Integrate contributor PRs #48-#58: eleven doctrine rules across three skills

[F-e-u-e-r]

Integrates the eleven open contributor PRs (#48–#58, all by @firaen22) into one branch via ascending-order merges, so the set lands atomically and each source PR is marked merged by reachability.

What the eleven PRs add

All eleven are doctrine rules mined from the contributor's private incidents, cited as shape per the README covenant's second branch, each carrying an in-body unprobed marker plus a Provenance entry recording the probe debt.

delegation-and-review (6 PRs, +307 lines in SKILL.md, 2 new references/ files):

• delegation-and-review §3: freeze the tree for a read-only review dispatch #48 — §3 settled-tree review: a fresh-context critic wave must read a settled tree; verdicts bind only the exact state whose immutability was enforced. Protocol body in references/settled-tree-review.md (baseline over the protected read set, enforced-copy-or-frozen-tree surfaces, two return comparisons, recovery).
• delegation-and-review §2: scope a sweep by every generator, not one grep pattern #50 — §2 sweep fields: a find-and-fix-every-instance sweep splits scope by what the invariant is (textual vs behavioral); search scope covers every generator surface with a mandatory producer/effect axis; the proof gate is the observed effect at every inventoried surface, and a zero-hit grep gates only the textual branch.
• delegation-and-review §1: isolated worktrees do not isolate ports #53 — §1 port contention: isolated worktrees share a port namespace; a displaced server plus a static port reference silently reaches a sibling's server with all-200 responses. Defenses: unique fixed port with fallback disabled, or runtime derivation; identity check last, non-mutating, through every consumer path.
• delegation-and-review §1: a pinned model string does not pin behavior #54 — §1 pinned-string drift: a pinned model string does not pin behavior; date-stamp behavioral measurements and re-probe at decision time; probe unavailable → route as unguarded.
• delegation-and-review §2: recurring sweeps carry ledgers (prior fixes + refuted classes) #55 — §2 recurring-sweep ledgers: a recurring dispatch packet names a stable campaign id and a durable ledger file (prior fixes / refuted classes / open / unresolved) reconciled against enumerated prior-round reports; the ledger is dedup context, never authority. Lifecycle in references/recurring-sweep-ledgers.md.
• delegation-and-review §1: labels are routes, listings are claims #56 — §1 labels are routes, listings are claims: a lineup listing is not callability (verify with a fixed trivial prompt through the same wrapper and read its route report); a wrapper's model string is not a provider ID (resolve the alias mapping from the wrapper's own config before any namespace crossing).

operational-rigor (2 PRs, +84 lines in SKILL.md, +125 in references/external-systems.md):

• operational-rigor §4: verify a scheduled process's side effects, not its self-report #49 — §4 scheduled-process rule: a recurring schedule's own "completed" report is not evidence its side effects landed — verify at the destinations, attributed to the invocation. Full arming/audit protocol added as the scheduled-process entry in references/external-systems.md (gates-first authorization, channel inventory, emission-positive verification, overlap/lock traversal, absence alarm outside the failure domain, inspection-only arm).
• operational-rigor §4: a check's name is not its coverage #57 — §4 check-name rule: a check's name is not its coverage — trace a cited check to its pass/fail oracle at the cited run's revision and assert only what the trace established.

skill-authoring (3 PRs, +155 lines in SKILL.md):

• skill-authoring §5: keyword-grep absence is not absence — read the section before concluding not-covered #51 — §5 keyword-grep absence is not absence: the dup-check greps the target and its siblings with the concept's name plus alternates, lists real outlines, reads candidate sections in full, and records the result line; the reviewer re-runs the searches against pre-addition text.
• skill-authoring §6: name the deployment runtime before the review concludes #52 — §6 deployment-runtime review gate: verify the target runtime before a skill review concludes; sweep for machine-bound assumptions (paths, launchers, host identity, clock/locale, tool prefixes, execution deps); each keeps a portable form, a target-scoped dispatch, or an in-file label.
• skill-authoring §3: capability-negative claims rot the worst #58 — §3 capability-negative claims rot worst: a stale "no such flag" fails silent and steers every later session into a degraded path; pin every capability-negative to the version observed and re-verify on version change.

