Triage the remaining stranded P2 review findings from the post-merge sweep

[kyle-sexton]

This was generated by AI while closing out the stranded post-merge review-findings effort.

Context

The stranded-findings sweep (morning-brief.sh --stranded-days N, merged in #1781) surfaced ~34 P2
findings posted on merged PRs after they merged, invisible to the thread-resolution merge gate. The
P0/P1 tiers of that backlog are now drained (#1784, #1786, and the eight findings from #1322/#1425
via #1886). The P2 tier was assigned to an agent that produced no durable output; this issue carries
the remaining work so a cold agent can pick it up.

Task

1. Enumerate: bash plugins/claude-ops/skills/morning-brief/scripts/morning-brief.sh --repo melodic-software/claude-code-plugins --stranded-days 6 — findings date ~2026-07-28..31, so widen --stranded-days (10, then 14) until the P2 rows appear.
2. Exclude PRs already drained: docs(topics): absorb the context-engineering source into a designed, dispatched instruction-surface pass #1322, feat(discovery)!: dispatch explore and research to purpose-built subagents #1425, feat(autonomy): deterministic lane-stop gate + operator STOP notification #969, feat(claude-config): add audit-pass, one coordinated resumable instruction-surface pass #1318, docs: source-control the loop-lane launch prompts #1320, feat(source-control): explicit-autopilot merge-rung exception with fresh-agent resolution #1405, feat(claude-ops): consume lane restart-requests via scheduled headless reader #1720, context-guard: an uninstalled plugin's tee keeps executing through the shim, plus a compounding sh -c wrap #1787, fix(guardrails): close git-alias-chain re-expansion bypass of both git guards #1503.
3. Verify each remaining finding against current origin/main as a claim to falsify (in the P1 pass, 0 of 8 were already fixed — do not assume the earlier ~25% already-fixed rate).
4. A finding the author fixed in the same PR without resolving the thread still reports as stranded (deliberate STRICT behavior); classify those as "fixed-in-PR, thread-only" and only resolve the thread.
5. Fix real findings in PRs batched by plugin, each independently verified before merge.
6. Resolve threads via the guarded wrapper plugins/source-control/bin/source-control-babysit-resolve-thread (vetted-thread mode: --resolve --thread-id <id> --expected-comment-count N --expected-last-updated <ts>). Threads with any human (User) participant are refused by design — list them for the human instead.

Acceptance criteria

• Every stranded P2 row from the sweep has a verdict (fixed / already-fixed / fixed-in-PR-thread-only) with evidence.
• Real findings fixed and merged, batched by plugin, each with independent verification.
• Bot-only threads resolved; human-participant threads listed for the operator.

Related

Refs #1322
Refs #1425
Refs #1786

[kyle-sexton]

Scope correction and run-plan guidance from an interrupted execution attempt (2026-08-05); recorded so the next executor starts from it rather than re-deriving.

Scope: the sweep returns 73 stranded P2 rows, not the ~34 estimated above. The window is rolling and the repo merges heavily, so freeze the sweep output as a snapshot at run start, drive every captured row to verdict, and re-run the sweep once at the end — newly-stranded rows found then are reported as follow-up, not chased.

Run-plan guidance (from an advisor pass over the issue, adopted):

• Resolution timing: REAL findings resolve only after the fixing PR merges. ALREADY-FIXED and FIXED-IN-PR-THREAD-ONLY resolve once the verdict survives independent audit. INVALID-CLAIM gets a one-comment rebuttal (verdict + file:line evidence + verified main SHA) before resolving — a silently-resolved wrong "invalid" verdict is the worst failure mode. Add a distinct file-deleted/obsolete-on-main verdict category.
• Non-REAL verdicts need independent audit too: they produce thread resolution with no diff, so a per-batch fix-diff verifier never sees them. Include each batch's non-REAL verdicts in the verifier brief with quoted current-main evidence, before any resolve call.
• Stale-base guard: record the verified main SHA per verdict; re-read the touched files at fresh origin/main immediately before opening each fix PR; rebase before arming auto-merge.
• Wrapper freshness: never resolve from snapshot metadata — any reply (including your own rebuttal) bumps commentCount/lastUpdated; re-fetch immediately before each resolve call. Treat refusals as park-and-refetch, never force.
• Human drift: a human can reply mid-run; wrapper refusal (skipped-human-thread) is correct behavior — classify as newly-parked for the operator.
• Batching: by plugin per the acceptance criteria; changelog-parity operates per plugin. Singletons stay per-plugin unless the operator approves a grouped-misc deviation.
• End-of-run: post one consolidated verdict table on this issue (thread ID → verdict → evidence → fixing PR) — that is the artifact satisfying "every row has a verdict with evidence". Your own fix PRs can strand new findings; the end re-sweep catches them — report, don't recurse.

