diff --git a/skills-staging/2026-07-30-campaign-ops/MANIFEST.md b/skills-staging/2026-07-30-campaign-ops/MANIFEST.md
new file mode 100644
index 0000000..608b009
--- /dev/null
+++ b/skills-staging/2026-07-30-campaign-ops/MANIFEST.md
@@ -0,0 +1,89 @@
+# MANIFEST — campaign-ops library (2026-07-30)
+
+Project skill library distilled from the 2026-07-29/30 contributor-PR
+campaign session (transcript: `chat-history.md`, local untracked file;
+public session id `session_01WMbbZQ8ZMMsNRGyV2nvJR8`, stamped in the
+session-authored commits' trailers — e.g. `f529014`; the
+GitHub-generated merge commits carry no trailer). Staged, NOT
+installed: nothing under `skills-staging/` auto-loads. Every
+repo-facing claim was re-verified against the repository on
+2026-07-30; what could not be verified is in `UNCERTAINTY.md`, not in
+the skills.
+
+Covenant scope: the README House covenant's probe/marker regime
+governs the pack roots (`skills/`, `design-pack/skills/`). This staged
+library ships under the retiring-architect MANIFEST/UNCERTAINTY regime
+instead — its rules are verified operating records, not covenant-bound
+pack doctrine, and this delivery is not precedent for shipping pack
+rules without markers.
+
+Delivery note: `skills-staging/` is gitignored by owner decision
+(`92314a6`). This library is delivered by explicit force-add on the
+owner's instruction for this delivery only; `.gitignore` itself is
+untouched. Once tracked, these files behave like any tracked file.
+
+One line per skill — loading condition ← session evidence ← repo
+evidence ← verification status:
+
+- **contribution-gate-playbook** — contributor PRs open + owner ask to
+  review/integrate/merge, or a live combined-*/integrate-* branch ←
+  the session's three campaigns end to end ← PR #90/#93/#97 bodies,
+  merge commits `a148180`/`92077a7`/`79ca49c`, `gh pr list/checks`,
+  reviewDecision fields, #91 createdAt vs #90 mergedAt, commits
+  `f529014`/`814d116` ← **verified** (cadence wording
+  partially-verified: UNCERTAINTY 2; comments: UNCERTAINTY 7).
+- **gate-adjudication-and-folds** — a lens verdict landed and needs
+  fold/reject/stop; opposite-fix convergence; non-converging lens;
+  runtime-semantics finding ← r1–r11 fold/reject decisions, the #91
+  precedence adjudication, the #96 execution chain ← commits
+  `6fa6154`/`5fa241b`/`83a038d`/`1560b97`, PR #90/#93/#97 bodies,
+  #91/#96 evaluation comments, cross-model-review §§3–4 ←
+  **verified** (the rejection specifics, the r8 pre-declaration, and
+  the fold-introduced attribution are transcript-recorded and labeled
+  so in place; the `→` count was re-run in-repo 2026-07-30).
+- **doctrine-change-conventions** — editing rule text in
+  skills/*/SKILL.md heading to commit/merge; Provenance work; a rule
+  with no probe ← integration fixes, covenant handling, width/glyph
+  rejections, no-bump observation ← README House covenant paragraph,
+  issue #92 body, `.github/checks.py` (read in full: no width check;
+  sweep scope), `git tag -l`, version badge lines, `32c6929`,
+  `f529014`/`814d116` ← **verified**.
+- **repo-boundaries-and-sync** — about to stage/commit; post-merge
+  sync; unexpected diff-loop output; untracked file at repo root ←
+  the session's staging discipline and closing privacy flag ←
+  `.gitignore` + `git check-ignore`, `git ls-files` top-level,
+  `92314a6` message, live-copy diffs run 2026-07-30, README
+  Maintainer Notes, evaluation comments ← **verified**
+  (no-owner-quotes directive is session/memory-recorded, corroborated
+  by the read comments' shape; skill-vetting install intent:
+  UNCERTAINTY 4).
+- **failure-archaeology** — about to hand-roll a CI wait, ship
+  runtime-semantics prose, adopt/delete leftover branches, or re-open
+  settled items ← the session's dead ends and residue ← same commits
+  and bodies as above plus `git branch -vv` and
+  `git branch --merged main` observed 2026-07-30 ← **verified**
+  (entry 1's loop internals are transcript-only — the loop never
+  landed anywhere; the entry carries that label in place).
+
+Claims rejected or downgraded during verification (full list with
+evidence in `UNCERTAINTY.md`):
+
+- "bump the version every substantive PR" (2026-07-12-era rule) →
+  **historically-valid**; contradicted by observed practice #84–#97
+  (0.1.16 unchanged; tags end at alpha-0.1.2).
+- The 2026-07 reviewer lineup recorded in the older local library
+  (gpt-5.5 xhigh + grok-4.5 max) → **historically-valid**; the
+  campaign lineup differs and is itself date-stamped volatile.
+- Session-memory phrase "evals/round4 … COMMITTED" → not supported by
+  the repo (`evals/` is gitignored, zero tracked files); treated as
+  loose wording, recorded in UNCERTAINTY 8.
+- Installed-copies "cmp-verified @ merge SHA" → **verified with an
+  exception the session itself disclosed**: `skill-vetting` absent
+  from both installs (expected state; UNCERTAINTY 4).
+
+Review status: three fresh-context reviews (factual / doctrine /
+usability) ran 2026-07-30 over these files; every BLOCKING and
+IMPORTANT finding was applied in place before delivery, and the
+dispositions are summarized in the delivering PR's description (the
+PR that adds this directory). Unresolved and evidence-dependent items
+live in `UNCERTAINTY.md`.
diff --git a/skills-staging/2026-07-30-campaign-ops/START-HERE.md b/skills-staging/2026-07-30-campaign-ops/START-HERE.md
new file mode 100644
index 0000000..c91fdcb
--- /dev/null
+++ b/skills-staging/2026-07-30-campaign-ops/START-HERE.md
@@ -0,0 +1,83 @@
+# START HERE — campaign-ops library (2026-07-30)
+
+What this library is: the distillation of the 2026-07-29/30 maintainer
+session that integrated contributor PRs #85–#89, #91, and #94–#96 into
+main through three combined gates — PR #90 (merge `a148180`), PR #93
+(`92077a7`), PR #97 (`79ca49c`) — and opened the probe-debt tracker,
+issue #92. The engineering core is not the doctrine content those PRs
+carried; it is operating the contributor-PR pipeline (combined branch →
+tri-lens cross-model gate → owner-authorized merge → post-merge duties)
+without breaking the house covenant, the privacy boundaries, or the
+published-record conventions.
