perf(B5): pr-review-receipt moves off pull_request to ad-hoc dispatch — 39.4% of the merge path for ONE verdict in 30 days

[noahgift]

Ticket — PMAT-1091, issue #3049 (bse-001). Claim — B5: the largest measured waste on the merge path, moved so it can actually finish.

Measured before (basis=)

machines/clean-room/ci-occupancy.sh, 30 days of aprender PR-event runs, 302-run newest-first sample. PR-event CI costs 286.56 runner-hours:

job	runs	med min	total h	share
pr-review-receipt	76	95.7	112.88	39.4%
workspace-test	80	44.6	67.99	23.7%
guard-runner-labels	58	26.6	29.02	10.1%

What those 112.88 hours bought:

cancelled n=71   110.83 h        success n=1   2.05 h
cancelled durations (min): p25=34.1  med=95.7  p75=150.3  p90=150.4  max=151.1

One verdict in thirty days. And the cancels are not all supersedes: p75 sits on the declared timeout-minutes: 150, and GitHub reports a timeout kill as cancelled, not failed — so 32 of 71 ran the full 2.5 hours holding a clean-room slot and were then killed. ~80 runner-hours a month of pure timeout on a 16-slot fleet.

PRREV-015's own comment above that timeout predicted it: "a job that times out the first time the runner is busy, and a cancelled job reads as a failure nobody can distinguish from a real one." This is that, 32 times a month.

Change

if: github.event_name == 'pull_request' → workflow_dispatch && inputs.pr != '', plus a pr input on workflow_dispatch.

gh workflow run ci.yml -f pr=<number>

Moved, not deleted. Deleting removes mutation coverage of the receipt guards. Moving lets it finish.

The trap this change would otherwise have walked into

PR_HEAD_SHA is removed from the job-level env. On workflow_dispatch there is no github.event.pull_request, so it would be the empty string — and the sweep does:

BASE=$(git merge-base origin/main "${PR_HEAD_SHA}" 2>/dev/null) || BASE=""

which tolerates an empty base. The job would have swept nothing and read green: the exact defect class this PR is motivated by, reintroduced by the fix. A first step resolves the head from the PR number, asserts a 40-char sha, asserts it is fetchable, and refuses otherwise. It runs before both PR_HEAD_SHA consumers (verified by line order: resolve at job-relative 61, consumers at 201 and 214).

Queries at the change site

All three are written into ci.yml above the if:, not left in a commit message:

1. required? gh api repos/paiml/aprender/rules/branches/main → ["gate","workspace-test"]. Not this job. (The classic branch-protection API says ci / gate — a different job. The ruleset is what blocks.)
2. needs:-ed? grep over every workflow on origin/main → 7 hits, six are comments, the seventh is its own job key. Nothing consumes it.
3. still covered? Yes, for 5% of the price: present in pr-review-quorum.yml (pull_request_target, base-owned) is 98 runs, 1.7-min median, 5.93 h, and pr-review-sign still signs on every PR. What moves is the mutation sweep over those guards.

Predicted after

~112.9 runner-hours per 30 days off the PR merge path, ~80 h of it pure timeout. Second-largest leg (workspace-test, 68.0 h) becomes the top row and the next measurement.

Writes

.github/workflows/ci.yml (one job's trigger + one new step + the input), docs/roadmaps/roadmap.yaml (+19/−0, added=1 reserialised=0 deleted=0, guard PASS). No script, no crate, no PP-066 row.

actionlint clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EeUX4ymyt8XhjXbtTuNH8L

[noahgift]

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

{
 "ticket": "PMAT-1091",
 "head": "0e8dd9df80b5beec44ce621e967f8acd4c364cff",
 "width": 3,
 "executor": "agy",
 "agreed": true,
 "lanes": [
  {
   "lane": 1,
   "verdict": "PASS",
   "findings": 0
  },
  {
   "lane": 2,
   "verdict": "PASS",
   "findings": 0
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 0
  }
 ]
}

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3052 head=a9d60adebc052dc8d253511caa12fbc1ff5ee0df 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): three PASS — agreed

{
 "ticket": "PMAT-1091",
 "head": "80ea9ce0780fc7a2ac0ae62602350b524182e774",
 "width": 3,
 "executor": "agy",
 "agreed": true,
 "lanes": [
  {
   "lane": 1,
   "verdict": "PASS",
   "findings": 0
  },
  {
   "lane": 2,
   "verdict": "PASS",
   "findings": 0
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 0
  }
 ]
}