Merge mechanics

Ascending merges #48 → #58, preserving each PR's commits. Conflicts were confined to the three files' Provenance tails plus one section-body adjacency (delegation-and-review §1, where #54 and #56 add bullets at the same anchor); every resolution keeps both sides in merge-order chronology. The closing re-check line of delegation-and-review/SKILL.md takes #54's expanded form (a superset of the old line, adding hosted-endpoint behavioral claims to the re-verify list).

Checks: python3 .github/checks.py passes on the merged tree; zero-width/bidi sweep over all added diff lines is clean; no conflict markers remain.

Current diffstat vs main: +856/−11 across 6 files (originally +787/−10 at integration, before the three review-fold commits; .gitignore untouched).

Cross-model review

Three rounds, two reviewers per round, both discovered and dry-run-verified at session time: grok-4.5, reasoning-effort high (grok CLI 0.2.101) and gpt-5.6-luna, reasoning effort ultra (codex-cli 0.144.6; banner route report confirmed the model each run). Round 1 reviewed the raw integration; rounds 2-3 reviewed with a cumulative disposition ledger in the packet. Full verdicts and dispositions in the review-trail comment below.

• r1: grok FIX 1,2,3 (+5 nits) / codex FIX 1-7 (+2 nits) → 12 items fixed in ae96cd7, 2 rejected with counter-evidence.
• r2: grok FIX 1 (+2 nits) / codex FIX 1-5 (+4 nits) → 9 items fixed in 8a94407, 3 rejected with counter-evidence. No prior rejection re-raised.
• r3: grok FIX 1,2 (+4 nits) / codex FIX 1-5 → 9 items fixed in ed47089 (two of codex's scoped to their valid core), 2 rejected with counter-evidence (one a re-raise of PR skill-authoring §6: name the deployment runtime before the review concludes #52's standing rejection, one directly contradicted by the reference text it attacked).
• Merge-integrity, covenant, and safety surfaces: clean from both reviewers in all three rounds.
• Terminal condition: the round cap (cross-model-review §4) is reached; every r3 item carries a disposition (fixed / rejected-with-recorded-counter-evidence); remaining valid findings across r2-r3 were exclusively polish of r1-r2 fix wording, never the eleven rules' substance. Residual risk: the r3 fold itself (one-sentence-level wording fixes) shipped without a fourth round, per the cap.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx

[F-e-u-e-r]

Cross-model review trail (combined gate over the eleven-PR integration)

Reviewers (discovered + dry-run-verified at session time, per cross-model-review §1): grok-4.5 / reasoning-effort high (grok 0.2.101) and gpt-5.6-luna / reasoning effort ultra (codex-cli 0.144.6; the exec banner's model:/reasoning effort: lines confirmed the route on every run). Packets were self-contained (combined diff + full post-merge text of all six touched files + seven-surface rubric); rounds 2–3 additionally carried the cumulative disposition ledger. Verdict files retained under the session's tmp (combined/r{1,2,3}/{grok,codex}/verdict.md).

Round 1 (raw integration, 5158b3c)

• grok: FIX 1,2,3 + nits 4–8. Codex: FIX 1–7 + nits 8–9.
• Fixed in ae96cd7 (12): settled-tree trigger broadened to write-capable waves (both files); labels pass = answer AND route report; silent-fallback → route unverified; callability session-scope self-standing; port identity-check ordering rebuilt (check after repairs, cleanup after check, non-disruptive markers only — the ✅ example had cleanup before the nonce check it needed); campaign-ledger path pinned to the dispatching side's repo; behavioral sweep gate requires the producing condition driven true; capability-negative restructured (trigger/probe/unknown-not-absent/Done/✅); dup-check joins router/entry files; hunt-log vs campaign-ledger naming; field title → dispatches; keyword-grep ✅ ends on duplicate-found; §6 review-record decoupled from §7's word-diff artifact.
• Rejected with counter-evidence (2): codex F7 (the §6 deployment-runtime gate governs §6 usability reviews and the added rules are pure instruction text under that rule's own runtime-agnostic branch — also a packet-limitation shape: absence from the packet is not absence); codex n9 (the rule already lists the wrapper's request trace among mapping sources).

