maintenance: revive PRs demoted by the 2026-08-18 WIP sweep with current exact-head evidence

[ll7]

Purpose

Tracker for PRs that were fully reviewed on 2026-08-18 (exact-head trailers published, hosted CI green at the live head) and then demoted by the factory closure-lane sweep (see ll7/codex-orchestrator#1696) or left "one label away" by the same run. Park, don't close: each entry names its exact head, what exists, and what remains.

Ready — evidence current at live head, needs only un-draft + merge-ready (re-verify head unchanged first)

PR	head	evidence at head	state after sweep	remaining
#7389	86124a6	gate-verdict, base-policy ordinary-cas, pr-metadata; 255 tests local; CI green	draft + deferred	un-draft, merge-ready, merge
#7390	38ee6f0	same set; exhaustive Bresenham equivalence 28,561 pairs vs inline algorithm	draft + deferred	same
#7104	0c37b0b	same set; domain-approved; 18 tests	draft + deferred	same
#7435	55b8c4d	same set at live head (delta vs 23bbb70 re-reviewed); overlap vs #7471/#7419/#7420 clean	needs-review	merge-ready, merge

Ready but base-sensitive — re-check freshness before label

PR	head	note
#7336	86c9d2c	base-sensitive marker file; base was fresh at review, current-base subset 22 tests passed, base-policy: current-base published; every main advance re-stales it — refresh, rerun subset, republish, label, merge in one motion

Needs one more step before label

PR	head	remaining
#7431	4b9dc20	conflicts with #7428/#7433 union-resolved, clean wheel yaml-import CI fix pushed; needs current-head legal/release exact-head review (comment on PR), CI green → label
#7486	39a0727	#7482/#7489 shared-baseline repair; factory rebased mid-preflight, reviews exist at 39a0727; land first — it unblocks readiness runs on every other PR
#7146	deed666	CI green; do not merge before #7486 and then bump EXPECTED_DIRECT_CHILD_COUNT 296→297 (both branches rewrite the same line, git merges it cleanly and the count ends one short)
#7144	f35d7a4	now draft; base-policy + pr-metadata at head; gate-verdict withheld pending determinism-gate and un-draft

Not in scope here (own packets exist)

Author-decision: #7046, #7081, #7334, #7343, #7236. Parked with fix on side branch: #7357 (review/pr-7357-observation-binding-20260818 @ 5949b70, #7454). Quarantined: #7364. Closed superseded by #7469: #7092, #7102.

Revival condition

Resolve after each row above is merged or its PR reaches a documented terminal state. If the closure lane demotes these again while their evidence is current, add the event here and to ll7/codex-orchestrator#1696.

[ll7]

Maintainer ruling — 2026-08-18

The 2026-08-18 closure-lane demotions of PRs whose live head carried current accepted exact-head evidence and green CI (draft + deferred + needs-review) are reversed. Those PRs re-enter the queue: re-verify head unchanged and CI green, un-draft, re-apply merge-ready, merge via the guarded merger. Evidence-stale relabels from the same sweep (predecessor-head verdicts, moved heads) stand.

Consequence: a config change limiting the closure lane to evidence-stale relabels only is being prepared against ll7/codex-orchestrator#1696. Reopen condition: the sweep recurs on evidence-current PRs after that change lands.

[ll7]

Revival progress update (2026-08-18 09:21Z)

Rows from the "Ready" table processed by the autopilot loop:

• refactor: share PySF state slice constants (clean replacement) #7389 — un-drafted, merge-ready applied, MERGED (commit 81b802d6b65017bdb09a08b168e44da81a47298a).
• refactor: share Bresenham rasterizer (clean replacement) #7390 — un-drafted, merge-ready applied, MERGED (commit 8859e0ae6981404fbcdb9d55fe9918e059785e2e).
• feat(research): add Open Dreamer model quality gate #7104 — un-drafted, merge-ready applied, MERGED (commit 97c8d4357cd01165863b972a311c6af76b6ec724).
• fix(ci): require exact-head changed coverage for admission #7435 — merge-ready applied, MERGED (commit 9a161d77a592e08407a6a325a59036489ae91744).

Remaining rows not processed by this run:

• fix(workflow): harden stale-base observation provenance (#7261) #7336 — base-sensitive, base now stale against current main; needs a dedicated refresh lane (rebase, rerun current-base subset, republish evidence, then merge-ready).
• fix(release): bind legal inputs to validators #7431 — needs current-head legal/release exact-head review and terminal CI.
• fix(dev): refresh benchmark namespace inventory (#7482) #7486 — queued after friction: refresh benchmark namespace inventory after merged trace-dossier module #7482/friction: dependency review test assumes the license action is the last workflow step #7489 shared-baseline repair; land first to unblock readiness runs.
• fix(benchmark): accept reviewed issue 5409 rerun receipts #7146 — do not merge before fix(dev): refresh benchmark namespace inventory (#7482) #7486, then bump EXPECTED_DIRECT_CHILD_COUNT 296 -> 297.
• fix(repro): bind exact-repeat verdicts to execution context #7144 — draft; needs determinism-gate and un-draft.

[ll7]

Merge lane (pass 3) — closed out under the 2026-08-18 ruling

All ten PRs in scope reached a terminal state. Nine merged, one blocked on a real gate failure.
Two merge lanes were live simultaneously (this one and the ll7 factory), so the lane column records
which one executed the merge — the review evidence each merge consumed is noted where it differs.