A clean read-only verification worktree from the interrupted run exists at C:/Users/KyleSexton/.claude-loop-worktrees/p2-verify-base (detached, clean) — reusable or ignorable.

[kyle-sexton]

Frozen run snapshot and scope correction (2026-08-06)

Posted by the executing session at run start. Supersedes the scope in the issue body and in the 2026-08-05 comment.

Sweep: morning-brief.sh --repo melodic-software/claude-code-plugins --stranded-days 14, run against origin/main at cff36aa5c7, frozen 2026-08-06T13:25:01Z.

Corrections to the stated scope

1. The tier split was misread. The sweep's per-PR [P1]/[P2] label is the PR's worst severity and the parenthesised number is its total finding count, so a [P1] ... (2 findings) row is not two P1s. Thread-level enumeration gives 16 P1 + 61 P2 = 77 threads, not "73 P2 rows".
2. Eight P1 threads are new and unaccounted for. The prior handoff recorded the P1 tier as drained. Eight P1 threads outside the nine known human-gated ones are live: on docs(conventions): add loop-lane convention with ci-workflows-work-loop contract slice #1147, feat(work-items): add work-loop and attend-queue loop-lane skills #1163, feat(context-guard): version-independent statusline shim for both guard plugins #1252, refactor(session-flow): restructure the handoff save-point taxonomy #1296, feat(claude-config): add audit-pass, one coordinated resumable instruction-surface pass #1318, fix(guardrails): close git-alias-chain re-expansion bypass of both git guards #1503, feat(autonomy): add C3 auto-merge cell with evidence predicate #1694, fix(autonomy,lanes): read the lane-stop gate's enable flag from trusted channels only #1865. fix(autonomy,lanes): read the lane-stop gate's enable flag from trusted channels only #1865 is this effort's own merged PR.
3. The exclusion list in step 2 of the issue body is stale. It names feat(claude-config): add audit-pass, one coordinated resumable instruction-surface pass #1318 and fix(guardrails): close git-alias-chain re-expansion bypass of both git guards #1503, both of which carry live P1 findings. The sweep predicate is unresolved AND first comment postdates the merge, so anything the sweep returns is by definition not drained. Sweep output supersedes the exclusion list.
4. Scope for this run is widened to P1-first, then P2 (operator-approved). Actionable set: 68 threads = 77 minus the 9 parked human-gated ones.

Predicate validation

All nine known human-gated thread IDs appear exactly where the handoff records them (…TzCBB/D/F/G/I on #1322, …TzjM_ on #1425, …TzKVB/…TzKVD on #1320, …T1lT5 on #1405), and per-PR counts reconcile against the sweep's own output. The enumeration reuses the sweep's structured severity matcher (badge alt-text / shields URL / leading bracketed marker), never a prose substring test.

Inventory

parked = human-gated, listed for the operator, not worked by this run.

