P0.6.3 — Step 5 Verification

Closes the 5-step chain for P0.6.3 ε Skill Capability Index — the final non-deferred Phase 0 code task. With this merge, the ε Skill Registry axis is complete (3/3 P0.6.* tasks done) and Phase 0 progresses to 26/28 (P0.5.1 + P0.5.2 remain parked behind ADR-005 as spec-only deferrals).

§1. Files shipped

Path	Kind	Commit
`docs/audits/p0-6-3-capability-index-audit.md`	Step 1 audit	`ad19b0e6`
`docs/contracts/p0-6-3-capability-index-contract.md`	Step 2 contract	`dab1d6ac`
`docs/packets/p0-6-3-capability-index-packet.md`	Step 3 packet	`bf124da5`
`src/domains/skills/capability-index.ts`	Step 4 source (new module)	`f7cd8d6a`
`src/domains/skills/repository.ts`	Step 4 source (edit)	`f7cd8d6a`
`src/__tests__/domains/skills/capability-index.test.ts`	Step 4 tests	`f7cd8d6a`
`docs/verification/p0-6-3-capability-index-verification.md`	Step 5 (this file)	pending

Scope fence verified — every changed path matches the allowlist in the dispatch prompt (src/domains/skills/, src/__tests__/domains/skills/, or docs/{audits,contracts,packets,verification}/p0-6-3-*).

§2. Gate results

All three gates run in the worktree E:/AMS/.worktrees/claude/p0-6-3-capability-index after Step 4 committed at f7cd8d6a.

2a. `npm run build`

> colibri@0.0.1 build
> tsc

Exit 0, zero TypeScript diagnostics.

2b. `npm run lint`

> colibri@0.0.1 lint
> eslint src

Exit 0, zero warnings, zero errors.

2c. `npm test`

Test Suites: 24 passed, 24 total
Tests:       1026 passed, 1026 total
Snapshots:   0 total
Time:        27.726 s

Exit 0. All suites green.

Test suite growth: +1 file (capability-index.test.ts), +25 tests (1001 baseline → 1026 post-P0.6.3; baseline count from MEMORY.md tail, verified 2026-04-18).

Coverage:

File	Stmts	Branch	Funcs	Lines
`src/domains/skills/capability-index.ts`	100%	100%	100%	100%
`src/domains/skills/repository.ts`	91.66%	78.94%	89.47%	92.22%
All files	92.25%	84.28%	89.63%	92.52%

capability-index.ts is 100% covered on every axis. repository.ts drops slightly from 92% lines (baseline) because the test-only reset path + new module-level state add lines that are only exercised from within this suite; existing repository coverage paths are unchanged.

2d. Known pre-existing flake

One test suite in the first npm test run flaked on startup-subprocess smoke:

FAIL src/__tests__/startup.test.ts
  ● startup — subprocess smoke › tsx src/server.ts boots and logs [Startup] Phase 1
  Expected pattern: /\[colibri\] starting/
  Received string:  ""

This flake is documented in MEMORY.md under “Pre-existing intermittent startup-subprocess smoke test flakiness (flagged by Waves F and G sub-agents; predates)”. It is not caused by this task — the flake disappeared on the very next npm test invocation, which returned a clean 24 suites / 1026 passed result (reported in §2c above). The verification evidence uses the second, fully-green run.

§3. Acceptance criteria — final trace

AC from dispatch prompt	Status	Evidence
`buildCapabilityIndex` returns `Map<string, Set<string>>`	PASS	`src/domains/skills/capability-index.ts` §buildCapabilityIndex; tested by `buildCapabilityIndex — empty` + 5 more suites
Multi-capability skills in every relevant bucket	PASS	INV1; tested by `buildCapabilityIndex — multi-capability (INV1)` (2 tests) + parity suite
`findSkillsByCapability` returns sorted `string[]`, `[]` for unknown	PASS	INV3 + INV4; tested by `findSkillsByCapability — sorted` (2 tests) + `findSkillsByCapability — unknown` (2 tests)
Index built once in loader, not lazily	PASS	INV8; index is built inline at the tail of `loadSkillsFromDisk`; confirmed by `getCapabilityIndex — before any load` (empty before first load) + `getCapabilityIndex — after loadSkillsFromDisk` (populated after first load)
Parity with `listSkills({capability})` confirmed by test	PASS	INV7; the `parity — INV7` suite iterates 6 capabilities (4 present, 2 absent including an unknown one) and asserts set equality
Zero new MCP tools (ε surface stays `skill_list`)	PASS	No `registerColibriTool` call in `capability-index.ts`; `registerSkillTools` signature unchanged; `grep -n "registerColibriTool" src/domains/skills/capability-index.ts` = 0 matches
No `src/domains/agents/`, no subprocess spawning	PASS	Directory does not exist (`ls src/domains/` confirms); no `child_process`/`worker_threads` import anywhere in diff
No file watcher, no hot-reload path	PASS	No `fs.watch`, no `chokidar`, no interval timer in diff; `grep -n "fs.watch\\|chokidar\\|setInterval" src/domains/skills/capability-index.ts` = 0 matches
`npm run build && npm run lint && npm test` green	PASS	§2a, §2b, §2c

All 9 acceptance criteria satisfied.

