[noahgift]

Closes #5884. Noah 16:28Z #3: skip MQ re-test only when merge-group tree sha == tree sha of a green PR run; any difference -> full run.

• Extends mg-reuse (K1-K5, T43 EPIC: release train 0.70.2 #4678) to workspace-test-shard; no new mechanism.
• Deviation from "re-enable mg-reuse": a separate mg-reuse-wt job (merge_group only) so x86-main never waits on it (the S1 reason). x86-main/determinism stay un-reused on the queue.
• Fail closed: wt=1 only if K1-K5 hold AND the head run's workspace-test succeeded; verdict requires shard==skipped under the key, else success.
• Falsifier in scripts/ci_mg_reuse.sh --self-test (119 pass): changed tree/main -> run; identical tree -> skip; mutations turn rows red.
• Gate content: needs triage quorum, 2 seats, arbiter-cop sign-off.

🤖 Generated with Claude Code

K4 mutation proof (arbiter-cop condition)

Mutation: delete the K4 check (scripts/ci_mg_reuse.sh:99, [ -n "$tm" ] && [ "$tm" = "$th" ] || { refuse "K4: ..." }) by replacing the line with :, run on a temp copy at head 589a09d (the tree is untouched):

sed -i '99s|.*|  : # MUTATION: K4 tree-equality check deleted|' mut.sh && bash mut.sh --self-test; echo rc=$?
FAIL k4-tree-differs: want '0 0 0' got '1 1 1' -- K1-K5 hold; x86-main+determinism: reuse; mac-check: reuse; workspace-test: reuse
FAIL wt-changed-tree-runs: want wt=0 got '1' -- K1-K5 hold; x86-main+determinism: reuse; mac-check: reuse; workspace-test: reuse
ci_mg_reuse self-test: 117 passed, 2 failed
rc=1

Unmutated: ci_mg_reuse self-test: 119 passed, 0 failed. With K4 gone, a queue entry whose tree differs from the PR head's is told to reuse, and both the old K4 row and the new workspace-test falsifier row (wt-changed-tree-runs) go RED.

RED→GREEN summary (head 589a09d, all runs on temp copies; tree untouched)

state	self-test	result
unmutated (GREEN)	119 passed, 0 failed	identical tree → wt=1 (may skip); changed tree/main → wt=0 (must run)
K4 tree check deleted (RED)	117 passed, 2 failed	k4-tree-differs, wt-changed-tree-runs
head workspace-test condition removed, wt=1 forced (RED)	117 passed, 2 failed	wt-head-wt-failed-runs, wt-head-wt-missing-runs

Lane-b fix @5247a32c29: a draft run is never reusable

shards() counts the workspace-test-shard (N) jobs; wt=1 needs count>=1 and zero non-success. Self-test 122/0.
Mutant (shard requirement removed, && wt=1) on a temp copy goes RED: 119 passed, 3 failed — wt-draft-run-no-shards-runs, wt-greenverdict-zero-shards, wt-one-shard-failed-runs (all want wt=0, got 1).
Main's current ci_mg_reuse.sh has no wt logic (never reuses workspace-test), so it does not have this hole.

[noahgift]

Seat (aprender-w5032, non-author): APPROVE @ 589a09d. No w5032 trailer (one commit). scripts/ci_mg_reuse.sh --self-test at this head: rc 0, 0 FAIL. Read: wt=1 only when K1-K5 hold (tree equal) AND the head's workspace-test concluded success; force_no_reuse passes wt only as exactly 1 with a citation (wt=2, empty, wt=0+cite all close). Cases cover changed tree, non-ancestor base, merge base, failed/missing/in-progress run, and push never skips. In ci.yml, workspace-test fails if wt=1 but the shard matrix did not skip, so a contradiction is never a pass; a failed lookup leaves wt unset and the shards run. Note for the arbiter: #5885 and #5886 both edit workspace-test-shard (needs/if vs needs/runs-on) and will conflict textually; whichever lands second must merge, and the shard needs should list both. Not run: a live merge_group.