Tier	PR	Thread	Location	Finding
P1	#1147	PRRT_kwDOTCGFQM6TRjqD	docs/conventions/loop-lane/README.md:53	Keep C3 merges human-gated
P1	#1163	PRRT_kwDOTCGFQM6TVrP0	plugins/work-items/skills/work-loop/SKILL.md:241	Evaluate the frontier exit against the start snapshot
P1	#1252	PRRT_kwDOTCGFQM6Tzj7k	plugins/context-guard/scripts/statusline-shim.sh:84	Stop resolving tees from uninstalled cache entries
P1	#1296	PRRT_kwDOTCGFQM6Tym33	plugins/implementation/skills/implement/SKILL.md:161	Bump the implementation plugin version
P1	#1318	PRRT_kwDOTCGFQM6TzY_h	plugins/claude-config/skills/audit-pass/reference/run-contract.md:277	Do not reclaim a live apply lock based on age alone
P1 parked	#1320	PRRT_kwDOTCGFQM6TzKVB	prompts/loops/loop-lane-prompts.md:475	Default reusable launches to human-only merging
P1 parked	#1322	PRRT_kwDOTCGFQM6TzCBB	docs/topics/context-engineering-claude-5/design/rerun-contract.md:698	Exempt fix-bearing pairs from comparability
P1 parked	#1322	PRRT_kwDOTCGFQM6TzCBD	docs/topics/context-engineering-claude-5/design/rerun-contract.md:701	Include external instruction surfaces in the run basis
P1 parked	#1322	PRRT_kwDOTCGFQM6TzCBF	docs/topics/context-engineering-claude-5/design/rerun-contract.md:826	Include the judging model in the run basis
P1 parked	#1322	PRRT_kwDOTCGFQM6TzCBG	docs/topics/context-engineering-claude-5/design/rerun-contract.md:701	Version the sweep machinery in the run basis
P1 parked	#1322	PRRT_kwDOTCGFQM6TzCBI	docs/topics/context-engineering-claude-5/PLAN.md:144	Remove the obsolete derived-subset acceptance gate
P1 parked	#1405	PRRT_kwDOTCGFQM6T1lT5	plugins/source-control/skills/babysit-loop/SKILL.md:296	Apply dimension floors before dispatching the resolver
P1 parked	#1425	PRRT_kwDOTCGFQM6TzjM_	plugins/discovery/skills/research-deep/SKILL.md:28	Verify single-topic Tier 2 artifacts before handoff
P1	#1503	PRRT_kwDOTCGFQM6T1tWv	plugins/guardrails/hooks/block-noncanonical-commit.sh:608	Exclude wrapper options from alias directory resolution
P1	#1694	PRRT_kwDOTCGFQM6T8HlP	plugins/autonomy/skills/setup/scripts/check-security-binding.mjs:349	Bump the autonomy plugin version to ship this cell
P1	#1865	PRRT_kwDOTCGFQM6WACHN	plugins/autonomy/hooks/lane-stop-gate.sh:167	Make the first-session arm claim atomic
P2	#1025	PRRT_kwDOTCGFQM6TyjYS	plugins/toolchain/skills/check/context/python.md:37	Align the Pyright prerequisite with the actual preflight
P2	#1030	PRRT_kwDOTCGFQM6TzHVg	plugins/markdown-format/hooks/markdown-format.sh:73	Handle missing Git before treating the file as out-of-tree
P2	#1033	PRRT_kwDOTCGFQM6TybdT	plugins/session-flow/skills/find-handoff/SKILL.md:84	Distinguish completed agents before retaining handoffs
P2	#1052	PRRT_kwDOTCGFQM6TxJRc	docs/topics/autonomy-ignition/PLAN.md:285	Restart the acceptance window after the cadence flip
P2	#1096	PRRT_kwDOTCGFQM6T02ir	plugins/skill-quality/.claude-plugin/plugin.json:5	Replace the old manifest version instead of duplicating it
P2	#1096	PRRT_kwDOTCGFQM6T02it	docs/PLUGIN-PHILOSOPHY.md:496	Reconcile the default worker with the normative rule
P2	#1124	PRRT_kwDOTCGFQM6TzBwz	plugins/disk-hygiene/skills/setup/scripts/python3_alias_probe.py:31	Support the bootstrap interpreters the setup permits
P2	#1147	PRRT_kwDOTCGFQM6TRjqH	docs/topics/ci-workflows-work-loop/babysit-PLAN.md:147	Make babysit-loop wait for the rate-limit contract
P2	#1252	PRRT_kwDOTCGFQM6Tzj7l	plugins/context-guard/skills/setup/SKILL.md:110	Unwrap the generated sh -c adapter on reruns
P2	#1260	PRRT_kwDOTCGFQM6TzDQs	plugins/discovery/skills/research/context/discipline.md:99	Do not treat a sitemap as proof that a rung is absent