+
+Authored 2026-07-30 by the retiring architect from the session
+transcript (`chat-history.md`, a local untracked file at the repo root)
+with every repo-facing claim re-verified against the repository on
+2026-07-30. Companion files: `MANIFEST.md` (per-skill evidence),
+`UNCERTAINTY.md` (everything NOT settled — read it before trusting a
+claim near its topics).
+
+## Canonical-source map (who wins on disagreement)
+
+| Topic                     | Canonical source                        |
+|---------------------------|-----------------------------------------|
+| Cross-model gate doctrine | `skills/cross-model-review/SKILL.md`    |
+| Authoring/covenant method | `skills/skill-authoring` + README       |
+|                           | ("House covenant" paragraph)            |
+| Execution discipline      | `skills/operational-rigor`              |
+| Practice records          | PR #90/#93/#97 bodies, per-PR           |
+|                           | evaluation comments, issue #92          |
+| This library              | repo-specific operating layer, dated    |
+|                           | 2026-07-30 — canon and records win      |
+
+## Current-state triage (as observed 2026-07-30)
+
+Facts a fresh session must not misread as its own breakage:
+
+- At authoring time: 0 open PRs; 1 open issue (#92) — #92 is DESIGNED
+  to stay open while any `unprobed` marker exists; it is a queue, not
+  a stale ticket. This delivery itself adds issue #100 (transcript
+  disposition; UNCERTAINTY item 1) and, while open, the delivering PR.
+- The repo root carries UNTRACKED private session transcripts
+  (`chat-history.md`, `security-enhancement.md`). Never stage them;
+  see `repo-boundaries-and-sync`. Owner disposition pending (#100).
+- `skills-staging/` is gitignored by owner decision (`92314a6`);
+  this library was force-added on an explicit owner instruction —
+  that delivery is not license to publish other ignored zones.
+- The README per-skill sync loop currently errors on `skill-vetting`
+  (present in `skills/`, absent from both live installs) — a known
+  pre-existing state, not drift you introduced.
+- Many local branches are residue of merged or abandoned campaigns —
+  do not adopt-and-finish them (see `failure-archaeology`).
+
+## Reading order
+
+1. `contribution-gate-playbook` — the pipeline end to end.
+2. `gate-adjudication-and-folds` — judging lens verdicts.
+3. `doctrine-change-conventions` — what any doctrine diff satisfies.
+4. `repo-boundaries-and-sync` — what may be staged, synced, published.
+5. `failure-archaeology` — dead ends and residue from this session.
+
+## Relationship to the older local library (2026-07-12)
+
+`skills-staging/` also holds a 12-skill library authored 2026-07-12
+(local-only, unpublished, gitignored — its existence is public via
+`92314a6`'s commit message). Overlap map, verified 2026-07-30:
+
+- Partially superseded by this library: cross-model-gate-ops,
+  release-and-publish, repo-state-and-sync (verified stale examples
+  in `UNCERTAINTY.md` item 5: the 2026-07 reviewer lineup, the
+  bump-every-PR rule, the `README.zh-TW.md` filename).
+- Old editing-skills-and-docs overlaps the new
+  doctrine-change-conventions — and is itself stale (its description
+  counts "the eight SKILL.md files"; `skills/` now has ten).
+- NAME COLLISION: the old library and this one BOTH carry a skill
+  named `failure-archaeology` (different content). Never install both
+  libraries into one skills directory without renaming one.
+- Everything unlisted (architecture-contract, build-and-env,
+  config-and-flags, debugging-playbook, diagnostics-and-tooling,
+  eval-operations, validation-and-qa) has the old library as sole
+  source — verify each claim against the repo before relying on it.
+
+Precedence for overlapping topics: canon skills first, then the newer
+dated statement, then the older library.
diff --git a/skills-staging/2026-07-30-campaign-ops/UNCERTAINTY.md b/skills-staging/2026-07-30-campaign-ops/UNCERTAINTY.md
new file mode 100644
index 0000000..fd0d081
--- /dev/null
+++ b/skills-staging/2026-07-30-campaign-ops/UNCERTAINTY.md
@@ -0,0 +1,132 @@
+# UNCERTAINTY — campaign-ops library (2026-07-30)
+
+Findings that are useful but cannot be stated authoritatively. Each
+entry: claim / where it appeared / repo evidence / status / why
+uncertain / what would resolve it / issue created? Every entry ends in
+a safe default a zero-context reader can act on.
+
+## 1. Untracked private transcripts at the repo root
+- Claim: `chat-history.md` and `security-enhancement.md` (both full
+  session transcripts — private content) sit untracked in the public
+  repo's working tree; any stage-everything command would publish
+  them.
+- Appeared: the session's closing report flagged
+  `security-enhancement.md` to the owner and asked for a disposition;
+  no answer had arrived by session end. `chat-history.md` appeared
+  after that session (it is that session's own transcript).
+- Repo evidence: both present and untracked on 2026-07-30
+  (`git status --short`); never in any commit; `92314a6`'s message
+  names the exact threat ("a stray `git add -A`").
+- Status: verified (existence/exposure risk); **user-must-provide**
+  (disposition).
+- Resolves it: owner picks a `.gitignore` pattern or relocates the
+  files out of the repo. `.gitignore` is outside this delivery's
+  allowed paths.
+- Issue: yes — #100 (created 2026-07-30, single consolidated issue):
+  https://github.com/F-e-u-e-r/opus-pack/issues/100
+- Safe default: never stage them; pathspec-only adds; never quote
+  their content anywhere public.
+
+## 2. Exact wording of the standing gate cadence
+- Claim: grok-4.5 high + gpt-5.6-luna ultra every round; gpt-5.6-sol
+  max on "every 3rd round + pre-commit" (session-memory phrasing).