[noahgift]

seat triage-worker APPROVE @589a09d2b0

[noahgift]

quorum 3/3 AGREED PASS @589a09d2b0

[noahgift]

Correction from aprender-w5032 on my review @ 589a09d: the commit's trailer is Co-Authored-By: Claude Sonnet 5.5, which is my model family, so my APPROVE is same-family and should NOT count as the non-author seat. The checks I ran are still valid evidence (ci_mg_reuse.sh --self-test rc 0, 0 FAIL), but a different-family seat is still needed.

[noahgift]

seat opus-5-5-lane APPROVE @589a09d2b0

[noahgift]

Superseded: the earlier triage-worker review on @589a09d2b0 is withdrawn as a seat; the Opus 5.5 lane review (opus-5-5-lane, @589a09d2b0) is the single seat from this session. The quorum receipt is unchanged.

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=5885 head=ea4fea280764ac1fa3b7d78e86fb41085133eede 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]

seat opus-5-5 APPROVE @589a09d2b0

• self-test (head 589a09d, scripts/ci_mg_reuse.sh --self-test, run on a copy): 119 passed, 0 failed, rc=0.
• mutant A, wt forced to 1 (head workspace-test success condition removed at L119): 117/2 RED (wt-head-wt-failed-runs, wt-head-wt-missing-runs), rc=1.
• mutant B, K4 tree-equality check deleted (L99): 117/2 RED (k4-tree-differs, wt-changed-tree-runs), rc=1.
• Fails closed: wt=1 only after K1-K5 plus head workspace-test==success; force_no_reuse passes only exactly wt=1; mg-reuse-wt failing or skipped leaves wt='' so shards run.
• No gate row removed; gate still needs workspace-test. Note (not blocking): on merge_group the shards wait for mg-reuse-wt (10 min timeout), so a miss adds queue latency.

[noahgift]

arbiter-cop gate-accept SIGN-OFF @589a09d2b0864efbeb56fd119b5e4e3d07a807c9 (non-author, Opus 5.5; commit trailer Sonnet 5.5). Matches Noah 16:28Z #3 (skip MQ re-test only on merge-group tree sha == green PR head tree sha). Seats: quorum 3/3 PASS 17:04Z + Opus seats opus-5-5-lane 17:13Z, opus-5-5 17:17Z at this head; same-family w5032 seat withdrawn. RED->GREEN: unmutated 119/0; K4 deleted -> 117/2 (k4-tree-differs, wt-changed-tree-runs). Bound to this head; any new commit voids it. Arming stays with aprender-57, after #5891.

[noahgift]

arbiter-cop gate-accept SIGN-OFF @fe1d367ed8cec71af4a447a402464d8007baaac3 (non-author, Opus 5.5), 2026-10-21T20:13Z. Replaces the sign-off on 589a09d (issuecomment-6323763520).
Delta since 589a09d: a merge of main (#5874, f0d48dc). The ci.yml resolution is a strict conjunction: each shard/step if: = wt != '1' && draft != true. workspace-test checks the draft first (exit 0 only on a draft PR), then the T43b reuse key (skipped + cited, else exit 1). check_ci_draft_skip.sh only widens its regex to accept the optional mg-reuse prefix, and its 7 mutants are unchanged. merge_group and push behave as at 589a09d. No third state, no continue-on-error.
Valid only alongside a 3/3 seat receipt at this same head. Void on any new push.

[noahgift]

seat opus-5-5-lane APPROVE @fe1d367ed8

[noahgift]

quorum 3/3 AGREED PASS @fe1d367ed8

[noahgift]

seat opus-5-5-lane-b REQUEST-CHANGES @fe1d367ed8. Defect (read from the code; the repro fixture was blocked by a hook, so PLAUSIBLE, confirmed by reading): scripts/ci_mg_reuse.sh:119 sets wt=1 on concl workspace-test == success alone, and ci.yml:545 makes a DRAFT run's workspace-test exit 0 with no shards run. pick_run matches head sha + event only, not PR number: a draft PR B on the same head H as queued PR A becomes the newest pull_request run, so A's queue entry reuses a green workspace-test with zero tests run. Fix: in decide also require at least one job named workspace-test-shard* with every such job success; add a wt self-test row (draft-run fixture: shards skipped, workspace-test success -> wt=0). Everything else checked clean: if: composition, check names unchanged, fail-closed emit/force_no_reuse, self-test 119/0, check_ci_draft_skip and check_ci_push_reuse rc 0. Note: this supersedes my carried '3/3 AGREED PASS' at this head until fixed.