P2	#1264	PRRT_kwDOTCGFQM6TzaPh	plugins/source-control/skills/babysit-prs/reference/safety.md:457	Exempt help runs from the safety contract
P2	#1275	PRRT_kwDOTCGFQM6TzGBZ	plugins/guardrails/hooks/block-dangerous-git.sh:386	Replay wrapper directory changes in the hash probe
P2	#1280	PRRT_kwDOTCGFQM6TzHGM	docs/adr/0005-bound-instruction-surface-work-by-question-not-population.md:186	Restrict the inventory to loadable supporting files
P2	#1280	PRRT_kwDOTCGFQM6TzHGN	docs/adr/0005-bound-instruction-surface-work-by-question-not-population.md:279	Define the plane without requiring pre-session residency
P2	#1296	PRRT_kwDOTCGFQM6Tym34	plugins/implementation/CHANGELOG.md:19	Describe the standalone fallback accurately
P2	#1301	PRRT_kwDOTCGFQM6T0Qjw	plugins/work-items/skills/setup/SKILL.md:205	Verify the repo before auto-binding GitHub
P2	#1316	PRRT_kwDOTCGFQM6Tz7bA	plugins/claude-config/skills/audit-instructions/SKILL.md:133	Honor CLAUDE_CONFIG_DIR for the default memory store
P2	#1316	PRRT_kwDOTCGFQM6Tz7bD	plugins/claude-config/skills/audit-instructions/reference/criteria.md:129	Exclude preloads from load-deferring destinations
P2	#1316	PRRT_kwDOTCGFQM6Tz7bF	plugins/claude-config/skills/audit-instructions/reference/conflict-criteria.md:107	Add agent-specific memory to the comparison corpus
P2	#1316	PRRT_kwDOTCGFQM6Tz7bG	plugins/claude-config/skills/audit-instructions/SKILL.md:157	Apply hook disablement before inventorying plugin hooks
P2	#1318	PRRT_kwDOTCGFQM6TzY_i	plugins/claude-config/skills/audit-pass/reference/run-contract.md:438	Exclude ambiguous collisions from exact suppression matching
P2	#1319	PRRT_kwDOTCGFQM6TxHZQ	plugins/guardrails/hooks/skill-reference-verify.sh:121	Honor manifest-declared skill paths
P2	#1319	PRRT_kwDOTCGFQM6TxHZS	plugins/guardrails/hooks/skill-reference-verify.sh:179	Preserve short substring edits during reconstruction
P2	#1319	PRRT_kwDOTCGFQM6TxHZU	plugins/guardrails/hooks/skill-reference-verify.sh:184	Scope reconstruction to references changed by this Edit
P2 parked	#1320	PRRT_kwDOTCGFQM6TzKVD	prompts/loops/loop-lane-prompts.md:376	Carry the discipline preamble into dispatched task briefs
P2	#1323	PRRT_kwDOTCGFQM6Ty6T0	docs/adr/0004-rightsize-instruction-surfaces-by-incumbent-first-arbitration.md:158	File the unenforced disk-hygiene guard defect
P2	#1323	PRRT_kwDOTCGFQM6Ty6T1	docs/adr/0004-rightsize-instruction-surfaces-by-incumbent-first-arbitration.md:276	Link dotfiles PRs to the dotfiles repository
P2	#1331	PRRT_kwDOTCGFQM6T0QzB	plugins/source-control/skills/babysit-prs/scripts/prune_babysit_worktrees.py:265	Restore the gitfile whenever the directory survives
P2	#1343	PRRT_kwDOTCGFQM6TzPgn	plugins/claude-config/skills/audit-instructions/SKILL.md:200	Keep nested project memory pairs in I15
P2	#1343	PRRT_kwDOTCGFQM6TzPgo	plugins/claude-config/skills/audit-instructions/scripts/conflict-scan.sh:130	Separate coordinated directives before classifying polarity
P2	#1343	PRRT_kwDOTCGFQM6TzPgp	plugins/claude-config/skills/setup/SKILL.md:30	Document awk and sort in the setup remediation source
P2	#1349	PRRT_kwDOTCGFQM6TzCLa	plugins/claude-config/skills/audit-instructions/reference/criteria.md:240	Include the alternate project-memory path in I14
P2	#1349	PRRT_kwDOTCGFQM6TzCLd	plugins/claude-config/skills/audit-instructions/reference/criteria.md:258	Resolve active imports before exempting supporting documents
P2	#1355	PRRT_kwDOTCGFQM6TzQHE	plugins/source-control/skills/babysit-prs/reference/orchestration.md:593	Scope live-head queries to the target repository
