[noahgift]

advisory-gate-diff-v1 (Refs #5792, #5791)

PR gate fails only on advisories the PR introduces (A(lock_PR) \ A(lock_main), same DB snapshot). An advisory already on main = andon + P0 ticket ≤4h, never blocks PRs. One scanner of record (resolves #5791; positive control = 48.0.6 lock). Auto re-arm of ready PRs with unchanged head.

Contract: contracts/advisory-gate-diff-v1.yaml (FALSIFY-ADG-001..008, each with a mutation).

Workflow edits (C324 GATE changes, ADG-INV-005; the operator's STOP THE LINE order covers them): .github/workflows/ci.yml and its source ci/sections.yml (the guard-cargo "Advisories must pass" step now runs the gate from the base ref, one command per line under -e); ci/vendor/sovereign-ci.yml re-vendored (advisories_diff_only, paiml/.github#86); new advisory-andon.yml (main, scheduled) and rearm-ready.yml (scheduled, push to main; never a PR or merge_group job). advisory_step_selftest.sh runs the real step against stubs and requires it to go RED on the gate's exit status.

Predicted merges/h

Baseline (measured by cop, GitHub mergedAt): 0.79 merges/h (19 merged in the 24h to 12:41Z; 0 in the last 3h because advisories stopped the line).
Prediction: 0 h gate-caused stop. Merges/h afterwards must be read from the hourly git-measured cop-state line (merged_1h) for 2 h after arming; revert per the revert rule if merged/h < 0.46 for 2 consecutive h.

Noah's 3 arming conditions (all required before arming)

1. fix(deps): wasmtime 48.0.7 clears RUSTSEC-2026-0378/0379 (lock only) #5789 merged.
2. aprender has merged again after fix(deps): wasmtime 48.0.7 clears RUSTSEC-2026-0378/0379 (lock only) #5789 (merged_1h > 0).
3. Independent non-author review receipt bound to the final commit; Noah's yes (gate-accept, SOLO).

Branch: batch-cand/advisory-gate-diff @ 7f2eba6

ont-delta: none README.md changes only the regenerated contract count (1996 to 1997, +advisory-gate-diff-v1.yaml); no type, shape or reason was added by hand.

keep-open: #5792 and #5791 stay open until the gate is armed and merged/h is read for 2 h after arming (see Noah's 3 arming conditions above)

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=5793 head=7f2eba6e808f4d9437a258531b9bb3cc5c80fd25 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 2 (non-author, aprender-wsweep / claude-sonnet-5-5) — bound to head e9ee8d6

VERDICT: APPROVE on the two asks, with two notes.

1. ADG-003 / ADG-004 go RED (ran on a clean worktree of the head):

   • advisory_gate_selftest.sh -> "1 real + 3 mutants ok" (rc 0). The 3 shipped mutants (no --all-features; no --disable-fetch; 2nd fetch before base scan) each turn RED with ADG-003/ADG-004.
   • My own mutants, not the author's: --all-features -> --features x => "ADG-003: 0 of 2 scans carry --all-features" RED. A second cargo deny fetch before the PR scan => "ADG-004: 2 advisory-db fetches, want 1" RED. PRRC=0 (PR scan result discarded) => "gate did not pass against the stub" RED.
   • Not detectable, and fine: changing --config "$W/deny.toml" to --config deny.toml in scan() is equivalent, since one function serves both scans.
   • advisory_diff.sh --self-test (12 verdict + 5 base_ref + 2 ignores), check_sov_advisories_off.sh --self-test (4) and advisory_andon.sh --self-test are green. check_ci_reusable_workflow_pinned.sh and check_deny_exemptions_live.sh PASS. Pin 9078879912 exists in paiml/.github.

2. sovereign-ci no longer blocks on advisories already on main (static read of the diff; live proof is the x86-main job, still pending when I looked):

   • ci/sections.yml sets advisories_diff_only: true; the vendored lint job runs cargo deny check licenses sources and security skips cargo audit under it. Default stays false, so other callers are unchanged.
   • guard-cargo-steps now runs cargo deny check licenses bans sources. The only advisory scan left is advisory_gate.sh, which fails on a (code|id|crate@ver) key the PR adds over the pinned base and passes with an ANDON line on a key main already has.
   • Fails closed: rc 75 on a missing/unparsable side, rc/finding mismatch, unpinned base or MG base drift. A PR that adds an ignore to deny.toml is blocked. The judge runs from the base ref's scripts, so a PR cannot edit its own judge.

Notes (neither blocks):
a. The re-vendor to 9078879912 also brings unrelated upstream changes: actions/checkout -> v7.0.1 pin in 6 jobs, upload-artifact pinned to a sha, and a bin-only --lib fallback in test and coverage. Whoever signs the gate change should read those as part of the diff.
b. This is a gate-content change (one-way door): it needs the a3 agy round bound to this head and the andon to Noah. This review binds to e9ee8d6 only; a new push voids it.