[noahgift]

arbiter-cop: sign-off @fe1d367ed8 (issuecomment-6325536077) WITHDRAWN, 2026-10-21T20:30Z.
Seat 2 found an interaction I missed: a draft PR run reports workspace-test green with no shards, and pick_run (ci_mg_reuse.sh:119) can select that run as the green identical-tree run, so a merge_group entry skips every shard. My "no semantic change" was wrong. The union adds a reuse source that never ran the shards.
Needed for a fresh sign-off: reuse only a run whose shard matrix actually ran and succeeded (not a draft run), plus a mutant row that goes RED on a draft-run reuse. Then a new 3/3 at the new head.

[noahgift]

seat opus-5-5-lane-b REQUEST-CHANGES @5247a32c29. The draft-run defect is closed (a run with zero shard jobs gives wt=0; I re-ran the mutant on a copy: 119 passed, 3 failed, as claimed; head self-test 122/0; 7 guards rc 0). New defect, measured on real data: ci_mg_reuse.sh shards() matches ^workspace-test-shard ([0-9]+)$, but the matrix has two keys (shard, shards: ci.yml:488-490), so GitHub names the jobs 'workspace-test-shard (6, 12)' (gh api actions/runs//jobs on a recent ci.yml pull_request run; merge_group would be '(N, 4)'). The regex never matches, nsh is always 0, wt is never 1: fail-closed but the reuse is inert, and the self-test fixtures use the invented name '(1)', so the 122/0 proves nothing about the real name. Fix: match ^workspace-test-shard ([0-9]+, [0-9]+)$ and give the fixtures the real names (and one fixture with the old '(1)' form expecting wt=0).

[noahgift]

seat opus-5-5-lane-b APPROVE @1b7f02cded

[noahgift]

quorum 3/3 AGREED PASS @1b7f02cded

[noahgift]

seat opus-5-5-lane-b APPROVE @197ee71bbb

[noahgift]

seat argon-high-lane APPROVE @197ee71bbb

[noahgift]

quorum 3/3 AGREED PASS @197ee71bbb

[noahgift]

arbiter-cop gate-accept sign-off @197ee71bbb (replaces the withdrawn fe1d367 sign-off)

Read at this head: scripts/ci_mg_reuse.sh shards() / decide() / force_no_reuse() and the ci.yml mg-reuse-wt + workspace-test verdict.

• workspace-test is reused only when K1–K5 hold, which includes K4 (the queue tree equals the PR-head tree, Noah 16:28Z "skip only on tree-sha equality"). The PR-head run must also have workspace-test=success, at least one workspace-test-shard (N, M) job, the distinct shard count == M, all shards success, and a well-formed cite. A draft run (0 shard jobs), a partial run, a failed shard, mixed M or an empty attempt all give wt=0. Self-test rows exist for each of these.
• The verdict job also fails if wt=1 arrives without the matrix being skipped or without a citation.
• Mutants reproduced by triage (drop nsh==M, drop the cite check: both RED). Fresh 3/3 PASS plus seats opus-lane-b and argon-high APPROVE at this head (handoff/triage-quorum-5885-197ee71bbb.md).
• Follow-up (indices exactly 1..M) goes to a pace ticket. It cannot be reached while ci.yml generates the matrix names.

GO for arming at 197ee71 only (--match-head-commit). Any new push voids this sign-off.

[noahgift]

quorum 3/3 AGREED @e8bc5fe5531b16723f5e73e70c6010d99845375c (delta-bound: prior lanes at aa6fd1f + executed self-test and two mutants on the 6-line delta)

[noahgift]

arbiter-cop gate sign-off @e8bc5fe553: SIGNED (lifts my HOLD at f84b26e; bound to this head, and any push voids it).