+- Repo evidence: #90's body "gpt-5.6-sol max from r3" (fold commits
+  show sol at r3 then r6–r11, the terminal candidates); #93's body
+  "r2 pre-commit (grok + luna + sol-max)"; for #97 the cadence line
+  lives in its constituents' evaluation comments ("gpt-5.6-sol max
+  from r2"), not the body. All consistent with the phrasing; none
+  states it as a standing rule.
+- Status: **partially-verified** (pattern verified across three
+  campaigns; the governing rule text is owner-set, session-recorded).
+- Resolves it: owner confirmation, or the next campaign PR body.
+- Safe default: run sol on every 3rd, every pre-commit, and every
+  terminal-candidate round; when unsure, include sol — an extra lens
+  costs a round, a missing lens costs the gate's meaning.
+
+## 3. Merge-bypass mechanics ("admin merge")
+- Claim: campaign merges used the owner's admin authority to bypass
+  required review.
+- Repo evidence: #90/#93/#97 all show reviewDecision REVIEW_REQUIRED
+  with zero reviews, yet merged — a bypass demonstrably occurred.
+- Status: bypass **verified**; the exact mechanism (gh `--admin`
+  flag, ruleset bypass list contents) **user-must-provide** (ruleset
+  config needs owner/admin access to read).
+- Resolves it: owner states the flag/ruleset arrangement.
+- Safe default: a session never merges without the owner's ask for
+  that campaign; if a merge is blocked, surface the blocker — do not
+  probe for bypass paths.
+
+## 4. `skill-vetting` absent from both live installs
+- Claim: deliberate pending-adoption state (the advisory hook is
+  opt-in; "Workstream B"), not an overlooked sync failure.
+- Repo evidence: `skills/skill-vetting/` exists; both live installs
+  lack it (diffs run 2026-07-30); README documents the hook as
+  optional. Intent is session/memory-recorded only.
+- Status: state **verified**; intent **user-must-provide**.
+- Resolves it: owner either installs it or records the deferral.
+- Safe default: do not install or uninstall it yourself; expect the
+  README sync loop to error on exactly this directory.
+
+## 5. The 2026-07-12 local library's stale spots
+- Claim: the older root-level `skills-staging/` library (local-only,
+  gitignored) is stale where it overlaps current practice.
+- Verified examples (2026-07-30): its release-and-publish requires a
+  version bump per substantive PR (contradicted by #84–#97 at
+  0.1.16); its re-verify command greps `README.zh-TW.md` (renamed
+  `README.zh-Hant.md` in `c15c29d`-era work — the old path no longer
+  exists); its gate-ops file records the 2026-07 lineup gpt-5.5 +
+  grok-4.5 max (differs from the campaign lineup; both are
+  date-stamped volatile by their own rules).
+- Status: **historically-valid**; the files are the owner's local
+  scratch, not touched by this delivery.
+- Resolves it: an owner-sanctioned refresh or retirement of the old
+  library.
+- Safe default: for overlapping topics, canon skills win, then this
+  dated library; verify any old-library command against the repo
+  before running it.
+
+## 6. The sol-unverified r11 folds in #90
+- Claim: the final two r11 folds (`git add --` hardening; the
+  fast-forward scope sentence) merged without sol re-review.
+- Repo evidence: the two folds are named in `83a038d`; #90's body
+  discloses that the final two folds land sol-unverified without
+  naming them; grok/luna never objected; CI green.
+- Status: **verified** as a disclosed, accepted residual.
+- Resolves it: nothing owed; a future finding against those folds is
+  a new finding.
+- Safe default: do not re-open #90's gate; treat the disclosure as
+  the record.
+
+## 7. Evaluation comments — verification depth
+- Claim: all nine constituent PRs (#85–#89, #91, #94–#96) carry a
+  maintainer evaluation comment.
+- Repo evidence: #85, #94, #96 read in full on 2026-07-30 (shape and
+  no-owner-quotes confirmed); the remaining six confirmed present the
+  same day with matching "Evaluation record (maintainer)" openers
+  (on #86 it is the SECOND comment — the first is the contributor's).
+- Status: **verified** (presence, all nine; full-text read, three).
+- Resolves it further: read the remaining six in full before citing
+  their details.
+- Safe default: when writing new evaluation comments, follow the
+  three fully-read ones' shape.
+
+## 8. Session-memory phrase "evals/round4 … COMMITTED"
+- Claim (memory, 2026-07-23): round-4 eval design "committed".
+- Repo evidence: `evals/` is gitignored with zero tracked files
+  (2026-07-30) — nothing eval-related is committed to this repo.
+- Status: memory claim **unverified as written**; repo state
+  verified. Likely loose wording for "written to the private evals
+  tree".
+- Resolves it: owner's private tree; not resolvable here.
+- Safe default: treat `evals/` as private and ignored; never cite a
+  tracked evals path.
+
+## 9. This library's own delivery mechanism
+- Claim/process note: delivering skills under a gitignored path
+  required `git add -f`, per the explicit retiring-architect task
+  instruction; `.gitignore` was not modified.
+- Status: verified (this delivery).
+- Safe default: this is NOT precedent for publishing other ignored
+  zones (`evals/`, `internal/`, `.claude/`, `guideline*.txt`,
+  `pack-eval-artifacts/`) — those stay private absent the same kind
+  of explicit owner instruction.
diff --git a/skills-staging/2026-07-30-campaign-ops/contribution-gate-playbook/SKILL.md b/skills-staging/2026-07-30-campaign-ops/contribution-gate-playbook/SKILL.md
new file mode 100644
index 0000000..5cd78db
--- /dev/null
+++ b/skills-staging/2026-07-30-campaign-ops/contribution-gate-playbook/SKILL.md
@@ -0,0 +1,158 @@
+---
+name: contribution-gate-playbook
+description: Load when contributor PRs are open on F-e-u-e-r/opus-pack and the owner asks to review, integrate, or merge them (any "follow our previous practice" form), or when a combined-*/integrate-* branch exists with its constituent PRs still open. Not for authoring doctrine (doctrine-change-conventions) or judging individual lens verdicts (gate-adjudication-and-folds).
+---
+
+# Contribution Gate Playbook
+
+The observed, repo-verified pipeline that took #85–#89 → PR #90
+(`a148180`), #91 → PR #93 (`92077a7`), and #94–#96 → PR #97 (`79ca49c`)
+on 2026-07-29/30. Canonical gate doctrine is
+`skills/cross-model-review/SKILL.md` — this file only pins how THIS
+repo runs it. On any disagreement, the canon skill wins.