P2	#1355	PRRT_kwDOTCGFQM6TzQHF	plugins/source-control/skills/babysit-prs/reference/orchestration.md:646	Move the live-head check after the base fetch
P2	#1357	PRRT_kwDOTCGFQM6Txl3A	docs/adr/0003-verification-guards-earn-default-on-by-measured-precision.md:31	Measure actual edits instead of whole-file prevalence
P2	#1369	PRRT_kwDOTCGFQM6Tz-a2	plugins/miro/skills/setup/SKILL.md:55	Register the marketplace at the selected scope
P2	#1369	PRRT_kwDOTCGFQM6Tz-a3	plugins/miro/skills/setup/SKILL.md:91	Remove the no-op -y requirement
P2	#1383	PRRT_kwDOTCGFQM6TzlNu	plugins/claude-ops/skills/lanes/scripts/lane-launcher.sh:332	Hash repository keys in the target repository
P2	#1383	PRRT_kwDOTCGFQM6TzlNw	plugins/claude-ops/skills/lanes/scripts/lane-launcher.sh:279	Reject non-string names before testing path components
P2	#1384	PRRT_kwDOTCGFQM6Tzy7H	plugins/work-items/skills/track/actions/start.md:77	Verify the fully qualified remote-tracking ref
P2	#1429	PRRT_kwDOTCGFQM6Tzdue	scripts/check-contract-slice-prune.sh:139	Normalize the configured root before matching paths
P2	#1429	PRRT_kwDOTCGFQM6Tzduh	scripts/check-contract-slice-prune.sh:92	Preserve literal hashes when parsing contract_dir
P2	#1429	PRRT_kwDOTCGFQM6Tzduj	scripts/check-contract-slice-prune.sh:183	Check overlapping contract roots independently
P2	#1431	PRRT_kwDOTCGFQM6Tzl-v	plugins/review/skills/fanout/context/findings-normalization.md:21	Select the plugin comment rather than the latest comment
P2	#1445	PRRT_kwDOTCGFQM6Tz-jR	scripts/check-contract-slice-prune.sh:119	Reject Windows-style absolute contract roots
P2	#1449	PRRT_kwDOTCGFQM6T1cvS	plugins/disk-hygiene/skills/clean/scripts/destructive_guard.py:900	Preserve fd 2 when the null-device open reuses it
P2	#1450	PRRT_kwDOTCGFQM6T0gnF	plugins/skill-quality/scripts/check-listing-budget.sh:167	Normalize accepted integer overrides to decimal
P2	#1450	PRRT_kwDOTCGFQM6T0gnG	plugins/skill-quality/scripts/check-listing-budget.sh:262	Strip YAML comments before measuring listing scalars
P2	#1466	PRRT_kwDOTCGFQM6T0gfE	plugins/guardrails/hooks/skill-reference-verify.sh:246	Restrict reconstruction to references overlapping the anchor
P2	#1466	PRRT_kwDOTCGFQM6T0gfF	plugins/guardrails/hooks/skill-reference-verify.sh:233	Avoid a full-file grep for every hunk line
P2	#1530	PRRT_kwDOTCGFQM6T1s2n	scripts/shell-portability-tokens.txt:98	Detect fully quoted --sort arguments
P2	#1534	PRRT_kwDOTCGFQM6T1rfG	scripts/shell-portability-tokens.txt:117	Stop sed context at shell command separators
P2	#1534	PRRT_kwDOTCGFQM6T1rfI	scripts/shell-portability-tokens.txt:117	Detect unsuffixed in-place clusters without -E
P2	#1580	PRRT_kwDOTCGFQM6T4tKN	plugins/typos-format/hooks/typos-format.sh:313	Detect concurrent changes before classifying rewrites
P2	#1590	PRRT_kwDOTCGFQM6T4klg	plugins/source-control/skills/commit/scripts/exec-bit-check.sh:204	Include copy destinations in the exec-bit scan
P2	#1633	PRRT_kwDOTCGFQM6U4QmR	plugins/source-control/skills/pull-request/SKILL.md:178	Define a fallback when no tracker is available
P2	#1738	PRRT_kwDOTCGFQM6Uwxwm	plugins/disk-hygiene/skills/clean/SKILL.md:81	Fall back when AskUserQuestion is denied
P2	#1738	PRRT_kwDOTCGFQM6Uwxwt	plugins/disk-hygiene/skills/clean/SKILL.md:77	Limit the exact-list bar to deletion questions
P2	#1865	PRRT_kwDOTCGFQM6WACHR	plugins/claude-ops/skills/lanes/scripts/lane-launcher.sh:600	Preserve explicit empty marker overrides
P2	#1865	PRRT_kwDOTCGFQM6WACHX	plugins/claude-ops/skills/lanes/scripts/lane-launcher.sh:618	Inject arm IDs only into requesting plugin entries