PR	head	merge SHA / blocker	lane
#7500	ff85bad12	416d43819	this lane — reviewed, labeled, merged
#7435	55b8c4db3	9a161d77a	already merged 09:06Z before this lane started
#7389	86124a68a	81b802d6b	factory, 09:17Z
#7390	38ee6f001	8859e0ae6	factory, 09:17Z
#7104	0c37b0b3d	97c8d4357	factory, 09:18Z
#7431	4b9dc208f	266a13090	this lane — reviewed, labeled, merged
#7144	f35d7a426	fe5a2c500	this lane — un-drafted, reviewed, labeled, merged
#7486	d3ca231a5	98ef78ca8	factory, 10:12Z — consumed this lane's merge-ready + review
#7336	298ab2316	c2690579e	factory, 10:32Z — consumed this lane's exact-head evidence at 298ab2316
#7146	740b79afc	BLOCKED — changed-coverage-gate fail	this lane — refreshed, reviewed, parked

#7146 — the one blocker

Refreshed onto main (incl. #7486), #7486 union conflict resolved, 31 of 32 checks green. One real,
PR-attributable failure:

- robot_sf/benchmark/campaign/campaign_checkpoint_preflight.py: 0.0%  (changed executable lines 0/1)
- robot_sf/benchmark/issue_5409_campaign_identity.py:           55.9% (changed executable lines 38/68)   min=80.0%

The uncovered lines (27, 31, 63, 77, 92, 95, 101, 106, 131-138) are the fail-closed branches the PR
body already advertises, so the fix is additive tests, not a scope change. Full diagnosis and the
three-step unblock are in my review on #7146. I deliberately did not write those tests myself —
authoring the coverage that clears my own gate verdict would defeat the exact-head contract. No
gate-verdict published, merge-ready not applied. Terminal state: BLOCKED_EXTERNAL.

What cost the most time, and is worth fixing

1. The main-refresh treadmill. #7486's head moved eight times in ~45 minutes — an automatic
refresher rebased it onto main every time any lane merged, while its CI takes ~9 minutes. Every
refresh voids the exact-head carriers and restarts the cycle, so no CI/carrier pair can settle. I
parked it rather than weaken the exact-head contract; the factory landed it once traffic eased.
#7486 was ordinary base policy, where a stale base is explicitly tolerated by compare-and-swap — so
those refreshes bought nothing and cost the merge. Suggestion: gate the auto-refresher on
check_base_sensitive_gates.py and refresh only base_sensitive PRs.

2. New follow-up filed: #7517. #7435's exact-head changed-coverage requirement is unsatisfiable
for docs-only PRs: ci.yml is paths-ignore-filtered on **/*.md and docs/**, so
changed-coverage-gate never produces a check run, and merge_queue_gate.py then fails closed with
changed_coverage_proof_missing. Harmless today (source-PR runs are --advisory, main is
unprotected) but it becomes a permanent deadlock for every docs-only PR the moment
Merge Queue Gate / merge-queue-gate is made a required check — which is the stated intent. Observed
on #7500 and #7431.

3. Two stale-narrative bodies caught before labeling — the #7448 / #7491 defect, now codified by
#7500's goal-pr-review hardening, which this lane merged first and then immediately needed:

• fix(release): bind legal inputs to validators #7431 pinned the superseded head 882d8b5f8 and ended "It remains unapproved and not
  merge-ready …" — that sentence was one label away from being squashed into main.
• fix(benchmark): accept reviewed issue 5409 rerun receipts #7146 carried a gate-verdict: accepted @ deed6665f trailer in the body for a superseded
  head, plus a "Domain-aware approval is still pending" line contradicting the recorded benchmark: #5409 launch-packet validator and paired-horizon report disagree on valid campaign IDs #7453 ruling
  stated earlier in the same body.

Both re-narrated via gh_pr_body_rest.py --reconcile before any label write.

4. The 297 trap in #7146. Git auto-merged EXPECTED_DIRECT_CHILD_COUNT to 296 because both
sides happened to say 296 — silently wrong for the merged tree (296 from #7486 + this branch's own
new direct child = 297). Only caught by running the test in a branch-attached worktree; a
detached one errors out spuriously because the inventory helper shells out to git symbolic-ref HEAD.

Verification notes

• Linked-issue semantics checked after every merge. bug(release): trigger licensing checks on all provenance and legal inputs #7300 correctly CLOSED/COMPLETED via fix(release): bind legal inputs to validators #7431's
  Closes #7300; reproducibility: bind exact-repeat cross-host verdicts to canonical execution-context provenance #7128 and friction(merge-gate): merge-ready + stale 'not merge-ready' body narrative reaches main (#7374, #7435) #7491 correctly still open (Refs/Relates, no auto-close); follow-up: measure risk-tiered stale-base policy under normal throughput #7261
  correctly open. bug(provenance): PR bodies carry fabricated exact-head SHAs in gate-verdict trailers #7448 / friction: refresh benchmark namespace inventory after merged trace-dossier module #7482 / audit(6905): regenerate the residual top-level benchmark namespace inventory after completed domain moves #7331 / friction: pin empty-changed-lines coverage verdict introduced by PR #7435 #7460 were closed manually by ll7, not by a merge keyword
  (no commit_id on the close events) — no accidental closures from this lane.
• Every merge used the guarded path: merge-ready + non-draft + base main + reconciled metadata
  digest + green CI at exact head + check_pr_current_base_cas.py passed + no unresolved threads +
  --squash --delete-branch --match-head-commit, with remote read-back. Two merges reported a local
  branch-deletion failure only (used by worktree at … in another lane's checkout); remote state is
  correct.
• API budget: core quota exhausted once at 09:26Z (5,000/hr shared across all lanes), costing ~4
  minutes. All worktrees created by this lane were removed; none used /dev/shm.