+
+## Step 0 — pin the campaign set
+
+Trigger: the owner's ask arrives.
+
+- The set = the PRs open at instruction time (`gh pr list --state
+  open`). Record the numbers before touching anything.
+- A PR that opens mid-campaign is NOT swept in. Precedent: #91 opened
+  2026-07-29T22:18Z, before #90 merged (23:34Z); it was excluded from
+  #90, flagged to the owner, and handled as its own campaign (#93).
+- The owner's ask authorizes THIS set; a later set needs a fresh ask
+  (operational-rigor §2, per-invocation grant).
+
+## Step 1 — verify each PR's verifiable claims
+
+- Every repo-facing claim in a PR's text (a commit id, a section
+  number, a "main already says X") is checked against the repo BEFORE
+  the gate. Precedent: #85's `8f8413f` incident claim was verified
+  in-repo, and the check is recorded in #85's evaluation comment.
+- A claim that fails verification is a finding to surface and
+  adjudicate, not something to silently fix or silently accept.
+
+## Step 2 — build the combined branch
+
+- One integration branch from current main, named by joining the PR
+  numbers — `combined-85-86-87-88-89`, `combined-94-95-96` (multi-PR)
+  or `integrate-91` (single PR) are the observed forms.
+- Bring each PR's head branch local first (`gh pr checkout <number>`
+  handles fork PRs too; the campaign merge commits show `pr-<number>`
+  local branch names). Merging every head into the combined branch is
+  what closes each constituent by reachability once the combined PR
+  merges (observed on #85–#89 and #94–#96; GitHub marks them MERGED).
+  Merge ASCENDING (lowest number first) — the house convention that
+  keeps Provenance-entry stacking orderly.
+- Provenance-tail conflicts are the expected conflict shape: stack
+  both entries in PR order and reconcile with zero content loss (read
+  both sides in full; nothing dropped silently).
+
+## Step 3 — integration fixes, before round 1
+
+Integrator additions are separate commits, named as integration
+fixes, and are gate-reviewed like all other content. The recurring
+three (commits `f529014`, `814d116`):
+
+- in-body `unprobed` markers on clauses carrying only a
+  Provenance-level label (the covenant's standing form);
+- rewrap added lines over 80 characters without changing words;
+- Provenance blank-line normalization (house continuous-prose style).
+
+## Step 4 — checks green at every commit
+
+- `python3 .github/checks.py` green before every push. The campaign
+  PR bodies record checks green at every (fold) commit as the norm.
+
+## Step 5 — run the gate rounds
+
+- Lens lineup as run 2026-07-29/30 — a dated OPERATING RECORD, not a
+  lineup to trust (the canonical rule is cross-model-review §1:
+  discover reviewers at session time; the concrete run recipes are
+  owner-personal and are NOT in this repo — ask the owner): grok-4.5
+  high + gpt-5.6-luna ultra EVERY round; gpt-5.6-sol max joining
+  every 3rd round, pre-commit rounds, and terminal-candidate rounds.
+  Each campaign's records state its own cadence (PR #90/#93 bodies;
+  for #97, its constituents' evaluation comments) — the newest
+  campaign's record is the freshest statement; the standing wording
+  is owner-set (UNCERTAINTY item 2).
+- Lenses run as background CLI jobs writing verdict files. An empty
+  or still-writing file is NOT a verdict; wait for the completion
+  signal, then read the file.
+- Each round's packet carries the current diff and the cumulative
+  disposition ledger of prior rounds' folds and rejections.
+- Judge what comes back per `gate-adjudication-and-folds`.
+
+## Step 6 — reach a terminal state
+
+- Normal close: TRIPLE PROCEED — every lens PROCEED in the SAME round
+  (#93 r2; #97 r3).
+- Non-convergence close: the bounded-loop close (cross-model-review
+  §4 is the canonical loop bound; `gate-adjudication-and-folds` has
+  this repo's precedents). Residuals are disclosed in the PR body.
+
+## Step 7 — combined PR, CI, merge
+
+- Push the branch; open ONE combined PR. Body shape: "What lands" per
+  constituent / "Review gate" (rounds, lenses, fold trajectory,
+  rejections) / "Integration notes" — the #90 shape; #93/#97 carry
+  the same three blocks with lighter headings.
+- Wait for the three CI checks (consistency, hook-suites,
+  gate-template) by reading state directly — `gh pr checks 97`,
+  substituting the campaign PR number; do not hand-roll wait loops
+  (`failure-archaeology` entry 1).
+- Merge as a merge commit, and ONLY on an owner ask whose words cover
+  MERGING this campaign — write the AUTH line first (operational-rigor
+  §2); this playbook is never that authorization. An owner ask to
+  merge does not skip the CI wait: report pending checks and merge on
+  green, or on the owner's explicit go given the pending state.
+  Observed record, not a norm to reuse: the three campaign merges
+  landed with reviewDecision REVIEW_REQUIRED and zero reviews — an
+  owner-side bypass whose exact mechanism is not readable from here
+  (UNCERTAINTY item 3). If a merge is blocked, surface the blocker;
+  never probe for bypass paths.
+
+## Step 8 — post-merge duties (all of them, in order)
+
+1. Evaluation comment on EVERY constituent PR — maintainer-record
+   shape: what was verified for that PR / gate folds applied to its
+   text / final decision + merge SHA. Never quote the owner's
+   messages (`repo-boundaries-and-sync`).
+2. Sync the live skill copies and cmp-verify at the merge SHA
+   (`repo-boundaries-and-sync` has the current expected state).
+3. Probe debt: new `unprobed` markers are ALREADY the live queue —
+   issue #92's canonical view is a grep, so there is no hand-kept
+   list to update.
+4. Update session memory with the campaign record.
+
+## Done definition
+
+Every constituent PR shows MERGED; the combined merge commit is on
+main with CI green; an evaluation comment sits on every constituent;
+live installs cmp-clean per the current expected state; no
+private-zone path appears in any pushed commit (`git show --stat` on
+each).
+
+## When NOT to use this skill
+
+- The owner asked a question about a PR (assess, don't integrate).
+- The change is yours, not a contributor's — the gate doctrine still
+  applies via cross-model-review, but the combined-branch and per-PR
+  comment mechanics here assume third-party PRs.