Total 77 threads; 9 parked; 68 actionable.

Follow-up filed separately

--stranded-days is absent from morning-brief.sh --help despite being the flag this issue instructs executors to use.

[kyle-sexton]

Generated by AI while working the stranded post-merge review-findings sweep. Ledger:
.work/stranded-review-verdicts.md (gitignored, local-durable).

Verdict coverage: 68 of 68 — complete

77 stranded threads enumerated (16 P1 + 61 P2). 9 parked as out of scope, leaving 68 actionable.
Every one now carries a verdict. The P1 tier is fully verdicted; the P2 tier's final 17 were
verdicted at origin/main = 14a38f893f — 14 REAL, 2 ALREADY-FIXED, 1 INVALID-CLAIM, all bot-only,
none human-gated.

Shipped

Seventeen PRs merged carrying discharges from this sweep:

PR	Carries
#1999	context-guard — 1 REAL, reopened after a verifier CHALLENGE
#2002	skill-quality — 3 REAL
#2003	claude-config — 9 REAL
#2016	disk-hygiene — 3 REAL
#2050	claude-ops — 3 REAL
#2064	repo scripts — 4 REAL
#2065	autonomy — 1 P1 (WACHN)
#2066	work-items — 2 REAL + 1 P1 (TVrP0)
#2070	miro — 2 REAL
#2075	work-items — follow-up: two defects in #2066's own machinery
#2077	review — fanout selects the plugin's own comment
#2085	scripts — comment correction (#2064 shipped an overstatement)
#2092	toolchain — TyjYS, documentary route
#2094	dometrain — --scope + no-op -y
#2095	typos-format — residual classification by token count
#2098	source-control — 4 REAL (T4klg, T0QzB, TzQHE/TzQHF, U4QmR)
#2100	guardrails — 6 REAL across #1275, #1319, #1466, including TzGBZ, the one security finding

No PR from this sweep is open. All seventeen have merged.

TyjYS (#1025) is resolved. Its fix reads as unmerged by SHA — git log origin/main..24d54388
still lists seven commits and git branch -r --contains is empty — but the content is on main:
git diff origin/main...HEAD and git show --format="" 1a26f760 produce byte-identical patches,
and the post-image blob hashes match the commit's own patch headers. A sixth squash-merge phantom.

Adversarial verification of #2098's five discharges — one did not survive

A fresh-context reviewer, briefed to refute rather than confirm, checked each of the threads #2098
claimed. Four discharged cleanly and are resolved: T0QzB, TzQHE, TzQHF, U4QmR. So is
TzaPh — with its rationale corrected, see below.

T4klg is only half discharged and stays open, tracked at #2118. #2098 widened the candidate set
to R*/C* destinations but gated both on the source being 100755, so a copy off a 100644
shebang source is still unreported — while the identical staged content under diff.renames=false
is reported. That config-dependence is the exact failure mode #2098's body says the fix removes.
Reproduced independently before accepting the finding. Worse, #2098 added a test pinning the gap
(case group 21, "a copy whose SOURCE was never executable is NOT reported"), so it now reads as
intended behaviour. I wrote both the fix and that test.

TzaPh's ALREADY-FIXED verdict was wrong — not the outcome, the reasoning. The thread was filed
2026-07-26T00:29:46Z, two minutes after #1264 merged at 00:27:44Z; the fix (72ee66e0, a
separate PR #1440) landed at 02:20:34Z. The defect was live on main for 1h 50m. It resolves as
fixed after filing, not as never a defect. The finding was correct when made.

In flight

• Three unowned REALs, one branch each: session-flow TybdT (feat(session-flow): add find-handoff skill to recover a lost handoff (#976) #1033), discovery TzDQs (feat(discovery): artifact ladder, large-doc fetch, absence enumeration #1260),
  markdown-format TzHVg (fix(markdown-format): skip out-of-tree Markdown when CLAUDE_PROJECT_DIR unset (#972) #1030). Confirmed still unresolved as of this comment.
• The six source threads fix(guardrails): replay a wrapper's chdir in the lease hash probe and rescope skill-reference-verify reconstruction #2100 discharged, TzGBZ included, still awaiting the same adversarial
  pass. None resolved.
• Test-only follow-up for fix(typos-format): classify residuals by token count so a moved finding is not disclosed as a rewrite #2095 — two gaps that merged with it (below).

Deliberately left open

Thread	PR	Why
PRRT_kwDOTCGFQM6TRjqD	1147	C3 per-invocation merge-authority gap — see #2087
PRRT_kwDOTCGFQM6Tzj7k	1252	partially fixed; the rate-limit half is still open, see #1849
TzHGM, TzHGN, the Decision-half of Txl3A	1280, 1357	governance-blocked — they need an ADR ruling, not a code change

Three of the nine non-REAL verdicts are additionally not resolvable because a human commented in
the thread (TzY_h, TzY_i on #1318; TxJRc on #1052).

Why TRjqD was overturned from AFFIRM to CHALLENGE

Two independent audit passes found the same facts and split on classification. The decider was
evidence only the second pass gathered:

• git grep -inE "this-run|per-invocation|single-run|single invocation" origin/main -- plugins/autonomy/
  returns zero hits. The governing guardrail matrix does not acknowledge, authorize, or bound the
  --merge c3-this-run widening at all.
• plugins/source-control/skills/babysit-loop/evals/evals.json:95-99 asserts as EXPECTED that a C3 PR
  merges while the tracked babysit_loop_merge is c2-mechanical — cell unpromoted, merges anyway.
• No ADR: git grep -n "c3-this-run" origin/main -- docs/adr/ is empty.

