feat(claude-config): ship I21 and I22 — the agnostic halves of the ratified Q8 split

plugins/claude-config/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://json.schemastore.org/claude-code-plugin-manifest.json",
   "name": "claude-config",
-  "version": "0.20.1",
+  "version": "0.21.0",
   "description": "Five audit skills for a repo's Claude Code configuration: audit (settings.json / .mcp.json / hooks / plugins / permissions drift), audit-automation-gaps (evidence-gated verdicts on automation gaps), audit-permission-grants (allow-rule / allowed-tools grants for auto-mode durability and portability), audit-instructions (locally-owned instruction surfaces vs current model capability — proposes removals/rewrites of instructions the model no longer needs, and detects cross-surface instruction conflicts), and audit-pass (one coordinated, ordered, resumable pass over a named target — three-scope inventory, run-time-derived exclusion set, stable finding identity, suppression memory, resume, one human gate — delegating every check to the plugin that owns it).",
   "author": {
     "name": "Melodic Software",

plugins/claude-config/CHANGELOG.md

@@ -3,6 +3,97 @@
 All notable changes to the `claude-config` plugin are documented here. Format follows
 [Keep a Changelog](https://keepachangelog.com/en/1.1.0/); this plugin uses semantic versioning.
 
+## [0.21.0]
+
+### Added
+
+- **`audit-instructions`: the two agnostic-mechanism rows from the IA-6 / IA-10-A2 ownership split —
+  I21 and I22** (criteria 1.7.0 → 1.8.0). Both source rules were **compounds**: an agnostic
+  mechanism fused to a consumer-state instance naming this fleet's own machines, files, and dates.
+  Routing either wholesale was wrong in both directions — outward it would ship our private state to
+  every consumer, inward it would strand a reusable staleness control. Each was split at the
+  mechanism/instance line; only the mechanism halves are here. The instance halves (a dated vet, a
+  chezmoi-managed fleet pin) are drafted for the consumer repository and deliberately ship nowhere
+  in this plugin.
+
+  - **I21 — effort level pinned across a model change with no re-sweep** (`mechanical`,
+    `ANTHROPIC-DOCS`, `warning`, unscoped). The effort scale is calibrated per model, so the same
+    level name does not carry the same underlying value across models, and a level measured against
+    one model then carried to the next is a pin nobody re-measured. The promotion gate is met on the
+    strong form: model configuration states the calibration property **with no model qualifier**, so
+    that page alone carries the row; the effort page's Opus 5 subsection is cited only for the
+    remediation's wording, and its per-model placement does not narrow a property stated generally.
+
+    **The model range is a Detect condition, not a `Model scope` annotation**, on I17's reasoning.
+    It is in Detect because the *consequence* varies: Claude Code applies a model's default effort
+    on first run of Fable 5, Opus 4.8, or Opus 4.7 and holds it, so a carried level there is
+    overridden harmlessly — while **Opus 5 has no such hold** and a stale pin actually reaches the
+    request. One thing is recorded as **unresolved rather than inferred**: the page names `/effort`
+    and `--effort` as *examples* ("such as") of the explicit choice that releases the hold, so
+    whether a settings-file `effortLevel` pin releases it is not stated anywhere read for this row.
+    The row therefore fires on the missing re-derivation regardless of model, and the hold is
+    severity context, never a fence.
+
+    Four fences keep it honest. A prescription of **`high` is exempt only where `high` is the
+    resolved target's default** — it is "Equivalent to not setting the parameter", so on such a model
+    it carries no measured calibration. **The exemption keys to the resolved target, never to the
+    wording**, which is what makes it correct: `high` is the default everywhere **except Opus 4.7,
+    which defaults to `xhigh`**, so when the target is 4.7 the exemption lifts and a broad
+    model-agnostic "always use `high`" naming no model is a finding — indeed the sharper case, since
+    a pin written where `high` was the no-op default becomes a silent step-down the moment it reaches
+    a model whose default sits above it. A resolved target always exists, because the skill body
+    aborts rather than run against an unresolved one, so the fence never guesses. A **per-task**
+    choice
+    (`ultrathink`, "reach for `xhigh` on hard problems") is not a durable pin. **`effort:`
+    frontmatter and `effortLevel` keys route to `claude-config:audit`** on I17's
+    instruction-text-versus-config discriminator. And **schema documentation and its illustrative
+    samples** are fenced **separately** rather than folded into the config fence, because a worked
+    example quoted inside documentation prose is not a key living in a config file and the config
+    fence would not have reached it — the level in a sample demonstrates syntax, not a measured
+    choice. That fence ends where the demonstration does: documenting the field *and then telling the
+    reader which level to put there* is prescribing, and stays in scope.
+
+  - **I22 — model-routing doctrine with no baseline named** (`mechanical`, `OPINION`, `info`,
+    default **off**, enabled by `--opinion`). First-party lane assignments derived from a reading of
+    vendor selection pages, written down with neither the baseline they came from nor an event that
+    re-opens them, become a claim about a model generation that has since passed, told in the
+    present tense. Its own contribution beyond "attach a trigger" is the **delta-not-re-run**
+    discipline: the action on a trigger is a targeted delta check against the named baseline, never
+    a re-derivation from scratch — a trigger nobody can afford to run is not a control.
+
+    **The row carries no baseline of its own, by design.** Naming a date or a vet here would hand
+    every consumer a foreign snapshot as their baseline, reproducing in their repos the exact drift
+    the check exists to catch.
+
+    Its third-party fence is stated **narrowly on purpose**: transcribed practice is out of scope
+    only because there is no vet to point at, **not** because a sync stamp makes it fresh. A stamp
+    tracks whether the transcription is current, never whether the transcribed advice still names a
+    live model — so a stale lane recommendation inside a faithfully synced pack stays stale. That
+    residual is the transcribing surface's to carry, and the fence says so rather than implying the
+    sync path has it covered.
+
+    Its non-duplication is stated in the row rather than assumed. **I19** covers a restated
+    *benchmark figure* and asks for the four-part record; it says nothing about lane assignments and
+    nothing about how to act when a trigger fires. **The catalog-wide recheck trigger** does not
+    reach it either — that trigger governs *this catalog's* staleness against its Sources, not an
+    audited surface's staleness against the pages its doctrine was read from. It ships
+    `Source: none` on I19's footing and adds no Sources entry for the same reason.
+
+### Changed
+
+- **`audit-instructions`: the model-configuration and effort Sources entries name what I21 depends
+  on** — the per-model calibration of the effort scale and the first-run default hold, and `high`'s
+  equivalence to omitting the parameter plus the carry-over sweep advice. The catalog's "the trigger
+  set is the source set" invariant makes these parentheticals load-bearing: a dependency the entry
+  does not name is a dependency nothing watches.
+
+- **`audit-instructions`: `--opinion` no longer restates which rows it enables.** The flag's
+  description in the skill body carried its own copy of the `OPINION` row set, which is a second
+  place to update on every new `OPINION` row and, when stale, silently narrows the flag below what
+  the catalog actually defines. The set is now read from the catalog at run time, where the
+  enablement policy already lives, and the run's tier-transparency line reports the count it found —
+  removing the drift class rather than correcting one instance of it.
+
 ## [0.20.1]
 
