perf(ci): pr-review-receipt is dispatch-only — 96 h of fleet runner-time for a job that gates nothing (PMAT-1078)

[noahgift]

What

The pr-review-receipt job is deleted from ci.yml, and check_pr_review_wiring.sh inverts to keep its falsifier alive. scripts/ci_job_minutes.sh is added so the runner-hours figure this rests on is re-derivable rather than asserted.

This supersedes the branch's first commit, which made the job dispatch-only. A three-lane quorum and an independent re-read both found that path broken: the diff gave PR_HEAD_SHA a || github.sha fallback and left PR_NUMBER: ${{ github.event.pull_request.number }} untouched at three sites. On workflow_dispatch there is no pull-request context, so PR_NUMBER is empty and scripts/check_pr_review_arm4.sh:372 (: "${PR_NUMBER:?PR_NUMBER is required}") refuses. The job would have been reachable and non-functional — worse than absent.

The query the deletion is gated on

Recorded at the deletion site in ci.yml, not only here:

$ grep -rn 'needs:' .github/workflows/*.yml | grep pr-review-receipt      # no match
$ grep -rn 'download-artifact' .github/workflows/*.yml                    # nightly.yml:141 only (its own build assets)
$ gh api repos/paiml/aprender/branches/main/protection --jq '...'         # ["ci / gate", "workspace-test"]

The job uploaded no artifact, so no consumer could read one, and gate had already stopped reading it (PP-066 C0-5, PRQ-013, #2982 — a head-defined receipt job is a check the PR under review can edit).

What still runs, and what does not

	where
check_pr_review_receipt.sh, check_pr_review_counts.sh, check_pr_review_arm4.sh	still run — they are scripts/check_*.sh, so guard_tree.sh's derived universe (git ls-files 'scripts/check_*.sh') runs them in guard-tree, which gate needs
the PR's own receipt (Arm 4 live)	already judged from the BASE by pr-review-quorum.yml, which consumed nothing from this job
scripts/mutate-guard.sh (Arm 3), scripts/mutate_quorum_arm.sh (Arm 6), the 43-row bats table (Arm 2)	no longer run in CI — no other workflow invokes them and they are not check_*.sh, so guard_tree does not reach them. Runnable by hand; the natural content of a nightly

That loss is the 150 minutes. It is stated here and at the deletion site rather than left to be discovered.

Deviation, stated: the standing falsifier outlived its subject

R4 asserted the job's if: polarity per event. Deleting the job removes R4's subject — and R1's and R3's, since they asserted that a job invoking the guard exists and carries a job-level if:. Asserting the presence of a deleted job is not a weaker guard, it is a broken one, so the file inverts instead of dying with it:

• R1 is now the standing falsifier: no ci.yml job may invoke the receipt guard. The case table carries the old if: github.event_name == 'pull_request' wiring verbatim, asserted FAIL, and also refuses a dispatch-only job — so restoring it is a deliberate change to this rule plus a new case-table row, never an edit to ci.yml alone.
• R2 unchanged (no workflow-level paths: filter in the file that carries both required checks).
• R3 is new and is R1's other half: the guard must remain a tracked scripts/check_*.sh. Without it, renaming or untracking the guard would leave R1 holding vacuously over a family that runs nowhere.
• R4 and its if: evaluator are deleted with the job.

bash scripts/check_pr_review_wiring.sh --self-test    # SELF-TEST PASSED (10/10)
bash scripts/check_pr_review_wiring.sh                # PASS against the real ci.yml

Both polarities on every rule; R3's two rows are driven against real git trees, since the rule is about git ls-files.

The figure is now measured, and it is 3.4x larger than the claim

scripts/ci_job_minutes.sh sums per JOB (completed_at - started_at) from the jobs API: runner time, not wall time (a run with four parallel jobs occupies four runners); started_at, not created_at, so queue wait is excluded; skipped jobs contribute nothing. It refuses rather than printing 0 h when gh is absent or the window is past run retention.

$ bash scripts/ci_job_minutes.sh --since 2026-08-08 --until 2026-09-07 --event pull_request --limit 400
=== runner-hours per job - paiml/aprender, runs created 2026-08-08..2026-09-07 (400 runs, event=pull_request) ===
     hours    runs  job
     326.8     261  pr-review-receipt
     210.2     272  workspace-test
      98.6     259  guard-runner-labels
      34.3     272  ci / lint
     850.6          TOTAL

326.8 h, not 96 h - and that is a floor, not the total: the run list returned exactly --limit runs, so the window is not fully covered. The claim this PR inherited understated its own case by more than 3x. The deletion does not rest on the figure either way; it rests on the query above.

Running the script is also what exposed a defect in it: gh run list --limit N truncates with no signal, so a sum over a capped list reads as a total. Equality with the limit is now exit 2 TRUNCATED naming a larger limit to retry with (self-test 6/6; the new row drives the real script and SKIPS with a printed line, never a silent pass, where gh cannot answer). The full-window re-derivation at --limit 1200 is still running; it can only move the number up.

bash scripts/ci_job_minutes.sh --self-test    # SELF-TEST PASSED (6/6)

Other checks

bash scripts/check_roadmap_diff_additive.sh origin/main HEAD
  roadmap-diff: base=818 head=821 added=3 lifecycle=26 reserialised=0 deleted=0    PASS
bash scripts/check_receipt_gate_base_owned.sh    # exit 0
actionlint .github/workflows/ci.yml              # exit 0
python3 -c "import yaml; yaml.safe_load(open('.github/workflows/ci.yml'))"   # parses; 9 jobs; no needs: edge to the deleted job

The first commit's pmat work re-serialisation (369 entries rewritten, two legacy prose-id entries re-folded) is collapsed by python3 scripts/roadmap_trim.py.