Round 2 (ae96cd7)

• grok: FIX 1 + nits 2–3 (and confirmed 11/12 r1 fixes hold; rejected items not re-raised). Codex: FIX 1–5 + nits 6–9 (all new items; no re-raise).
• Fixed in 8a94407 (9): the sweep ✅'s leftover "round ledger" noun (the one r1 fix that had missed its example — grok caught it); port defenses numbered (1)/(2); capability Done split by branch; §1 parallel-writers line carries the settled-tree carve-out; pinned-string re-probe requires route attribution first; hosted-model behavioral negatives scoped out of version-pinning to the pinned-string rule; settle-by-commit stages exactly the content under review; PRIOR-FIXES narrowing scoped to re-flagging only; explicit ledger write-back transitions.
• Rejected with counter-evidence (3): codex F4 (sweep packet scope vs whole-project twin search — different duty-bearers; the worker's reported gap is the disclosed limitation and the dispatcher owns the whole-project duty, escalation path already in the text); codex n6 (the miss-is-costly loop runs all axes each round, so a new producer category makes the round non-empty); codex n9 ("not found" binds to the fact under dup-check; hit-bearing sections are force-read).

Round 3 (8a94407 + ledger)

• grok: FIX 1,2 + nits 3–6 — both must-fixes were residue of the r1/r2 folds themselves (the labels fix's unverified-route block would have bricked wrappers with no attribution channel by design; the pinned-string Done/✅ had not absorbed the r2 attribution clause). Codex: FIX 1–5.
• Fixed in ed47089 (9): the two grok must-fixes (channel-class split: channel-present-but-unattributed blocks dispatch, no-channel-by-design yields a recorded reachability-only pass; Done/✅ now carry attribution); grok nits 3–6 (carve-out names both duties; check-name trigger covers any cited property; port recovery moved inside the address-in-use branch; reference cites the field's actual title); codex F2 scoped to its valid core (attribution = a report of what ANSWERED; a requested-slug echo is configuration); codex F3 scoped (probe example exercises the claimed-absent capability; runtime-scoping stays §6's gate — no duplicate home); codex F4 (regressed-prior-fix transition: re-flagging spawns a new OPEN finding, historical entry stays with a pointer).
• Rejected with counter-evidence (2): codex F1 (mutate-and-restore on an unenforced copy is the exact case the reference already rules: "an unenforced reader can mutate and restore without a trace; no endpoint comparison proves the read stayed clean" → the review runs provisional, never a clean gate pass); codex F5 (re-raise of PR skill-authoring §6: name the deployment runtime before the review concludes #52's two-strike standing rejection on gate placement — §1's placement line already gates authoring start and §6 is the named enforcement backstop, adjudicated in that PR's own per-PR review; the frontmatter loader is pre-existing structure outside this diff).

Terminal condition and residual risk

Cross-model-review §4 caps the loop at 2–3 rounds; round 3 is the cap. Every round-3 item carries a disposition (fixed, scoped-and-fixed, or rejected with recorded counter-evidence). The valid must-fixes of rounds 2–3 were exclusively wording-residue of earlier folds — the eleven rules' substance drew no new confirmed defect after round 1 — and the merge-integrity, covenant, and safety surfaces were clean from both reviewers in all three rounds. Residual risk, stated: the round-3 fold (sentence-level wording fixes, each verified in place against the finding it answers) ships without a fourth review round, per the cap; and the per-PR treadmill state recorded in the prior session (pending per-PR v-folds for #48–#58) is superseded by this combined gate rather than individually folded — the combined rounds re-covered those files in their merged form.

Per-invocation authorization for this merge: the owner's instruction to summarize the set, cross-check it, and "may be using a new pr to commit, push and merge it with admin access".

[F-e-u-e-r]

Gate continuation (cap lifted by owner): rounds 4–8 ran post-merge with the same reviewer pair and cumulative disposition ledgers; their folds landed via PR #60 (merged 3ba02a3). Trajectory: r4 grok 1 + codex 8 must-fix → r5 1+2 (one shared) → r6 1+PROCEED → r7 PROCEED+1 → r8 double PROCEED, zero findings — terminal. 15 fixed, 2 rejected with counter-evidence, no accepted re-raises. Full per-round dispositions in #60's body.