R76.H6 Audit — 22→23 skill count reconciliation + docs health

Dual-purpose. This document is both Step 1 of the 5-step executor chain AND the R76 Wave 2 gate evidence report for Sigma.

Context

R76.H5 (PR #157) added colibri-sigma-orchestrator as the 23rd canonical .agents/skills/colibri-* skill. The on-disk inventory, previously 22, is now 23. The H5 verification doc flagged a set of files whose “22 canonical” claims became stale at merge time. R76.H6 reconciles them and performs a read-only docs health scan so the exit state of Wave 2 is fully documented.

Part A — 22 → 23 reconciliation targets

A.1 Ground truth

$ ls -d .agents/skills/colibri-*/ | wc -l
23

23 canonical colibri-* skills as of 92ad99a7.

A.2 Files to edit (live canon)

Each row has the exact occurrence(s) to change. Replacement is a literal 22 → 23 in the context of “canonical colibri-* skills” only. Enumerated skill lists also get a colibri-sigma-orchestrator row inserted.

#	File	Line(s)	Type of change
1	`CLAUDE.md`	159	zone-manifest cell: `22 canonical` → `23 canonical`
2	`AGENTS.md`	27	section heading: `(22 canonical ...)` → `(23 canonical ...)`
3	`.agents/README.md`	7, 15, 20	R75 note + table cell + section heading; insert `colibri-sigma-orchestrator` in the PM & Orchestration list (§Skills)
4	`.claude/README.md`	17, 18, 23, 27	ASCII tree (2 lines) + “Skills (22 total)” + “Canonical (22 colibri-*)” subheading; insert `colibri-sigma-orchestrator` row in table
5	`README.md` (root)	158	ASCII tree line: `22 canonical colibri-* skill definitions` → `23 canonical ...`
6	`docs/colibri-system.md`	54	inventory table row: `22 canonical colibri-` → `23 canonical colibri-`
7	`docs/world-schema.md`	242	ASCII tree: `22 colibri-* skills in .agents/skills/` → `23 ...`
8	`docs/3-world/execution/index.md`	69, 127	`22 canonical skills` → `23 canonical skills` (two occurrences)
9	`docs/reference/cross-reference.md`	238	`22 canonical colibri-` → `23 canonical colibri-`
10	`docs/reference/glossary.md`	95, 281	ε definition + Skill definition: both `22 canonical` → `23 canonical`
11	`docs/guides/quick-start.md`	57	`22 canonical colibri-` → `23 canonical colibri-`

Count: 11 files, ~17 literal 22 → 23 edits, plus 2 enumerated list insertions (.agents/README.md and .claude/README.md).

A.3 Explicit NOT-to-edit

These matched the grep but are deliberately preserved because they are historical records that document the 22-skill state at an earlier commit:

docs/audits/chore-rename-repo-facing-polish-audit.md (R75 Wave F state)
docs/audits/p0-6-1-skill-schema-audit.md (R75 Wave C state)
docs/contracts/chore-rename-repo-facing-polish-contract.md (R75 Wave F state)
docs/packets/chore-rename-repo-facing-polish-packet.md (R75 Wave F state)
docs/packets/r75-plan-red-packet.md (R75 state)
docs/verification/chore-rename-repo-facing-polish-verification.md (R75 Wave F state)
docs/verification/r75-plan-red-verification.md (R75 state)
docs/verification/r76-h5-sigma-skill-verification.md (this PR’s own ancestor — documents the 22-state pre-bump by design)
.agents/spawns/colibri-research/r2/source-universe-ams.md (heritage/research)
docs/3-world/execution/scale.md (lines 724, 739 — “22 skills × any number of agents” is a heritage AMS-donor scale analogy; scale.md is a spec-scale document, not a live enumeration)
docs/guides/self-build-loop.md line 526 — “All 22 skills exist…” is a narrative claim about a donor state; file is a guide about the self-build lifecycle, treat as semi-historical

(We also leave docs/PLAN-RED.md and R75 history statements in CLAUDE.md’s changelog tail unchanged per task scope.)

A.4 Out-of-scope numeric matches (confirmed NOT 22-skill)

Regex sweep surfaced no false positives — no “22” anywhere in live canon refers to something other than skill count or the explicitly-historical exceptions above.

Part B — Docs health findings

B.1 Directory doc counts (verified)

Zone	Context block claim	Measured (R76.H5 head = `92ad99a7`)	Status
`data/`	6	2	DRIFT — context claimed 6 (stale from CLAUDE.md §9.2); actual `*.md` count is 2. The CLAUDE.md §9.2 figure is a ballpark from Wave H and may include staging artifacts; the tracked-doc count is 2. Defer reconcile to R77 (not safety-critical — `data/` is HERITAGE zone).
`docs/`	~242	274	DRIFT — grew ~32 docs across Waves H + I + R76 waves 1&2 (audits/contracts/packets/verification + ADR expansions). CLAUDE.md §9.2 still says `~242`. Defer to R77 (`~` already signals approximate; R77 can refresh the number).
`.claude/`	30 (after H5)	30	OK — matches post-H5 expected value.
`.agents/`	258 (after H5)	242	DRIFT — context expected ~258 but actual is 242. Likely the “258” figure in context was speculative; CLAUDE.md §9.2 says 257 which is also stale. Defer to R77 (`.agents/` is mostly heritage spawns; count fluctuations are normal).

