[jonahgabriel]

Closes OMN-18664. Parent OMN-18595.

The defect

The release train's lab premise asked for a lab-pass-receipt-compose-dev-<sha> naming the EXACT dev head. runtime-rebuild-trigger.yml emits that artifact only when the merge was runtime-affecting — its receipt job is gated on published == 'true' && runtime_lane == 'dev' — so a workflow-only or scripts-only merge landing last has no receipt, at any time, and never will. Seven of the eight dev merges preceding run 35316314098 carried none, by design. The cut premise was unsatisfiable on a head that adds nothing at all to the release it blocked.

Fixed in the TRAIN, not the trigger. Emitting a receipt for a sha the lane never rebuilt would mean fabricating deployed_revision or inventing a "no rebuild needed" verdict — the train's decision relocated, not removed.

What changed

AC1 — nearest runtime-affecting ancestor. The premise resolves the newest commit R at or below the head such that no commit in R..HEAD is runtime-affecting, and asks for R's receipt. The CUT detail names both shas and the count of commits walked past. One runtime-affecting commit in that gap voids the inheritance and the premise refuses, naming that commit. The release is still cut at the head; inheriting a lane proof never moves the commit being released.

AC2 — one classifier. The predicate moved to scripts/runtime_change_classifier.py, and trigger_rebuild_on_merge.py now re-exports every name from it — the same objects, not copies, so all 87 of that file's existing tests pass unchanged. The train reads the same module. A second path list would let the train inherit a lane proof across a commit the trigger thought changed the lane. The move was required: the trigger imports click and pydantic at module scope, while the train is stdlib-plus-pyyaml on bare python3 before any project install. A test greps the train for every literal pattern and fails if one appears.

AC3 — pending is not absent. lab_receipt_pending is returned when a trigger run for the candidate is still in flight. The old detail said the sha "has not been exercised on the compose dev lane", which was false in that state and sent a lane on a four-hour investigation of a working system.

AC4 — the policy note. lab_evidence_note claimed a receipt "for every merge" and printed that on every decision row. Corrected for omnibase_infra and omnimarket; a test asserts the phrase cannot return.

AC5 — keyed on the merge commit. Receipt lookups take a default-branch commit, never a workflow run's head sha. On a squash-only repo those differ by construction, so a query keyed by the run's head is a false zero that reads exactly like a finding. A test proves the pull request head sha is never passed to the artifact query.

Fail-closed, and no override

A classifier that cannot be loaded, a branch listing that raises, an exhausted walk, and a pending probe that raises all fall back to the exact-head premise — the pre-change behaviour, which refuses rather than inherits. The workflow's new sibling checkout is continue-on-error, so a failed fetch degrades the premise instead of failing the decision, and the row says so. There is no force input, no skip input and no way to assert the premise from outside; a test reads the parser's own option strings so adding one is a red test rather than a review catch.

RED first: the new test file failed 10 of 15 against the unmodified train (the 5 passing were the grep-shaped AC2 assertions and the unchanged fallback). Green after: 17/17, and 250 tests across every file touching the train, the trigger, the clean-root validator and the incident replays. pre-commit run --all-files exit 0. ruff check, ruff format --check and mypy clean on all three scripts.

Live proof, against the real clone, the real omniclaude validator and the live artifact API. The imported classifier reproduces the trigger's own published/not-published decision on all eight of the dev commits the diagnosis tabulated: 22745243f runtime-affecting, the other seven not — exactly matching the Verify dev lane applied the redeploy => skipped readback on each of those runs.

Running the train's own plan for omnibase_infra against real dev HEAD:

| omnibase_infra | CUT | 0.38.32 | `22745243f49b` | unreleased_work_lab_proven |

The inheritance path was exercised end to end by replaying the state at 06:40Z, when dev HEAD was the workflow-only 011d3c3ca: the walk skipped 7 non-runtime commits, resolved the runtime-affecting ancestor cf1645ddc, found its receipt (live artifact query, total_count=1), downloaded it, and returned lab_receipt_fail naming the six checks that did not pass. Before this change that same head could only ever have returned lab_receipt_absent — an unactionable reason on a sha that could never carry a receipt.

Rule 24(d) — lab statement

This is a CI decision script, a shared classifier module and a policy note. It runs on the runner, not on the dev lane: nothing here is in the runtime image, no compose service changes, and no tenant-scoped path is touched. The lab pass therefore does not cover this change and none is claimed. The premise it decides was instead proven directly against the live lab-pass receipt surface and the real dev history, as above.

Not dispatched, deliberately

A workflow_dispatch of the train on this branch would run the cut job, and omnibase_infra currently verdicts CUT — that would open a real release PR on unreviewed logic. The local run above exercises the identical code path; the workflow supplies two arguments and nothing else. The first dispatched run is the scheduled 07:11 UTC one after this merges.

Concurrency

Inventoried before starting: open PRs #3741 (scripts/deploy-runtime.sh, lane omn18656), #3742, #3743, #3410, #3247, #2658 touch none of these files, and neither does the omn18662 tagger lane. CLAIM row appended via scripts/ledger_lock.py, token LCT1-4189480-4000-feaf8ee20c47-2026-09-18T07:00:30Z.

Evidence-Ticket: OMN-18664
Evidence-Source: OCC#10173

[github-actions]

⚠️ Hostile Reviewer — DEGRADED (informational)

Critical findings: 0
Major findings: 0
Total findings: 0
Models succeeded: glm-review

Note: Fewer than 2 reviewer models succeeded. Degraded results are informational (OMN-8468/OMN-8524) and do not block merge. Error: cli_review exit 2 (fewer than 2 models succeeded — partial/total outage)

---

Semantics (OMN-17492 — the model finds, thread resolution gates)

Surface	Meaning	Blocks merge?
Review threads	Per-finding, posted by the reviewer	No (informational)
Hostile Review Thread Gate	Deterministic: unresolved hostile-reviewer threads exist	Fails until resolved (not yet a required context)
degraded verdict	Fewer than 2 models succeeded (infra)	No

Powered by omniintelligence.review_pairing.cli_review — multi-model adversarial review: qwen3-review, qwen3-review-b, glm-review (OMN-8468/OMN-8524/OMN-17492)