Leaving the thread open costs one open thread; resolving it wrongly buries a live merge-authority
gap. Asymmetric, so it stays open.

Follow-ups filed

#2078, #2086, #2087, #2088 — every banked item from the sweep is tracked. Two more are outstanding
as test-only gaps that #2095 merged without: a deep single-token fixture (SCALE_N=500 spreads
across two tokens, giving roughly a twentieth of the per-key depth needed to catch the quadratic it
was meant to catch) and a reflow telemetry assertion that never ran, because HOOK_TELEMETRY_SINK
was handed a file path where it requires an executable.

Method findings worth carrying forward

• The merge gate is required_review_thread_resolution, not CI. Nine PRs sat BLOCKED with
  every check green. Watching CI alone burned two monitor cycles.
• A review thread is an argument to evaluate, never an instruction to obey. One reviewer wanted
  gh auth status probed before gh repo view; the code deliberately does the opposite, because
  gh auth status tests every account on every known host and exits 1 if any has a problem — an
  unrelated stale credential would skip the probe and let a bad binding through. Resolved against.
• A fixture that cannot fail proves nothing. An auditor's reproduction returned identical results
  on the pre-fix and post-fix trees. Only a pre-fix control exposed it. Adopted for the rest of the
  sweep, and it changed a verdict.
• A pre-fix control proves sensitivity, not additivity — a distinct trap. Stripping
  exec-bit-check.sh's shared src_mode gate fails the new copy negative and the existing rename
  negative, because both arms read one if. A comment claiming the new case was the only thing
  catching that strip was corrected before merge.
• Prose asserting more than the artifact does is a recurring defect class here. It drew CHALLENGE
  on context-guard twice, again as disk-hygiene's R1, and again on scripts (docs(scripts): stop overstating what the process-substitution gap readmits #2085).
• Self-cited fix SHAs are unreliable — five were pre-squash branch commits absent from main.
  Trace shipped text with git log -S.
• Filed line anchors are stale far more often than not — 9 of 15 in one freshly-checked batch,
  worst off by ~311 lines. The complaint text was authoritative every time.
• Version collisions are the norm. Every branch re-conflicted at least twice; main shipped the
  same version number as a branch three separate times.
• A ledger goes stale inside a single session. fix(toolchain): document the real pyright-absent outcome instead of a skip the runner never performs #2092 merged the toolchain discharge while an
  agent was being briefed to open a PR for it. Re-read PR state from the API before dispatching, not
  from notes.
• Check who is already on a PR before dispatching to it. A duplicate agent sent to threads
  another session was working is the most likely cause of fix(guardrails): replay a wrapper's chdir in the lease hash probe and rescope skill-reference-verify reconstruction #2100's four threads being resolved with
  no verifier verdict on record. Dispatch collisions are not just wasted tokens; they can satisfy a
  merge gate.
• check-contract-clause-coverage.py does not enforce that a canonical clause and its
  restatements agree. Reverting either side alone leaves the gate green — it counts tagged
  surfaces, it does not compare text. Any discharge argument leaning on that gate as proof both
  halves shipped is unsupported; the pairing has to be read.
• A verifier attacking the orchestrator's own work is where the value is. The single finding
  that did not survive this sweep's verification pass — T4klg, now fix(source-control): exec-bit-check skips a copy destination whose source was never executable #2118 — was a gap in a fix I
  wrote, locked in by a test I added. Rationale withheld and instructed to refute is what surfaced
  it.

Evidence-quality notes

Two branches in this set carry weaker evidence than the rest. Recorded rather than smoothed over.

• fix(claude-ops): key lane markers in the target repo, and scope stop-gate arming to the installs that asked #2050 merged on the orchestrator's own verification plus green CI. Its agent never reported a
  fresh-context verifier verdict.
• fix(guardrails): replay a wrapper's chdir in the lease hash probe and rescope skill-reference-verify reconstruction #2100 (merge commit 1cbb4bc9; 1a8333ff is the branch head, not the merge) merged with all
  four of its own review threads resolved while its fresh-context verifier was still running, so no
  verifier verdict is on record for those four. The owning session states it resolved nothing; the
  likeliest actor is a duplicate agent dispatched onto the same threads by mistake and killed
  shortly after. resolvedBy cannot discriminate — every agent here acts under one account — and
  GraphQL exposes no resolvedAt, so this is not provable from the API. The verdict is still being
  collected and becomes a follow-up if it hits. The six source threads that PR discharges are held
  open pending independent verification, so they are unaffected.