+
+## Provenance
+
+Distilled 2026-07-30 from the 2026-07-29/30 session transcript
+(`chat-history.md`, local) and verified same-day against: PR
+#90/#93/#97 bodies and merge commits `a148180`/`92077a7`/`79ca49c`;
+evaluation comments on all nine constituent PRs (three read in full,
+six confirmed present with matching openers — UNCERTAINTY item 7);
+issue #92; commits `f529014`/`814d116`/`6fa6154`/`5fa241b`/`83a038d`/
+`1560b97`; `.github/checks.py`; `.github/workflows/checks.yml`;
+`gh pr view <n> --json reviewDecision,reviews`; `gh pr view 91 --json
+createdAt`.
+
+Re-verify: `gh pr view 97 --json body --jq .body | head -45` — and if
+a newer combined/integrate PR exists, read the newest body instead;
+practice evolves campaign by campaign and the PR bodies are its
+freshest public record.
diff --git a/skills-staging/2026-07-30-campaign-ops/doctrine-change-conventions/SKILL.md b/skills-staging/2026-07-30-campaign-ops/doctrine-change-conventions/SKILL.md
new file mode 100644
index 0000000..913b36b
--- /dev/null
+++ b/skills-staging/2026-07-30-campaign-ops/doctrine-change-conventions/SKILL.md
@@ -0,0 +1,104 @@
+---
+name: doctrine-change-conventions
+description: Load when adding or editing rule text under skills/*/SKILL.md or design-pack/skills/*/SKILL.md that is heading for a commit or merge — including applying an already-adjudicated gate fold, writing a Provenance entry, adding a rule that has no probe yet, or resolving a Provenance merge conflict.
+---
+
+# Doctrine Change Conventions
+
+What any doctrine diff in this repo must satisfy beyond the canon
+skills (`skills/skill-authoring` is canonical for method; the README
+"House covenant" paragraph for the probe policy). Every item below was
+verified against the repo on 2026-07-30.
+
+## The covenant pairing (in-body marker + Provenance entry)
+
+- README House covenant, quoted verbatim (2026-07-16): "a new
+  behavioral rule ships with the probe or trap that would have failed
+  without it, or it ships explicitly labeled `unprobed`."
+- The standing form here is BOTH pieces: an in-body marker at the
+  clause — ``(`unprobed` — see Provenance.)`` — AND a Provenance
+  entry. A clause carrying only the Provenance-level label gets the
+  in-body marker added as an integration fix (`f529014`: "#94 and #96
+  already carried theirs"; `814d116` added six for the #85–#89
+  batch).
+- Provenance source labels in current use: `contributor-reported, not
+  linkable` (private evidence) and `repo-verifiable` (cite the
+  commit/PR). External measurements are additionally glossed "cited
+  as shape" — a qualifier phrase, not a label — when another
+  project's numbers are described without restating them as this
+  repo's own evidence.
+- Probe debt needs no ledger edit: issue #92's canonical live view is
+  `grep -rn 'unprobed' skills/*/SKILL.md`. The issue stays open while
+  any marker exists; a marker leaves only by a probe-result
+  write-back or a recorded demotion/decline (see #92's body). Note
+  the grep's scope is `skills/` only — a design-pack marker would
+  escape it; extend the sweep when editing design-pack doctrine (the
+  gap is #92's mechanism, recorded here).
+
+## Line and glyph norms
+
+- Added lines stay ≤80 characters, counted in CHARACTERS. This is a
+  review-enforced house norm recorded in the campaign PR bodies (#90:
+  "added lines all ≤80 characters"; #93/#97: "added lines ≤80 chars")
+  — it is NOT enforced by checks.py (verified 2026-07-30: checks.py
+  contains no width check). Rewrap without changing words.
+- Glyphs: checks.py bans invisible/bidi/Tag-Block characters across
+  ALL tracked files, failing closed on anything it cannot decode.
+  Ordinary visible Unicode is house-normal — a gate finding claiming
+  `→` was banned was rejected on an occurrence count plus the
+  checks.py class read (worked example and counts:
+  `gate-adjudication-and-folds`, Rejections section).
+
+## Provenance house style
+
+- Continuous prose: no blank lines inside a Provenance section
+  (normalized during the #86 merge resolution; recorded in
+  `814d116`).
+- Entries append chronologically; a merge conflict at the section
+  tail is the expected shape — stack entries in PR order.
+
+## Version policy (date-stamped 2026-07-30)
+
+- Doctrine-only merges do NOT bump the version: `0.1.16` was set
+  inside PR #83 (commit `32c6929`) and is unchanged through #84–#97.
+  The 2026-07-12-era "bump every substantive PR" convention is
+  historical, superseded by observed practice.
+- Git tags still end at `alpha-0.1.2`; versioning lives in the README
+  badge AND "Early alpha" callout lines (both languages),
+  `plugin.json`, and `marketplace.json` — six sites per `32c6929`'s
+  own message; checks.py check 2 enforces their agreement, so a bump
+  touches all sites together or CI fails.
+
+## Pre-push check
+
+- Run `python3 .github/checks.py` locally before every push. Know its
+  scope: frontmatter checks cover the marketplace-declared skill
+  roots (`skills/`, `design-pack/skills/`) via the working tree; the
+  hidden-directive sweep covers every git-tracked file; there is no
+  line-width or prose-style check — those are review duties.
+
+## Done definition
+
+checks.py green; every new behavioral clause carries the in-body
+marker paired with a Provenance entry that names its source label;
+added lines ≤80 characters; no version bump unless the owner asked
+for one.
+
+## When NOT to use this skill
+
+Judging reviewer findings about your text →
+`gate-adjudication-and-folds`. Campaign mechanics →
+`contribution-gate-playbook`. General authoring method (triggers,
+dup-check, compaction) → `skills/skill-authoring` (canonical).
+
+## Provenance
+
+Distilled 2026-07-30 from the session transcript; verified same-day
+against: README House covenant paragraph; issue #92 body; commits
+`f529014`/`814d116`/`32c6929`; `.github/checks.py` (read in full for
+the width/glyph scope claims); `git tag -l`; the version badge lines
+in both READMEs; PR #90/#93/#97 bodies.
