feat: add three-layer swarm command conflict prevention system

[zaxbysauce]

Summary

• Three-layer defense preventing Claude Code command confusion inside swarm agent sessions
• Layer 1 (Conflict Registry): Pure data module with 9 command conflict entries, CI gate with ratchet mechanism, clashesWithNativeCcCommand field on CommandEntry, frozen ReadonlySet of 105 CC built-in commands
• Layer 2 (Prompt Hardening): COMMAND NAMESPACE blocks in all 4 agent prompts, skill files, session contract, and commands documentation
• Layer 3 (Runtime Intercept): Message pipeline hook that hard-blocks /reset/clear, soft-corrects /plan, logs HIGH advisories — 45 tests (23 unit + 22 adversarial), under 5ms design target

Test plan

• bun run typecheck — clean
• bunx biome ci . — clean
• 109 tests pass across 6 new/modified test files
• 22 adversarial tests cover 8 attack vectors (case variation, whitespace, code fences, URLs, comments, multi-command, corrected-form bypass, case evasion)

Invariant audit

• 1 (plugin init): not touched
• 2 (runtime portability): not touched
• 3 (subprocesses): not touched
• 4 (.swarm containment): not touched
• 5 (plan durability): not touched
• 6 (test_runner safety): not touched — all test runs use explicit file paths, no broad scopes
• 7 (test writing): touched — tests follow bun:test patterns, mock.module not used
• 8 (session state): not touched
• 9 (guardrails/retry): not touched
• 10 (chat/system msg): touched — cc-command-intercept hook operates on messages.transform chain
• 11 (tool registration): not touched
• 12 (release/cache): not touched

Pre-existing failures

• 148 test failures in tests/unit/commands/ (resolveCommand returning undefined for some keys) and tests/unit/config/ (project-init timing) predate this PR — verified against origin/main