feat(session-close): detect self-improve-worthy signals (Phase 8.7)

[bborbe]

What

Adds Phase 8.7 to /vault-cli:session-close — the friction-scoring sibling of Phase 8's reflect detection. When the session shows tooling friction, session-close now nudges /coding:self-improve, exactly the way it already nudges /vault-cli:reflect for knowledge.

Why

/coding:self-improve (just shipped in the coding plugin) reviews a session and proposes ≤2 durable improvements to commands/agents/rules. Without a nudge it rots unused — discoverability is the whole reason Phase 8 exists for reflect. This wires the parallel signal.

Design

• Suggest-only, never auto-run — same contract as reflect (never auto-self-improve added to the Phase 9 guard).
• Separate signal set: reflect = durable knowledge; self-improve = tooling friction (general corrections, repeated instructions, misfired commands, rule violations, uncovered manual workflows). Scored independently; threshold ≥ 3.
• Phase 9 wiring: mode 2 broadened to "reflect and/or self-improve"; outstanding-list gains a parallel item.

Files

• commands/session-close.md — Phase 8.7 + Phase 9 wiring + inline-description accuracy
• CHANGELOG.md — ## Unreleased feat bullet

Depends on /coding:self-improve (bborbe/coding#64, merged).

[ben-s-pull-request-reviewer]

I've completed the review. Here's my analysis:

Files changed (from the PR): CHANGELOG.md and commands/session-close.md

Key findings:

1. ❌ CRITICAL — Version alignment violation: CHANGELOG.md has ## v0.94.0 at the top, but both .claude-plugin/plugin.json and .claude-plugin/marketplace.json still show "version": "0.93.0". Per CLAUDE.md, all four version strings MUST always equal. This will fail make precommit's check-versions.

2. ✅ Phase 8.7 scoring threshold (score ≥ 3) — correctly implemented in lines 294–310. Five signals with weights (+2, +2, +1, +1, +1) and threshold ≥ 3.

3. ✅ Reflect and self-improve are separate — Phase 8 (reflect, lines 207–228) and Phase 8.7 (self-improve, lines 294–310) score independently. No cross-contamination.

4. ✅ Suggest-only contract upheld — Line 296: "suggest-only — never auto-run"; line 310: "Do NOT run /coding:self-improve from here. Only surface the suggestion."

5. ✅ Mode 2 wording — Line 339: "reflect and/or self-improve signals fired" — correctly handles cases where only one fired.

6. ✅ Mode 3 self-improve item — Line 362: correctly appended as a parallel outstanding item.

7. ⚠️ No test coverage — Phase 8.7 scoring logic (signal weights + threshold) has no test file. This is a command/skill markdown file, not Go code, so it can't be unit-tested directly — the scoring behavior would only be verified via scenario testing.

{
  "verdict": "request-changes",
  "summary": "Phase 8.7 self-improve detection is correctly implemented with proper suggest-only contract and independent scoring. However, a critical version alignment violation must be fixed before merge: CHANGELOG.md declares v0.94.0 but plugin JSON files still show 0.93.0.",
  "comments": [
    {
      "file": "CHANGELOG.md",
      "line": 1,
      "severity": "critical",
      "message": "Version mismatch: CHANGELOG.md top entry is '## v0.94.0' but .claude-plugin/plugin.json and .claude-plugin/marketplace.json both show version '0.93.0'. These four version strings must match per CLAUDE.md MANDATORY version alignment. Update plugin.json and marketplace.json to '0.94.0' (both metadata.version and plugins[0].version), then run 'make precommit' to verify."
    },
    {
      "file": "commands/session-close.md",
      "line": 308,
      "severity": "minor",
      "message": "Phase 8.7 scoring logic has no test coverage. The five-signal weighted scoring (score ≥ 3 threshold) cannot be verified automatically. Consider scenario-testing the self-improve nudge output under various signal combinations, or adding a comment in the source noting this is tested via scenario/smoke tests only."
    }
  ],
  "concerns_addressed": [
    "correctness: Phase 8.7 scoring threshold is score >= 3 — correctly implemented in lines 294-310",
    "correctness: reflect and self-improve scored independently — separate phases with no cross-contamination (Phase 8 vs 8.7)",
    "security: design contract suggest-only, never auto-run — confirmed at lines 296 and 310",
    "security: never auto-self-improve guard — confirmed, Phase 9 only surfaces the suggestion, never invokes it",
    "correctness: Phase 9 mode 2 wording 'reflect and/or self-improve' — correctly handles partial signals at line 339",
    "correctness: mode 3 outstanding list self-improve item — correctly appended at line 362",
    "tests: no test files modified — Phase 8.7 is command markdown (not testable via unit tests); scoring verified via doc review"
  ]
}