during the finalize sweep by passing the run's own computed verdict ($conclusion / $title). The reference lane, fork-first-principles-review.yml, still passed a hardcoded neutral / "review incomplete (advisory)" at its sweep call site -- the shape #5928 fixed in the four lanes it ported the sweep to, deferred there only because that PR's spec pinned the reference lane as unmodifiable.

Why it matters

A hardcoded conclusion at the sweep site can do one of two bad things on a fork PR:

• Outvote a genuine green re-run: pr-readiness reads check-runs with fail-precedence, so a swept-to-neutral run sitting beside a real success degrades the lane's reported verdict.
• Launder a genuine BLOCK: a run whose own finalize PATCH lost both attempts falls through to the sweep, which would publish "review incomplete (advisory)" over a verdict that was actually a premise-concern BLOCK -- the one state the advisory lane exists to surface.

What changed (motivation -> approach -> change)

The sweep's job is to close the stranded-run wedge, not to decide the verdict. The run already computes the correct verdict: the case statement earlier in the same step assigns $conclusion / $title on every path (unconditional *) arm), and with no verdict computed they already hold the neutral/"review incomplete" posture. So the fix is the same one-line shape the ported lanes carry: the sweep call passes "$conclusion" "$title" instead of the hardcoded pair, with the ported lanes' justifying comment mirrored above it. A genuinely-stranded run's advisory behavior is unchanged; only the case where a real verdict exists changes -- it now propagates.

The TestForkLaneStrandedRunSweeps ratchet (added by #5928) is extended: the reference lane joins FORK_SWEEP_LANES (all three existing pins -- PR-scoped external_id at creation, sweep presence, external_id-scoped selection -- hold on it unchanged), and a new test_sweep_completes_with_the_computed_verdict asserts all five lanes pass the computed verdict and the hardcoded-neutral sweep shape cannot come back.

Tests

• New ratchet assertion is mutation-verified: red on the unfixed workflow (exactly the reference-lane param fails), green with the fix.
• test/test_ai_review_workflows.py: 168/168 pass; TestForkLaneStrandedRunSweeps 20/20 (4 tests x 5 lanes).
• isort / flake8 clean; full backend pytest run -- the only failures are environment-owned (byte-identical failure set on pristine main in the same environment, A/B verified via stash).
• Both pre-push review lanes (GPT, Opus) returned NO BLOCKING FINDINGS; the Opus lane additionally executed the extracted step body under stubbed gh/jq across all VERDICT values (BLOCK/PASS/CONCERNS/SKIPPED/empty/UNKNOWN/WITHHELD) with set -u: no unbound variable on any path, and the negative assertion does not false-positive on the legitimate POST fallback's -f conclusion="neutral".

Manual verification

Verified all five lanes now carry the identical sweep call literal (fork-design:644, fork-first-principles:683, fork-gpt:785, fork-opus:520, fork-ux:836), and that $conclusion / $title are assigned on a flat path with no early exit between the case and the sweep.

Screenshots / video

CI-workflow + backend-test change only; no visual surface.

Related Issues

Closes #5949

Checklist

• Single conventional commit
• Tests added/updated and mutation-verified
• No user-facing strings, no spec-doc surface affected

Contribution License Agreement

By submitting this pull request, I confirm my contribution is made under the terms of the repository's license.

[github-actions]

Design Review (Fable 5) — ✅ PASS

Design-level review of edc99edf3b4777348b993e8a569672f6d76df0be — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: PASS

One-line convergence to the shape the four ported lanes already carry, with a ratchet test that pins all five — sound and minimal.

[DESIGN-REVIEWED] edc99ed

[github-actions]

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of edc99edf3b4777348b993e8a569672f6d76df0be and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] edc99ed

False positive or not applicable? A repository writer can comment:
/ai-review override gpt edc99edf3b4777348b993e8a569672f6d76df0be: <one-sentence reason>

[github-actions]

Opus 4.8 Review — ✅ no blocking findings

Reviewed edc99edf3b4777348b993e8a569672f6d76df0be — this comment is updated in place on each push.

Review details

The diff is fully contained in one run: shell block. conclusion and title are assigned unconditionally by the case statement (lines 629–637, with a *) default), so they are always in scope and set -u-safe at the sweep site (line 683). Values are properly quoted; both originate from the workflow's own VERDICT output, not attacker text. The sweep is external_id-scoped to this PR, so it never touches a sibling PR's check-run, and completing a same-head stranded run with this run's verdict is consistent because both runs review the same commit. The test changes are test-only (and test/ is exempt regardless). Nothing survives falsification and I found no grounded new defect.

No findings.

[OPUS-REVIEWED] edc99ed

Verdict parsed from the review's SHA-scoped output markers for commit edc99edf3b4777348b993e8a569672f6d76df0be.

False positive or not applicable? A repository writer can comment:
/ai-review override fable edc99edf3b4777348b993e8a569672f6d76df0be: <one-sentence reason>

[github-actions]

First Principles Review (Fable 5) — ✅ PASS

Premise-level review of edc99edf3b4777348b993e8a569672f6d76df0be — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

All claims verified. Ready to emit the review.

First-Principles-Verdict: PASS

One-line change deletes the last hardcoded sweep verdict; every item is the fix, its ratchet, or its pin — nothing rides along.

What this change ships

Intent: make the reference fork lane's stranded-run sweep publish the verdict the run actually computed instead of a hardcoded neutral — a FIX (#5949).

1. Stranded check-runs now complete with the run's real verdict, not "review incomplete" — justified
2. Reference lane joins the five-lane sweep ratchet (FORK_SWEEP_LANES) — justified, but re-pins two shapes older tests already pin
3. New test pins the computed-verdict sweep and forbids the hardcoded-neutral shape, all five lanes — justified
4. Comment above the sweep mirroring the ported lanes' rationale — justified
5. Ratchet-class docstring updated to name all five lanes — justified

The fix is the deletion this lane would have asked for: it removes the one divergent spelling rather than guarding it. Grepped complete "$id" "neutral" across .github/workflows/ — 0 remaining; all five lanes carry the identical "$conclusion" "$title" literal. $conclusion/$title are assigned by an unconditional case (fork-first-principles-review.yml:629-637) with a flat path to the sweep, and the sweep's external_id scope means the verdict applies only to this PR's own runs on this head. Description and diff match exactly.

Subtractions

• Delete the two reference-lane-only tests test_fork_finalize_sweeps_stranded_check_runs (test/test_ai_review_workflows.py:659) and test_sweep_only_completes_check_runs_this_pr_created (:671): now that the reference lane is in FORK_SWEEP_LANES, the parametrized tests at :1320/:1330 assert the same literals (completing stranded check-run, the external_id select, the PR guard, the unscoped-sweep negative), and the retry/warning strings are pinned by test_the_helper_matches_the_reference_lane (:1081). Two spellings of one pin will diverge.

[FIRST-PRINCIPLES-REVIEWED] edc99ed

[NicholasRBowers]

Tier 1 auto-approve: fix (2 files). Criteria: no conflict, no requested changes, security path denylist clean, design-doc gate clean, SAST annotations clean, security checklist all-NO, AI reviewers green. Category: fix(ci) with a clear root cause — the fork first-principles finalize sweep passed a hardcoded neutral instead of the run's computed verdict; one-line call-site fix mirroring the four already-ported lanes, plus a mutation-verified ratchet test pinning all five lanes.