+
+Re-verify: `python3 .github/checks.py && gh issue view 92 --json
+state --jq .state` (expect "all checks passed" and OPEN while any
+`unprobed` marker exists).
diff --git a/skills-staging/2026-07-30-campaign-ops/failure-archaeology/SKILL.md b/skills-staging/2026-07-30-campaign-ops/failure-archaeology/SKILL.md
new file mode 100644
index 0000000..6afeef9
--- /dev/null
+++ b/skills-staging/2026-07-30-campaign-ops/failure-archaeology/SKILL.md
@@ -0,0 +1,114 @@
+---
+name: failure-archaeology
+description: Load when about to wait on CI in a hand-rolled loop, ship a doctrine line that states runtime behavior, adopt or delete a leftover local branch, add a contributor clause without a probe marker, or re-open something this session settled (version bumps, mid-campaign PRs, the sol r11 residue, the untracked transcripts).
+---
+
+# Failure Archaeology — 2026-07-29/30 session
+
+Dead ends, traps, and residue from the contributor-PR campaign
+session, so no future agent re-walks them. Entry fields per the
+project-skill template: disposition / what happened / mechanism /
+standing rule / residue location.
+
+## 1. dead — hand-rolled CI wait loop
+
+- What: an until-loop ("wait until no check is pending") was launched
+  to wait for PR #90's CI, then stopped; the session's own live note
+  hedged that an inverted loop condition would return early. A
+  transient API error landed in the same window. (Transcript-recorded
+  only — the stopped loop left no repo artifact.)
+- Mechanism: a hand-written wait predicate is unverifiable from the
+  outside — when it exits you cannot distinguish "done" from
+  "predicate wrong", and a mid-wait interruption leaves no state.
+- Standing rule: read CI state directly and re-derive on every
+  resume — `gh pr checks <n>` — instead of encoding the wait
+  condition in a loop you cannot test. The session's later CI waits
+  (#93, #97) completed as plain background status checks.
+- Residue: none (the loop was stopped; nothing landed).
+
+## 2. dead — untested runtime-semantics text in #96 (fixed in gate)
+
+- What: contributor text asserted `os.environ["KEY"]` "silently reads
+  empty" in non-interactive shells, and its ❌ example used `'\\n'`
+  inside single quotes.
+- Mechanism: prose about runtime behavior that nobody had executed.
+  Both lenses converged on the KeyError correction (r1); sol+luna
+  refined `.get()` → `None` (r2); a lens executed the printf example
+  and showed `'\\n'` emits a literal backslash-n, so the example
+  could never produce the incident it described.
+- Standing rule: runtime claims in doctrine get executed verification
+  before shipping (`gate-adjudication-and-folds`).
+- Record: `6fa6154`, `5fa241b`, PR #97 body, #96's evaluation
+  comment.
+
+## 3. recurring-trap — covenant markers missing on contributor clauses
+
+- What: #95's clause carried only the Provenance-level label
+  (`f529014`); the #85–#89 batch needed six in-body markers added
+  (`814d116`); a sweep found five Provenance entries missing the
+  contributor-reported label (#90 body).
+- Mechanism: contributors satisfy the covenant at the Provenance
+  level and miss the in-body standing form; every batch re-imports
+  the same gap.
+- Standing rule: sweep new clauses for marker + label pairing as an
+  integration fix BEFORE gate round 1 (`doctrine-change-conventions`).
+- Tripwire: a new clause whose opening sentence lacks the marker
+  while its Provenance entry says unprobed.
+
+## 4. recurring-trap — your own folds introduce defects
+
+- What: #90's gate caught the staged-diff gate's circular trigger
+  (the catch is named in #90's body; that an earlier fold introduced
+  it is session-recorded).
+- Standing rule: folds are attack surface for the next round; no
+  terminal close on unreviewed folds outside the disclosed
+  bounded-loop residual (`gate-adjudication-and-folds`).
+
+## 5. residue, not in-progress work — local branch graveyard
+
+- What (observed 2026-07-30): local branches with deleted upstreams
+  (`[origin/…: gone]`: integrate-pr62-65, integrate-pr68-69,
+  phase-a…d, hook-*, delegation-*) plus never-pushed mine-* and
+  pr3x-review branches from earlier mining campaigns.
+- Standing rule: residue — do not adopt-and-finish, do not delete
+  without the owner's ask (cleanup mutates the owner's workspace;
+  operational-rigor §2). Observed mechanics: this repo merges with
+  TRUE merge commits, so `git branch --merged main` DOES list all of
+  these branches as merged (verified 2026-07-30: 37 of 37) —
+  reachability is the authoritative signal here. The canonical
+  squash-merge caveat ("squash defeats `--merged`") applies to
+  squash-based repos, not to this one's observed history. The
+  never-pushed mine-*/pr3x-review branches are also listed as merged
+  — their content landed via other branches; still residue, not live
+  work.
+- Residue location: `git branch -vv` on the owner's machine.
+
+## Deliberately-not-done (do not "helpfully" complete)
+
+- No version bump for #84–#97 — doctrine-only merges do not bump
+  (`doctrine-change-conventions`). Do not "fix" the badge forward.
+- #91 was NOT merged into #90 although it was open before #90 merged
+  — mid-campaign arrivals stay out of the pinned set
+  (`contribution-gate-playbook` step 0). It got its own campaign
+  (#93) instead.
+- No r12 for #90 — sol's r11 residue closed under the bounded-loop
+  precedent and ships labeled `sol-unverified` in the PR body. Do not
+  re-open the campaign to "finish" sol's convergence; a future
+  finding against those two folds is a NEW finding, handled fresh.
+- `security-enhancement.md` and `chat-history.md` stay untouched at
+  the repo root — owner disposition pending (issue #100;
+  `UNCERTAINTY.md` item 1). Do not move, stage, or edit them.
+
+## Re-verify
+
+`git branch -vv | grep -c ': gone'` (nonzero = residue still
+present) and `gh pr view 90 --json body --jq .body | grep -n
+'sol-unverified\|bounded'` (the disclosed residual is still the
+record). Entry 5's branch list is machine-local — re-observe it
+rather than trusting this file's snapshot.
