fix: consistency audit batch — H2, M3, M7 from #142

[dean0x]

Summary

Fixes all HIGH, MEDIUM, and 2 of 3 LOW findings from #142 (post-Wave 2 consistency audit). Only L1 (skill line counts) remains — skipped by design.

Commit 1: H2, M3, M7 code fixes

• H2: ambient-router SKILL.md description updated to standard "This skill should be used when..." format
• M3: simplifier.md agent gains ## Output and ## Boundaries sections (matching scrutinizer pattern)
• M7: Added homepage, repository, license, keywords to 3 plugin.json files (devflow-ambient, devflow-audit-claude, devflow-self-review)

Commit 2: L2, L3 documentation and description sync

• L2: Updated CLAUDE.md to document intentional allowed-tools exceptions (git skills add Bash, interactive skills add AskUserQuestion, persistence skills add Write)
• L3: Synced plugin descriptions across all 3 sources (plugin.json as canonical → plugins.ts + marketplace.json). 9 plugins had divergent descriptions with missing details.

Also resolved in issue (no code changes needed)

• M1: Closed as acceptable — agents under 350-line threshold
• M4: Closed as by-design — specify doesn't load skills, consistent with non-teams variant
• M6: Closed as by-design — agents: [] is correct; claude-md-auditor is plugin-specific, not shared

Previously fixed (PRs #145-147)

H3, H4, H6, M2, M5, M8

Remaining (skipped)

• L1: 15 skills exceed 150-line target — acceptable, all have references/ directories

Test plan

• npm run build — 45 skill copies + 21 agent copies, no errors
• npm test — 250 tests pass
• Spot checks: ambient-router description, simplifier sections, plugin.json metadata, description sync

Closes #142