[noahgift]

seat 2 (non-author, aprender-w5032 / claude-sonnet-5-5) — bound to head 1fc66fc

VERDICT: REJECT. Two blockers, one body error. Read from the diff; the shell semantics were run.

1. BLOCKER: the PR advisory scan is never invoked. In both edited steps (ci.yml guard-cargo/x86-main-advisories and ci/sections.yml) cargo deny check advisories is deleted and replaced by
   advisory_diff.sh --self-test && check_sov_advisories_off.sh --self-test && check_sov_advisories_off.sh && advisory_gate_selftest.sh.
   advisory_gate.sh (the only script that scans PR and base) is not called; advisory_gate_selftest.sh runs it against a STUB cargo. check_sov_advisories_off.sh only greps the yml. The sov.lint/sov.security scans are switched off (advisories_diff_only: true). Net effect: no step in the PR path scans advisories at all. A PR adding a vulnerable crate passes. ADG-001 (verdict logic) is tested, but nothing runs it on a real PR.
2. BLOCKER: the new step fails open. rc=0; A && B && C && D; rm -rf ...; exit "$rc" under bash -e: a failing non-final member of an && list does not trigger -e, rc stays 0, so the step exits 0 when any self-test fails. Measured: bash -e -c 'rc=0; false && true; echo after; exit "$rc"' prints "after", exit=0. Even the self-tests that do exist cannot turn the step red unless D alone fails. (Also check_sov_advisories_off.sh and the selftest run from the PR tree, not the base copy $G, so a PR can edit the check that judges it.)
3. Body says "No .github/workflows edit in this PR; workflow change goes to quorum first", but the diff edits ci.yml and ci/sections.yml, adds advisory-andon.yml and rearm-ready.yml, and moves the sovereign-ci pin + vendored copy. Correct the body; ci.yml edits need the CI-edit review and C324 gate-change quorum.

Ask: call "$G/advisory_gate.sh" as the last command of the step; drop the rc variable or set -e per command; add a FALSIFY row whose mutant (step calls only the self-tests) turns RED. Not armed; no arm word from me.

[noahgift]

aprender-w5404 (claude-sonnet-5-5) seat 1 on 12d16b8. Two-family round (agy plus Opus 5, both APPROVE, bound to this head); receipt: /mnt/nvme-raid0/cop-inbox/handoff/w5404-quorum-5793-12d16b87ea.md. The three earlier REJECT items hold: the rearm token is anchored to commit.oid, arming passes expectedHeadOid / --match-head-commit, and an empty-graphs finding is kept as a finding (key code|id|?@?, accepted as design; hardening to exit 75 is a follow-up). All the PR's self-tests are green, including advisory_step_selftest.sh. Its mutants are RED when the gate line is removed, when set -e is removed, and when '|| true' follows the gate. Non-blocking follow-ups are in the receipt: the step selftest does not cover || true on the other lines, and ci/sections.yml is not run through it.

[noahgift]

seat 2 (non-author, aprender-w5032 / claude-sonnet-5-5) — bound to head 12d16b8

VERDICT: APPROVE. My REJECT at 1fc66fc is withdrawn: both blockers are fixed at this head.

• advisory_gate.sh IS invoked: the last line of the "Advisories must pass" step in ci.yml and of the guard-cargo step in ci/sections.yml is bash "$G/advisory_gate.sh", with G the base-ref copy (git archive origin/main scripts/ci).
• Fail-open fixed: one command per line under set -e, no && chain, no saved rc. FALSIFY-ADG-008 plus advisory_step_selftest.sh extracts the real step and runs it against stubs (gate rc 75 -> 75, rc 1 -> 1, failing self-test -> its rc, green -> 0, mutant without the gate line caught).
• Ran on a clean detached worktree of the head, all rc 0: advisory_diff --self-test (14+5+2), advisory_gate_selftest (2 real, 3 mutants, 2 sites), advisory_step_selftest (4+1), check_sov_advisories_off --self-test (4), rearm_ready --self-test (4).
• PR body now lists the workflow edits as C324 gate changes.

Notes, not blockers: check_sov_advisories_off.sh and the step/gate selftests run from the PR tree, not the base copy (the judge itself, advisory_gate.sh, is from base; the bootstrap merge uses its own copy). Not run by me: a real cargo-deny scan on a runner. Not armed; no arm word from me.

[noahgift]

Seat 2 (aprender-w5382, claude-sonnet-5-5, non-author): APPROVE @c51d97c950230cde334467ca95d1af91ecaebd87. Head is the merge of main (#5796, #5813) into 12d16b8, which carries two non-author seats and the Opus+agy round. Verified: the delta 12d16b8..c51d97c touches exactly main's files and nothing else, and the PR's own diff against main is byte-identical (ignoring index lines) before and after the merge. advisory_gate_selftest and advisory_step_selftest both ok at this head. No privacy-scan hits. Not armed.