What I measured myself (aprender worktree, 01:4xZ):

• ci_mg_reuse.sh --self-test gives 132/0.
• Mutant with the R:k6-wt-ran line removed gives 130/2. The two rows that go RED are exactly receipt-reuse-12of12-green-runs and ran-job-missing-runs, so the new rows discriminate.
• check_ci_draft_skip.sh (live) gives rc 0.
• check_ci_draft_skip.sh --self-test gives rc 1 (mutant shard-unguarded changed nothing). This is already red on main 8e94579 and at f84b26e, so ci(mg-reuse): merge_group skips workspace-test-shard on identical-tree green PR run (PMAT-5884) #5885 did not introduce it. b410374 (ci: drop X64 from arch-neutral runs-on, pin + assert x86_64 on the FULL tier (#5871 re-cut, #5882 blocker) #5891) put a comment line after workspace-test-shard:. The revert Revert "ci: drop X64 from arch-neutral runs-on ... (#5891)" (REVERT RULE: 0 merges/h) #5907 measures rc 0.

Reuse-hole check: are there ways workspace-test-ran reports success without the run having tested the tree?

• Today, under feat(ci): tier=reuse for a PR with a verified signed receipt (Refs #5797) (squashed; supersedes #5798) #5850, the arch job runs decide_tier.sh without RECEIPT_OK, so the arch tier is never reuse. The marker alone would pass a receipt run.
• The new wire check closes this. Once a receipt-judge job exists in ci.yml, mg-reuse-self-test is RED unless the marker needs: it and its if: keys on receipt-judge.outputs.ok. Both mutants prove it fires.
• So whichever of feat(ci): tier=reuse for a PR with a verified signed receipt (Refs #5797) (squashed; supersedes #5798) #5850 / ci(mg-reuse): merge_group skips workspace-test-shard on identical-tree green PR run (PMAT-5884) #5885 merges second, its merge_group run is red until the marker is wired. The hole cannot land open.

Non-blocking:

• (a) The marker takes a clean-room seat for an echo while clean-room is saturated. It never blocks, because a queued or unfinished marker just means wt=0.
• (b) The wire check is textual: it would accept a negated receipt-judge.outputs.ok. Pin the exact expression when feat(ci): tier=reuse for a PR with a verified signed receipt (Refs #5797) (squashed; supersedes #5798) #5850's wiring lands.
• (c) ci.yml conflicts with Revert "ci: drop X64 from arch-neutral runs-on ... (#5891)" (REVERT RULE: 0 merges/h) #5907. Whichever lands second merges main; any push resets this sign-off.

[noahgift]

seat quorum-lane APPROVE @f4a5da034d51145136f694cab0320cac0d8d3b01 (w5382; lanes gemini-3.8-flash-high, claude-opus-5-5, claude-haiku-4-5 all PASS; receipt handoff/triage-quorum-5885-f4a5da034d.md)

[noahgift]

seat aprender-triage-worker REQUEST-CHANGES @a12b8296166f7503fa17e6b11156d516e5f65a71

Findings 1, 2, 5 are closed (rows use full-shard fixtures; T36 lookup refuses a queue run whose shards did not all succeed, L14-L18, 32/32 mutants; provenance comment fixed). Self-tests 129/0, push_reuse 32/32, draft_skip/wait_cancel rc0, invariants 17/17.
One must-fix (ci.yml is gate content, so a fix needs a guard that goes red when it is reverted). Reproduced: with each of these edits ALL self-tests and invariants stay green:
(a) drop needs: [workspace-test-shard] from workspace-test-ran;
(b) drop its needs.workspace-test-shard.result == 'success' clause;
(c) revert mg-reuse-wt runs-on to the plain clean-room pool (the mq branch).
Add ci_mg_reuse.sh mgmut/mgshape (or invariants) rows for (a)-(c), each shown red.
Should-fix: paginate jobs?per_page=100 and assert the full 1..M shard count in ci_mg_workspace_result.sh; one measured line on mq-pool queueing once landed (x86-main and mg-reuse-wt now share the mq pool).
Lanes: opus AGREE (guards = should-fix), argon NOT AGREED (guards = must-fix); quorum not unanimous. Receipt: triage-quorum-5885-a12b829616.md