None of these drifts are safety-critical. They belong to the same class as the “22 canonical” drift this PR fixes — number decay per feedback_number_decay.md. R77 hygiene candidate.

B.2 Link integrity

grep for ](file:// → zero hits in docs/.
grep for ](/ absolute-from-root patterns → zero hits in docs/.

No broken link patterns introduced by R76 waves 1 & 2 merges (#152–#157). Link integrity clean.

B.3 Mirror divergence inventory

diff -q of every .agents/skills/colibri-*/SKILL.md vs .claude/skills/colibri-*/SKILL.md:

Content-diverged (3) — the known H3-deferred set:

colibri-growth-strategy
colibri-mcp-server
colibri-tier1-chains

These three were explicitly left out of R76.H3 (PR #155) mirror resync scope because they have complex donor-era tool-name annotations that need careful review. Carry into R77 resync wave.

Cosmetic YAML quote-style diverged (3):

colibri-autonomous — canon description: is double-quoted; mirror is unquoted. Both parse.
colibri-docs-sync — canon description: is single-quoted (contains "); mirror is unquoted. Both parse.
colibri-observability — canon description: is double-quoted; mirror is unquoted. Both parse.

These 3 are style-only divergences — both versions convey identical meaning and parse as identical YAML. Not safety-critical. Defer to R77 resync wave (best folded into a single batch with the 3 above).

Total diverged: 6 (3 content + 3 style). Context block expected 3; actual is 6. The delta is entirely style-quote-only YAML, which is why H3’s resync may have considered them “effectively identical”. Flagged for R77.

B.4 Phase 0 reality — dangerous phrases

Grep for Phase 0 not started: one match inside CLAUDE.md changelog tail explicitly recording it was purged — acceptable (historical prose).

Grep for server_info / task_transition / audit_session_end / server_shutdown in live SKILL.md files: many matches, all in explanatory-reality-stamp context (e.g., “server_info from the R74.5 plan was never implemented” — citing ADR-004 as authority). No live doc actively claims these are shipped tools. Pattern is safe and intentional per R75 Wave H purge.

B.5 YAML frontmatter safety

.agents/skills/colibri-sigma-orchestrator/SKILL.md (the new R76.H5 file):

---
name: colibri-sigma-orchestrator
description: "Colibri Sigma Orchestrator (T1) contract-reference skill — Phase A (plan manifest) then Phase W (watch waves) then Phase B (seal). Sigma plans rounds, dispatches PM, gates waves, and seals with audit_verify_chain + merkle_finalize. Sigma does NOT write code. Use when the user opens a new round, asks for round planning, wave gate verification, or round seal. Greek letter alpha — System Core + round choreography. No executable tool dispatch in Phase 0 — agent runtime deferred to Phase 1.5 per ADR-005. Contract at docs/agents/sigma-orchestrator.md."
round: R76
updated: 2026-04-18
---

Description is double-quoted (wraps a string containing : — matches feedback_yaml_frontmatter_hazard.md). Parses cleanly via gray-matter. ✓

Summary for Sigma (Wave 2 gate evidence)

Wave 2 tasks (R76.H4, R76.H5, R76.H6)

Task	PR	Status	Gate evidence
R76.H4 (inventory-r74-2 ref cleanup)	#156	Merged	Listed in `.agents/spawns/r76-phase-0-to-1-bridge/manifest.md`
R76.H5 (colibri-sigma-orchestrator skill)	#157	Merged	`.agents/skills/colibri-sigma-orchestrator/SKILL.md` exists + test assertion bumped to 23
R76.H6 (docs health + 22→23 reconcile)	this PR	Open	This document + build/lint/test green

Wave 2 exit criteria

H5 skill present on disk, YAML clean, test suite assertion bumped to 23
11 live-canon files updated from 22 → 23 (this PR)
Repo-wide grep for "22 canonical \colibri-`”` returns zero hits in live canon (historical records excluded)
Link integrity clean
Mirror divergence inventory enumerated (6 total; 3 H3-deferred + 3 style)
Build + lint + test green
No safety-critical phrase regressions

Deferred to R77

.claude/skills/ mirror resync for the 6 diverged skills (3 H3-deferred content + 3 style-only YAML quote)
Directory doc-count reconciliation (CLAUDE.md §9.2 .agents/, docs/, data/ rows)
docs/PLAN-RED.md audit (currently labelled superseded)
.agents/spawns/colibri-research/r2/source-universe-ams.md heritage-banner review

Wave 2 is gated OK. Wave 3 (P1 κ prompts, P2 δ Phase 1.5 plan, P3 roadmap update) authorized.