Refs PMAT-1078, BSE-15, BSE-001 §4 wave 5

🤖 Generated with Claude Code

https://claude.ai/code/session_01X1cW8TnmNe3hgjragNjUxr

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3046 head=6d641b982aa7f796aad8c20e1dbd34154ad8341d verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.

[noahgift]

quorum-review (AD-04): NOT agreed

{
 "ticket": "PMAT-1078",
 "head": "77ba7d2fd14ad6d9c33d8f97ce5c5fc8041373ad",
 "width": 3,
 "executor": "agy",
 "agreed": false,
 "lanes": [
  {
   "lane": 1,
   "verdict": "FAIL",
   "findings": 2
  },
  {
   "lane": 2,
   "verdict": "FAIL",
   "findings": 2
  },
  {
   "lane": 3,
   "verdict": "FAIL",
   "findings": 1
  }
 ]
}

[noahgift]

Quorum: NOT AGREED — 3 lanes, 3 FAIL (agy, width 3, head 77ba7d2f; artifact posted above). pmat-merge 3046 --auto refuses at this head, so the PR is not armed. Both findings were re-read against the branch by the orchestrator rather than taken from the lanes.

1. PR_NUMBER has no dispatch fallback — the job becomes reachable but cannot run (all three lanes; confirmed here)

The diff gives PR_HEAD_SHA a fallback and leaves its neighbour alone:

+      PR_HEAD_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
       PR_NUMBER:   ${{ github.event.pull_request.number }}      # ci.yml:1860, :2185, :2337 — unchanged

On workflow_dispatch there is no pull_request context, so PR_NUMBER is empty, and scripts/check_pr_review_arm4.sh:372 is : "${PR_NUMBER:?PR_NUMBER is required (the pull request number)}" — a hard refusal on empty. So the "Still reachable / gh workflow run ci.yml --ref <branch>" section is not backed: after this change the job is dispatchable and fails at the step it exists to run. The PR body already notes arm4 "requires PR_NUMBER from CI context" and calls that pre-existing; it is pre-existing only while the job runs on pull_request, which is exactly what this PR removes.

Remedy: give PR_NUMBER a dispatch input (inputs.pr) or resolve it from github.sha, and add a self-test row asserting a dispatch run reaches arm4 with a non-empty PR number.

2. The roadmap diff is a pmat work re-serialisation — ci / gate will be RED (all three lanes; measured here)

$ bash scripts/check_roadmap_diff_additive.sh c04eda87d origin/bse-prr-dispatch-only
roadmap-diff: base=818 head=819 added=3 lifecycle=26 reserialised=369 deleted=2      # exit 1

The 2 "deleted" are legacy prose-id entries whose id strings were re-folded (Sovereign AI Integration Specification v1.1 …, Phase 2: Bayesian Examples & Documentation …), so at the YAML level nothing is lost — but G-6 (PMAT-980, #2874) refuses the churn either way, and since BSE-001 PR-A this guard runs through guard_tree.sh's universe (git ls-files 'scripts/check_*.sh') in guard-tree, which gate needs. The one real addition is PMAT-1078.

Remedy the guard itself names: python3 scripts/roadmap_trim.py, then re-run the guard.

Note on the ticket id. PMAT-1078 is this PR's, minted and InProgress. agent/R-5 also stamps it in contracts/apr-publish-cascade-v1.yaml:4; that one is the stray and is being renumbered (#2908).

[noahgift]

quorum-review (AD-04): NOT agreed

{
 "ticket": "PMAT-1078",
 "head": "3115c74e1bd77ae4bbf019f3428b579f35179e91",
 "width": 3,
 "executor": "agy",
 "agreed": false,
 "lanes": [
  {
   "lane": 1,
   "verdict": "FAIL",
   "findings": 2
  },
  {
   "lane": 2,
   "verdict": "FAIL",
   "findings": 2
  },
  {
   "lane": 3,
   "verdict": "FAIL",
   "findings": 2
  }
 ]
}

[noahgift]

quorum-review (AD-04): NOT agreed

{
 "ticket": "PMAT-1078",
 "head": "8a2f0656e969b23a0fd644341890d1e920aaeed5",
 "width": 3,
 "executor": "agy",
 "agreed": false,
 "lanes": [
  {
   "lane": 1,
   "verdict": "FAIL",
   "findings": 2
  },
  {
   "lane": 2,
   "verdict": "FAIL",
   "findings": 2
  },
  {
   "lane": 3,
   "verdict": "FAIL",
   "findings": 2
  }
 ]
}

[noahgift]

quorum-review (AD-04): NOT agreed

{
 "ticket": "PMAT-1078",
 "head": "bb2dca50b153ae9c3e425c2d9e47af2d1497e605",
 "width": 3,
 "executor": "agy",
 "agreed": false,
 "lanes": [
  {
   "lane": 1,
   "verdict": "FAIL",
   "findings": 1
  },
  {
   "lane": 2,
   "verdict": "FAIL",
   "findings": 1
  },
  {
   "lane": 3,
   "verdict": "FAIL",
   "findings": 1
  }
 ]
}

[noahgift]

quorum-review (AD-04): three PASS — agreed

{
 "ticket": "PMAT-1078",
 "head": "6d641b982aa7f796aad8c20e1dbd34154ad8341d",
 "width": 3,
 "executor": "agy",
 "agreed": true,
 "lanes": [
  {
   "lane": 1,
   "verdict": "PASS",
   "findings": 4
  },
  {
   "lane": 2,
   "verdict": "PASS",
   "findings": 0
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 4
  }
 ]
}