+
+## Provenance
+
+Distilled 2026-07-30 from the session transcript; commits, PR bodies,
+and comments cited inline were re-read in-repo the same day; the
+branch list was observed directly (`git branch -vv`, 2026-07-30).
diff --git a/skills-staging/2026-07-30-campaign-ops/gate-adjudication-and-folds/SKILL.md b/skills-staging/2026-07-30-campaign-ops/gate-adjudication-and-folds/SKILL.md
new file mode 100644
index 0000000..526b2a4
--- /dev/null
+++ b/skills-staging/2026-07-30-campaign-ops/gate-adjudication-and-folds/SKILL.md
@@ -0,0 +1,121 @@
+---
+name: gate-adjudication-and-folds
+description: Load when a cross-model lens verdict has landed for a doctrine change in this repo and you must decide fold, reject, or stop — including when two lenses propose opposite fixes for one finding, when one lens keeps producing must-fixes after the others hold PROCEED, or when a finding asserts runtime or tool semantics.
+---
+
+# Gate Adjudication and Folds
+
+Canonical verdict rules: `skills/cross-model-review/SKILL.md`
+(findings are claims, §3; the loop is bounded, §4). This file pins the
+adjudication patterns the 2026-07-29/30 campaigns used, each with its
+public record. On disagreement, the canon skill wins.
+
+## Fold discipline
+
+- Reproduce a finding against the actual file/repo before folding it.
+- Fold commits are named `Fold gate rN: <what>` — single-PR campaigns
+  prefix the target, e.g. `Fold #91 gate r1: …` — and state which
+  lens produced each item and why the fix is right (see `6fa6154`,
+  `5fa241b`, `83a038d`, `1560b97`).
+- Your own folds are attack surface for the NEXT round: #90's gate
+  caught the staged-diff gate's circular trigger, a defect an earlier
+  fold itself introduced (the catch is listed in #90's body; the
+  fold-introduced attribution is session-recorded). Do not close on a
+  round whose folds no lens has seen, except under the bounded-loop
+  close below, where exactly that residual is disclosed by name.
+
+## Rejections carry in-repo counter-evidence
+
+- A reviewer finding you reject gets a recorded counter-fact in the
+  fold commit and PR body. #90's body records the aggregate — "7
+  findings rejected with in-repo counter-evidence"; the two worked
+  examples below are session-recorded, their counter-facts re-derived
+  in-repo 2026-07-30:
+  - a lens called `→` a banned glyph — rejected by reading checks.py's
+    banned class (invisible/bidi/Tag-Block only; U+2192 is not in it)
+    and counting occurrences on main (plentiful: 107 hits in
+    `skills/*/SKILL.md` at the pre-#90 main);
+  - a lens called a line over-wide — rejected with the character
+    count (78; the house norm counts characters, not bytes).
+- No counter-evidence found = it is not a rejection; it is still an
+  open finding.
+
+## Runtime-semantics findings get EXECUTED verification
+
+- When doctrine text asserts what a runtime does, execute it before
+  the claim ships. Precedent chain on #96: the draft said the bracket
+  form "silently reads empty" — execution shows `os.environ["X"]`
+  raises KeyError; `.get()` returns `None`; an unguarded shell `$X`
+  expands empty (r1 converged must-fix + r2 refinement, `6fa6154`,
+  `5fa241b`). A lens also executed the ❌ example and showed `'\\n'`
+  inside single quotes emits a literal backslash-n — the example
+  could never produce the incident it described.
+- Rule: a doctrine line stating tool/runtime behavior is a claim to
+  verify by execution, not prose to wordsmith. This applies to YOUR
+  folds too — the r2 refinement corrected an r1 fold's attribution.
+
+## Opposite-fix convergence → adjudicate by the protected property
+
+- Two lenses converged on the SAME gap in #91 (the durable/ephemeral
+  overlap for "gitignored but externally archived" paths) with
+  OPPOSITE precedence fixes. Adjudication anchored on what the rule
+  protects — a later reader's ability to resolve the citation — so
+  check-ignore wins: an ignored working path is ephemeral even when
+  an external archive preserves it; cite the archive itself. Recorded
+  in PR #93's body, the fold commit `1560b97`, and #91's evaluation
+  comment — the comment publicly states that the two lenses proposed
+  opposite precedences.
+- Rule: when converged fixes conflict, derive the answer from the
+  rule's protected property, not from either lens's framing — and
+  record the adjudication where the next maintainer will read it.
+
+## Terminal conditions
+
+- TRIPLE PROCEED: every lens PROCEED in the same round (#93 r2,
+  #97 r3). The normal close.
+- Bounded-loop close (non-convergence): canon caps the loop and sends
+  a non-converging gate to a human with the trail (cross-model-review
+  §4 — on disagreement it wins). The house application, owner-set in
+  the 2026-07-23 twelve-round campaign (PR #61 era) and re-applied in
+  #90: a campaign may run past canon's 2–3-round cap only under the
+  owner's standing campaign mandate; when one lens keeps producing
+  must-fixes while the gate otherwise converges (#90's public record:
+  grok PROCEED r5–r11, luna PROCEED r7/r9/r11, sol still finding
+  last-mile items), declare the close condition before the terminal
+  round (session-recorded for #90 at r8; the public record is the
+  close itself), execute it, apply the final folds labeled
+  `<lens>-unverified`, and put the close AND its named residuals in
+  front of the owner — in the PR body's Review gate section and the
+  campaign report (`83a038d`; #90 body). A close that leaves
+  residuals is surfaced, never silent.
+- The two failure shapes this bounds: unbounded re-review (no
+  terminal proof ever arrives) and silent early close (residue
+  undisclosed). The bound plus named-residual disclosure avoids both.
+
+## Done definition
+
+Every lens finding for the round is folded (commit cites it),
+rejected (counter-evidence recorded), or explicitly carried into the
+next round's packet; a terminal state was reached by one of the two
+closes; the PR body's "Review gate" section states the trajectory,
+the rejections, and any unverified residuals by name.
+
+## When NOT to use this skill
+
+Pipeline mechanics (branching, CI, merge, comments) →
+`contribution-gate-playbook`. What the doctrine text itself must
+satisfy → `doctrine-change-conventions`.