 ### Fixed

plugins/claude-config/skills/audit-instructions/SKILL.md

@@ -19,7 +19,7 @@ locally-owned instruction surfaces, cites each finding to current official promp
 it by how confident the evidence can be, and packages proposed removals or rewrites as a human-gated
 diff — so instruction surfaces shrink as models get better instead of only ever growing.
 
-The check catalog — the checks I1–I20, their evidence tier, authority tag, severity, per-surface
+The check catalog — the checks I1–I22, their evidence tier, authority tag, severity, per-surface
 applicability, and the `OPINION`-tier enablement policy — lives in
 [reference/criteria.md](reference/criteria.md); the deterministic pre-scan is
 `${CLAUDE_PLUGIN_ROOT}/skills/audit-instructions/scripts/instruction-scan.sh`.
@@ -43,7 +43,7 @@ concerns its siblings already cover — route rather than re-answer:
   portability is `claude-config:audit-permission-grants`.
 
 On **memory-layer surfaces** (CLAUDE.md, CLAUDE.local.md, `.claude/rules/`, `~/.claude/rules/`),
-this skill runs only the model-era checks I6–I20. It never runs or reports the hygiene checks
+this skill runs only the model-era checks I6–I22. It never runs or reports the hygiene checks
 I1–I5 (line-necessity, length, placement, inferable content, rule-to-hook) on these surfaces —
 that instruction-memory hygiene layer belongs to the `claude-memory` plugin. When that plugin is
 installed, route memory-layer hygiene to its `audit` skill; when it is not installed, emit a single
@@ -105,8 +105,12 @@ scope; non-matching ones are inert and the report lists them as `skipped-for-tar
 
 Two flags govern the `OPINION` tier, whose enablement policy the catalog defines:
 
-- `--opinion` — also run the `OPINION`-tier checks that emit findings (I16 and I19 today). Off by
-  default; their findings are capped at `info` and are never applied.
+- `--opinion` — also run the `OPINION`-tier checks that emit findings. Off by default; their
+  findings are capped at `info` and are never applied. **Which rows those are is read from the
+  catalog at run time and deliberately not restated here**: the catalog owns the enablement policy,
+  so a second copy of the set in this file is one more thing to keep in sync on every new
+  `OPINION` row, and a stale copy silently narrows the flag. The run's tier-transparency line
+  reports how many it found.
 - `--no-stopping-condition` — disable the `OPINION`-tier stopping condition that bounds I6 and I8.
   It is on by default because it withholds findings rather than emitting them, so turning it off
   makes both trimming checks more aggressive, not the audit more conservative.