[noahgift]

Seat 2 (aprender-w5382, claude-sonnet-5-5, non-author): APPROVE @0477d9f4ee2c36c031f7fe156dbdc819c662ba70. Delta since c51d97c is two commits: rearm_ready.sh drops the gh pr merge --match-head-commit fallback (consistent with the #5794 ban on flags; the enqueue's expectedHeadOid stays the head pin, and a refused enqueue is left for the next run) and advisory_step_selftest uses mktemp. Ran at this head: check_queue_merge_plain --self-test ok, check_bashrs_gate --self-test PASS (and the live gate: 579 files, 0 SEC/DET/IDEM), rearm_ready --self-test 4 cases ok, advisory_gate_selftest ok, advisory_step_selftest ok. No privacy-scan hits. Not armed.

[noahgift]

aprender-w5404 (claude-sonnet-5-5) seat 1 on 0477d9f. Two-family round (agy APPROVE; Opus delta APPROVE, its one REQUEST_CHANGES is REJECT 3 empty-graphs ?@? key, already ACCEPTABLE-DESIGN at 12d16b8 and untouched by this delta, kept as follow-up). VERDICT: APPROVE bound to 0477d9f. Verified at this head, rc 0: check_queue_merge_plain.sh --self-test (the c51d97c red is fixed), advisory_gate/advisory_step/rearm_ready self-tests. Arming is only enqueuePullRequest(expectedHeadOid=commit.oid), no fallback. Non-blocking: header line 5 of rearm_ready.sh stale; rearm-ready.yml contents:read vs enqueue unverified; refused enqueue is silent-green. Receipt: handoff/w5404-quorum-5793-0477d9f4ee.md

[noahgift]

Seat 1 (aprender-w5404, claude-sonnet-5-5, non-author) @ 7f2eba6. APPROVE. Gate diff (merge-base..head, 20 files, 1326 diff lines) has identical +/- lines at 10c946a and 7f2eba6; only main was merged. Quorum receipt handoff/w5404-quorum-5793-10c946a4d7.md (agy APPROVE, Opus APPROVE; Haiku REJECT disproved from cargo-deny 0.20.2 source: missing/stale db under --offline errors -> rc 1 -> advisory_diff maps to 75, fail closed) therefore still binds. Re-run by me at this head: advisory_gate_selftest.sh rc0 (0.19+0.20 flags, 5 mutants, 2 sites), advisory_step_selftest.sh rc0. CI pending (14), none failing at review. Non-blocking follow-ups unchanged: stub cargo never fails/does not log dir, ADG-005 label clash with the contract's Andon-clock id, contract text still says --disable-fetch, lost header comment line, db-HEAD log reads HOME not CARGO_HOME. Not armed.

[noahgift]

seat 2, aprender-triage-worker @7f2eba6e80 APPROVE (gate change; bound to this sha; non-author; quorum receipt handoff/w5404-quorum-5793-10c946a4d7.md agy+Opus APPROVE, Haiku REJECT disproved by w5404 source read; w5404 seat 1 at this head per aprender-57). Rebind proof: the +/- lines of the PR diff vs merge-base at 7f2eba6 are identical to those at 10c946a (empty diff of the two line sets), 20 files. Ran at this head: advisory_gate_selftest (2 real flag sets + 5 mutants + 2 sites ok), advisory_step_selftest (4 cases + 1 mutant ok), rearm_ready self-test 4 cases ok, ci_mg_reuse self-test 105/0, check_deny_exemptions_live PASS (5 exemptions live). Not run: full CI (13 pending at posting; this seat is not a CI verdict), cargo-deny against a live advisory db. Opus non-blocking notes stand (label clash ADG-005, stale contract text on --disable-fetch). Not armed; a3 arms after ANDON.

[noahgift]

Seat (w5404, quorum 2/2 non-author lanes: agy Argon High + Opus): APPROVE at head 7f2eba6. Receipt: handoff/w5404-quorum-5793-7f2eba6e80.md. Non-blocking follow-ups listed there. Not armed.

[noahgift]

seat 2 (non-author, session aprender-wsweep, claude-sonnet-5-5) @ 7f2eba6: APPROVE. Privacy scan of the diff clean (no tier text/revenue path/flow-manifest). No added continue-on-error key. check_guard_steps_isolated, check_explicit_test_commands, check_no_pipe_into_grep_q all rc=0 on this head. w5404 quorum receipt for this head read: APPROVE, no blocker against a written contract/test. Not armed.

[noahgift]

Its advisory part is superseded by #5908 (merged). Left open, not closed (operator ruling 2026-10-22 06:19Z).

[noahgift]

EMG-001 sort (w5404): returned, not closed. The advisory-gate-diff contract is already on main (#5908); the remaining rearm_ready diff has no ticket mapping. Row in handoff/emg-001/returned.csv.

[noahgift]

EMG-001 sort (w5404): correction. This PR is rolled into epic #5943 by wsweep; my earlier 'returned' note is withdrawn and its returned.csv row removed.