[kyle-sexton]

Addendum, 2026-08-10. Supersedes the "In flight" and "Evidence-quality" sections of the comment
above. Posted as a new comment rather than an edit so the earlier state stays readable.

The verification pass found three defects in the fixes, and one in itself

Two sessions cross-verified each other's work — each independent of the code it was checking, though
neither fresh-context. That arrangement caught errors in both directions.

Found in work I had shipped or approved:

• T4klg (feat(source-control): harden /commit — exec-bit backstop, trailer SSOT, composition resilience #1590) — fix(source-control): discharge four stranded P2 review findings #2098's discharge was half a fix. C* destinations are admitted only when
  the source was 100755, so a copy off a 100644 shebang source is still unreported — while the
  identical staged content under diff.renames=false is reported. That config-dependence is the
  exact failure fix(source-control): discharge four stranded P2 review findings #2098's body says it removes. Worse, fix(source-control): discharge four stranded P2 review findings #2098 added a test asserting the gap as
  intended behaviour. Thread left open; fix(source-control): exec-bit-check skips a copy destination whose source was never executable #2118 filed.
• TzGBZ (fix(guardrails): block --force-with-lease forms that state no expected value #1275) — I resolved the security thread and was wrong to. My round of 16 cases all
  behaved. The session that wrote the fix then handed over the forms they had not tried, and
  several bypass origin/main today: the payload's cwd is never read (no wrapper needed at all —
  block-noncanonical-commit reads it, this guard does not), and env -S splices the chdir past
  the parser. Thread reopened; fix(guardrails): block-dangerous-git still clears an unsafe --force-with-lease via payload cwd and env -S #2124 filed with the reviewed design.
• fix(markdown-format): stop reading a missing git as "file is out of tree" #2121 — the nested-file finding against my own PR was correct, and two further review
  findings on it were also real. Config discovery stops at hook::repo_root, which without git
  returns the edited file's own directory, so a repository whose markdownlint config sits at the
  root stopped linting everything below it.

Found in the verification itself: a claimed second relative -C bug did not survive retest and
was withdrawn; a telemetry/slow-sink failure read as a defect turned out to be load
(#2125, filed as a flake note, not a bug).

The method finding, which outlasts these fixes

Fresh-context is the weaker property; independence is the one that matters. A fresh verifier can
only attack what its brief tells it to, so the brief's blind spots become the verification's blind
spots — and the author writing that brief is precisely the person who cannot see them. Three
separate fresh-context verifiers were briefed on TzGBZ; none found the .cwd hole. What found it
was a reviewer who knew the code and handed over the negative space of their own testing. That
input cannot exist in a fresh context by construction.

Practical consequence: a verification brief should carry the author's untested space as a
first-class section, not just the findings.

One defect explains almost every error, in five costumes

Nobody confirmed the fixture reached the thing under test.

1. env FOO=1 -C <dir> git push read as a live bypass — GNU env stops option parsing at the first
   NAME=VALUE, so rc=127 and git never ran. Caught before filing.
2. Two fixtures that errored, read as fail-closed.
3. Case 12's oracle — a correct answer could not pass it either, because two expectations were
   gated on branches the fixture never reaches.
4. repo21 — asserted the defect as intended behaviour.
5. A scale benchmark whose hunk lines tripped a span cap after 40 anchors, so it measured the capped
   path and never the quadratic one, reporting flat numbers that looked like success.

Countermeasure, in its general form: a liveness column in every probe table — not "prove the
command executed" but "prove the fixture exercised the path under test." Costume 5 is why the
weaker phrasing is not enough.

Open, and needing a human rather than more verification

Two PRs merged with their review threads resolved by an actor neither working session can account
for — #2100 and #2115. resolvedBy is a single shared account so it cannot discriminate, and
GraphQL exposes no resolvedAt. Both sessions have stated their complete write surface; neither
called resolveReviewThread. Nothing bad shipped — #2115's verdict was PASS on hard evidence — but
the thread-resolution merge gate is being cleared by something unidentified, and until that is
understood an open thread is not a reliable hold. Holds are now stated in thread text as well.

Current state

Item	State
#2100, #2115	merged
#2121	three findings fixed, threads answered, CI in flight
#2116	oracle unverified — the two-sided grading has not returned, and a half-answer is not an answer
TzGBZ	reopened, hold recorded, #2124
T4klg	open, #2118
T0gfE, TxHZS	pinned by discriminating pre/post failures
TxHZQ, TxHZU	no discriminating evidence yet
T0gfF	suggestive only — a suite timeout, not a controlled measurement, and this host has already produced false timing conclusions

Filed this pass: #2113, #2118, #2119, #2124, #2125.