§4. Forbidden checks

Per CLAUDE.md §9.1 + dispatch prompt “FORBIDDENS” section:

Check	Command	Result
No `src/domains/agents/`	`ls src/domains/`	absent (expected)
No new MCP tool registration in this task	`grep -n "registerColibriTool\\|register.*Tool" src/domains/skills/capability-index.ts`	0 matches
No `skill_get` / `skill_reload` / `agent_*`	`grep -rn "skill_get\\|skill_reload\\|agent_spawn\\|agent_status\\|agent_list" src/domains/skills/`	0 matches
No file watcher / hot-reload	`grep -rn "fs.watch\\|chokidar\\|fsPromises.watch" src/domains/skills/`	0 matches
No subprocess spawning	`grep -rn "child_process\\|worker_threads\\|spawn(" src/domains/skills/`	0 matches
Scope fence	`git diff --name-only main...HEAD`	all paths in allowlist

All forbidden patterns absent.

§5. Invariant spot-check

Invariant	Test suite	Status
INV1 (multi-capability → all buckets)	`buildCapabilityIndex — multi-capability (INV1)`	PASS
INV2 (Set semantics inside bucket)	`buildCapabilityIndex — duplicate capabilities in one skill (INV2)`	PASS
INV3 (sorted return)	`findSkillsByCapability — sorted (INV3)`	PASS
INV4 (unknown → `[]`)	`findSkillsByCapability — unknown (INV4)`	PASS
INV5 (pure functions)	No logger / no fs / no console in `capability-index.ts` (static audit)	PASS
INV6 (atomic replace, not mutate)	`getCapabilityIndex — rebuild after second load › prior callers that captured the old reference still see the old Map (INV6)`	PASS
INV7 (parity with `listSkills({capability})`)	`parity — listSkills({capability:C}) ≡ findSkillsByCapability(index, C) (INV7)` (2 tests)	PASS
INV8 (build-once per load)	`getCapabilityIndex — before any load` (empty until first load) + `getCapabilityIndex — after loadSkillsFromDisk` (populated after first load)	PASS
INV9 (no agent runtime)	Forbidden checks §4	PASS

All nine invariants from the contract have at least one corresponding test or static audit.

§6. What this closes

ε Skill Registry axis: complete (3/3 P0.6.* tasks done)
- P0.6.1 Skill Schema: shipped R75 Wave C (6c26bb58)
- P0.6.2 Skill CRUD + Discovery + skill_list: shipped R75 Wave D (cb9befaf)
- P0.6.3 Skill Capability Index: this task
Phase 0 code tasks: last non-deferred task. After merge, Phase 0 = 26/28 (P0.5.1 + P0.5.2 are spec-only per ADR-005).
MCP tool surface: unchanged at 14 tools — this task is library-only, per design (Phase 0 ε surface = skill_list only, S17 §1).

§7. Phase 0 carry-over (for Wave H / Phase 0 seal)

This task addressed exactly its own scope. The carry-over items listed in MEMORY.md (CLAUDE.md drift, ADR-006 graduation path, phantom server_info, etc.) are not touched and remain for the Phase 0 seal round.

§8. Writeback

Task ID: P0.6.3
Branch: feature/p0-6-3-capability-index
Worktree: .worktrees/claude/p0-6-3-capability-index
Implement commit: f7cd8d6a (this file will become the final commit; verify commit SHA recorded at push time)
Tests run:
- npm run build → PASS
- npm run lint → PASS (zero warnings)
- npm test → PASS (24 suites / 1026 tests; +25 new tests)
Summary: Implemented the ε Skill Capability Index as a pure in-process reverse lookup Map<capability, Set<skillName>> built once at the end of loadSkillsFromDisk, exposed via a module-level getCapabilityIndex() getter on repository.ts. Parity with listSkills({capability}) verified by a dedicated test suite. No new MCP tools; no src/domains/agents/; no file watcher; no subprocess spawning.
Blockers: None.
PR URL: recorded on PR creation.