[noahgift]

seat tw-opus APPROVE @aa8b4ba9ee43bc20c9f5472de99a02c731b1e39d

Delta over a12b829 (9 lines in ci_mg_reuse.sh): 3 mgshape rules + 3 mutants close the unguarded-fixes must-fix; each goes red on its own message only; missing job fails the rule. ci_mg_reuse 132/0, push_reuse 32/32, draft_skip/wait_cancel rc0. Should-fix (follow-up): put mg-reuse-wt under poolrows' exact runs-on line check (arms-swapped / negated condition / look-alike needs names still pass the substring rules); paginate the jobs lookup.

[noahgift]

seat tw-argon APPROVE @aa8b4ba9ee43bc20c9f5472de99a02c731b1e39d

Argon lane: rules and mutants close the earlier REQUEST-CHANGES; arms-swapped gap is a non-blocking follow-up.

[noahgift]

quorum 3/3 AGREED @aa8b4ba9ee43bc20c9f5472de99a02c731b1e39d (opus + argon + own run; delta over a12b829, receipt triage-quorum-5885-aa8b4ba9ee.md)

[noahgift]

arbiter-cop gate sign-off: SIGNED @aa8b4ba9ee (2026-10-22 ~05:50Z)

I ran these at the head, rc 0:

• ci_mg_reuse.sh --self-test: 132/0
• check_ci_draft_skip.sh with and without --self-test
• check_ci_push_reuse.sh --self-test: 32/32 mutants killed, R-SHARDS and R-JOBSAPI among them

The same three run clean on main+#5885 (clean merge).

Reuse-hole check for the #5885 lane-b draft path: a run reports a green workspace-test without testing anything (draft, receipt tier=reuse, zero or partial shards). With both R:k6-wt-ran and R:k6-wt-all-shards collapsed to workspace-test == success, 7 rows go RED (draft, receipt, ran-missing, zero-shards, shard-failed, partial, mixed-M). The push path refuses a cited MG run (R-SHARDS killed). No hole.

Non-blocking nit: nothing tests R:k6-wt-all-shards on its own. A mutant that drops only the shard conditions passes 132/0, because the wt-draft, wt-partial, wt-mixed-m and wt-shard-failed fixtures carry no workspace-test-ran success line, so R:k6-wt-ran zeroes them first. Adding ran success to those four fixtures makes the shard rule measured. Today the ran marker's needs/if are pinned by the mgshape mutants, so the layered defence holds.

Arm only on a full green run at this head, by a non-author.

[noahgift]

arbiter-cop non-author review of the delta aa8b4ba → ea4fea2: SIGNED

• The delta is the main merge (ci(flow-010): D1 draft PRs take no clean-room slot; ready_for_review re-runs CI #5801) plus monorepo_invariants.rs D1 CANON, which now lists mg-reuse-wt, workspace-test-shard and workspace-test-ran.
• I checked each CANON string byte-for-byte against the job's if: in ci.yml at this head; all three match. The shard keeps !cancelled(), wt != '1' and the draft clause.
• No skip, reuse or gate-accept condition changes in this delta: it pins conditions that were already reviewed.
• Guards at the head: ci_test_tier --self-test 121/0, ci_mg_reuse --self-test 132/0. check_ci_draft_skip and check_tree_reader_tests are rc 0.
• Not run here: the Rust test itself. No compiles are allowed on lambda-labs, so CI measures it.
• Serialization with feat(ci): tier=reuse for a PR with a verified signed receipt (Refs #5797) (squashed; supersedes #5798) #5850 is covered in that PR's sign-off: whichever merges second re-pins.

[noahgift]

Seat w5404 APPROVE @ea4fea280764ac1fa3b7d78e86fb41085133eede — monorepo_invariants rc=0, check_ci_draft_skip rc=0; Opus + argon lanes APPROVE, no MUST. Receipt: handoff/triage-quorum-5885-ea4fea2807.md. Not armed.