chore(work-issues): fold the 2026-10-13 run's lessons into claim, verify and ship

.claude/skills/work-issues/references/claim.md

@@ -43,10 +43,11 @@ gh issue comment <n> --body-file "$SCRATCH/standdown-<n>.md"
 ```
 
 For EACH issue you start — PROMOTING a QUEUED one included — first re-check
-`gh issue view <n> --json state`, §2's open-PR `files` query, and §3's premise
-check on CURRENT `origin/main` — in a call BEFORE the claim, never chained with
-it: after TRIAGE a peer can close a queued issue, or open a PR holding its
-lines, before its turn (#3979). Then:
+`gh issue view <n> --json state`, §2's open-PR `files` query and "working on
+this" claims on OTHER issues (extensions and this run's own too, #4857), and
+§3's premise check on CURRENT `origin/main` — in a call BEFORE the claim, never
+chained with it: after TRIAGE a peer can close a queued issue, or open a PR or
+claim holding its lines, before its turn (#3979). Then:
 
 ```bash
 gh issue comment <n> --body "Working on this in PR/branch <ref> — touching <files>. \
@@ -95,9 +96,8 @@ stand down naming the winning branch and pick another issue, unasked. Escalate
 when timestamps cannot settle it (go-to-k/cdkd#1446).
 
 **A QUEUED comment IS a claim, and its `createdAt` is what the tie-break reads.**
-Re-read the thread to the END before the tie-break. A missing branch is
-not absence (signals show LIFE only, §9); a claim has no TTL, so one you think
-dead goes to arbitration.
+Re-read the thread to the END before the tie-break. A missing branch is not
+absence (signals show LIFE only, §9); no claim has a TTL: escalate a dead-looking one.
 
 **Nothing makes the tie-break's LOSER re-read**: re-read the claims before you
 PUSH; if yours is later, stand down even with code written.

.claude/skills/work-issues/references/ship.md

@@ -10,17 +10,17 @@ CLEAN (one lane at a time), then run `/run-integ` and
 auto-mode classifier can refuse the PARENT's calls too: any refused call goes
 to `AskUserQuestion`, never a reworded retry (go-to-k/cdkd#4059); only
 `/run-integ` step 9's one retry (marker set or ledger push) is exempt.
-While that tree is busy with a follow-up branch, ship from a SECOND sibling
-on the PR's branch
+Ship from a SECOND sibling on the PR's branch while that tree holds a follow-up
 (`git -C <MAIN_CHECKOUT> worktree add <MAIN_CHECKOUT>/.claude/worktrees/<name> <branch>`),
 running BOTH the integ and the merge there (the marker is per tree). A fix
 round DURING that tree's integ edits a `--detach` sibling, pushes
 `HEAD:<branch>` (bash reads `verify.sh` live), and fast-forwards that tree
 before the marker, re-running an integ the fix touches (#4561). Never two
-lanes' integs or merges at once; a turn blocked on a fixture fix yields to
-the next merge-ready lane. A turn holding for HOURS: tell the user
-unasked its ETA and queue. A subagent running a turn's integs, marker
-and ledger must block until EVERY run ends and report ONCE.
+lanes' integs or merges at once, nor a merge during another lane's integ: it
+can move `main` under that batch (#4864). A turn blocked on a fixture fix
+yields to the next merge-ready lane. ANY turn holding for HOURS tells the user
+unasked its ETA and queue. A subagent running a turn's integs, marker and
+ledger must block until EVERY run ends and report ONCE.
 
 - The `integ-destroy` marker is read from the tree the command runs in, so a
   merge from the main tree consults the WRONG store (go-to-k/cdkd#2363). Its

.claude/skills/work-issues/references/verify.md

@@ -12,20 +12,21 @@
 
 ### 8-b. Integ ordering vs review rounds and rebases
 
-**Run the integ LAST — after the final edit to any `integ-destroy`-scoped file.**
-A lane cannot touch AWS, so only the parent finds an AWS-side rejection: run a
-NEW arm once EARLY too, as soon as the lane has written it and between review
-rounds (never during a `src` fix round); that run sets no marker and is no
-§8-d proof — a new arm's first AWS contact can reject its shape (#4618).
-Sequence: dispatch reviewers → (new arm: early run) → apply
-EVERY finding, nits included → rebase → integ → marker. A UNIT-test-only fix
-round may overlap the parent's integ in the lane's tree. Brief the lane: no build, `/check` or `/verify-pr` there (a
-build rewrites the live fixture's `dist/`), no rebase or install, never the
-live fixture's directory, and stage only its own test files, since the
-integ's ledger row lands in the same tree (#4302). A lane REPORTS
-`mise exec -- markgate status integ-destroy --explain`'s line, never "integ
-not needed": comment-only edits count (#3873). Bare `markgate` can resolve to
-a stale PATH copy that dies `unknown hash "diff"` (#4477).
+**Run the integ LAST — after the final edit to any `integ-destroy`-scoped
+file.** A lane cannot touch AWS, so only the parent finds an AWS-side rejection:
+run a NEW arm once EARLY too, as soon as the lane has written it, between review
+rounds (never during a `src` fix round); that run sets no marker and is no §8-d
+proof — a new arm's first AWS contact can reject its shape (#4618). Sequence:
+dispatch reviewers → (new arm: early run) → apply EVERY finding, nits included →
+rebase → integ → marker. A UNIT-test-only fix round may overlap the parent's
+integ in the lane's tree. Brief the lane: no build, `/check` or `/verify-pr`
+there (a build rewrites the live fixture's `dist/`), no rebase or install, never
+the live fixture's directory, and stage only its own test files (the integ's
+ledger row lands there, #4302). A round overlapping the §8-d probe (clean tree;
+it restores `src` from `HEAD`) edits in scratch until the parent frees it. A
+lane REPORTS `mise exec -- markgate status integ-destroy --explain`'s line,
+never "integ not needed": comment-only edits count (#3873); bare `markgate` can
+be a stale PATH copy dying `unknown hash "diff"` (#4477).
 The gate's `hash: diff` stales on a rebase only when main changed a scoped
 file THIS branch changes too, so a set marker on a MERGEABLE PR needs no
 rebase (`mise exec -- markgate status`) — unless main changed code the fixture
@@ -183,9 +184,8 @@ stacks. After teardown, sweep for orphans it cannot reach (`/aws/lambda/*` log
 groups, RETAIN resources, Secrets in recovery, KMS keys pending deletion), then
 run AGENTS.md's leftover check, which the `deployments/` store survives.
 
-**`/run-integ` records `integ-destroy`, a marker gate on `gh pr merge`; the
-`main` ruleset's checks are the other merge condition.**
-`/verify-pr` and `/review-pr` record nothing; run them anyway.
+**`/run-integ` records `integ-destroy`, which `gh pr merge` needs beside
+`main`'s ruleset checks**; `/verify-pr` and `/review-pr` record nothing — run them anyway.
 
 **The independent review round is the ORCHESTRATOR's; a LANE's own reviewers
 never substitute for it** (go-to-k/cdkd#2383) — they are its children and