+
+## Provenance
+
+Distilled 2026-07-30 from the session transcript; verified same-day
+against PR #90/#93/#97 bodies, the #91/#96/#85 evaluation comments,
+and commits `6fa6154`/`5fa241b`/`83a038d`/`1560b97`. Session-recorded
+specifics (public anchors in parentheses): the `→`/78-char rejection
+details (#90's body carries the aggregate rejections line), the r8
+close pre-declaration (the body carries the close itself), and the
+fold-introduced attribution of the circular trigger (the body lists
+the catch). The `→` occurrence count was re-run in-repo 2026-07-30.
+
+Re-verify: `git log --oneline --grep='Fold' -10` and the newest
+combined PR body's "Review gate" section.
diff --git a/skills-staging/2026-07-30-campaign-ops/repo-boundaries-and-sync/SKILL.md b/skills-staging/2026-07-30-campaign-ops/repo-boundaries-and-sync/SKILL.md
new file mode 100644
index 0000000..97c45e2
--- /dev/null
+++ b/skills-staging/2026-07-30-campaign-ops/repo-boundaries-and-sync/SKILL.md
@@ -0,0 +1,116 @@
+---
+name: repo-boundaries-and-sync
+description: Load before staging or committing anything in this working tree, after any merge that touches skills/ (live-copy sync duty), when a sync or diff check reports an unexpected missing directory, or when an untracked file sits at the repo root.
+---
+
+# Repo Boundaries and Sync
+
+## Boundary map (verified 2026-07-30)
+
+Published (tracked):
+
+- `skills/`, `design-pack/` — pack source (the marketplace roots).
+- `hooks/`, `.github/`, `.claude-plugin/` — enforcement, CI,
+  manifests.
+- `reviews/` — public review and threat-model records.
+- `README.md`, `README.zh-Hant.md`, `LICENSE`,
+  `THIRD-PARTY-NOTICES.md` — docs; the two READMEs mirror each other.
+
+Gitignored (private — does not publish by default):
+
+- `evals/`, `internal/`, `pack-eval-artifacts/`, `guideline*.txt` —
+  fixtures burn on exposure; owner-private notes and drafts. These
+  NEVER publish.
+- `.claude/` — live install + settings.
+- `skills-staging/` — retiring-architect scratch (`92314a6`);
+  publishes only on explicit owner instruction.
+
+## The working tree is permanently dirty-adjacent
+
+- The repo root can carry UNTRACKED private session transcripts. On
+  2026-07-30 it holds `chat-history.md` and `security-enhancement.md`
+  (both full session transcripts — private content). Neither has ever
+  entered a commit.
+- Standing defense: stage by explicit pathspec, always
+  (`git add -- <path>`); never any stage-everything form. The
+  gitignore commit for skills-staging states the exact threat in its
+  own message: it "keeps a stray `git add -A` from publishing"
+  private content (`92314a6`).
+- The transcripts' disposition (ignore pattern vs relocation) is the
+  owner's pending decision — tracked in issue #100 (`UNCERTAINTY.md`
+  item 1). Do not stage, quote, relocate, or edit them; quoting them
+  in any public artifact also breaks the rule below.
+
+## No owner quotes in public artifacts
+
+- Standing owner directive (recorded 2026-07-24, applied throughout
+  2026-07-29/30): the owner's message wordings never appear in PR
+  bodies, commit messages, PR/issue comments, or any published file.
+  Evaluation comments describe facts, findings, and decisions only.
+  AUTH quotes live in session reports to the owner, nowhere public.
+- Verified against practice: the sampled evaluation comments on
+  #85/#94/#96 contain no owner quotes.
+
+## Live-copy model and sync duty (three locations)
+
+- `skills/` — the publish source (tracked).
+- `.claude/skills/` — repo-local live install (gitignored).
+- `~/.claude/skills/` — global install (outside the repo).
+- The README Maintainer Notes' sync contract is canonical (on any
+  divergence from this file, the README wins): after editing any
+  SKILL.md, sync (`cp -R skills/. .claude/skills/`) and run the
+  per-skill diff loop before pushing.
+- TEMPORARY overlay on that contract (delete this bullet when
+  UNCERTAINTY item 4 is settled): today the blanket `cp -R` would
+  ALSO install `skill-vetting`, closing the known gap without the
+  owner's decision. Until the owner settles it, copy the changed
+  skill directories explicitly (`cp -R skills/<name>
+  .claude/skills/`), skipping `skill-vetting`; do the same for the
+  global copy (`~/.claude/skills/`), then cmp-verify both at the
+  merge SHA (campaign practice, recorded in the close reports).
+- Current expected state (2026-07-30) — this file is the SINGLE HOME
+  for it; START-HERE/MANIFEST/UNCERTAINTY point here. Update this
+  file first when the state changes: `skill-vetting` exists in
+  `skills/` but in NEITHER live install, so the README loop errors on
+  exactly that one directory. Known pre-existing state (the advisory
+  hook ships opt-in; install pending an owner decision) — not drift
+  you introduced. Every other skill dir cmp-clean at `79ca49c`.
+- Removing or renaming a published skill: delete its old dir from the
+  live installs by hand in the same change — `cp -R` never deletes
+  (README Maintainer Notes).
+
+## Done definition
+
+Nothing from a gitignored/private zone appears in the staged set
+(`git status` read, not assumed); every commit was staged by
+pathspec; after a merge, both live installs match `skills/` except
+the known `skill-vetting` gap — or that gap was closed deliberately
+by the owner, in which case update this file's expected state.
+
+## When NOT to use this skill
+
+Deciding whether content may ship at all (covenant, markers, width)
+→ `doctrine-change-conventions`. Campaign mechanics →
+`contribution-gate-playbook`.
+
+## Provenance
+
+Distilled 2026-07-30 from the session transcript; verified same-day
+against: `.gitignore` + `git check-ignore`; `git ls-files` top-level
+enumeration; `92314a6`'s commit message; the two transcripts'
+presence and untracked status (`git status --short`); README
+Maintainer Notes; live-copy diffs run 2026-07-30 (both installs
+differ from `skills/` only in `skill-vetting`); evaluation comments
+(three read in full, all nine confirmed present).
+
+Re-verify:
+
+    git status --short
+    for d in skills/*/; do
+      diff -rq "$d" ".claude/skills/$(basename "$d")"
+    done
+
+(expected today: the two `??` transcripts, one error line for
+`skill-vetting`, and — until the delivering PR merges — this
+library's own tracked entries; anything else is new state —
+investigate before acting).