chore(repo): resolve subagent narration + tool-registration failures

[Sawmonabo]

Summary

• YAML tool registration: Claude Code's session-start parser handled tools: ["Read", ...] flow-form inconsistently — same byte format parsed as a real array for some agents (plan-execution-implementer, plan-execution-spec-reviewer) but as a literal string for plan-execution-housekeeper and plan-execution-contract-author, leaving zero tools registered (No such tool available on every Read/Edit/Write). Switched to block-list YAML form so the parser registers the array reliably.
• Narration anti-pattern removal: Agent bodies contained literal `Tool: <Name>\n{...}` strings in failure-mode warnings and "do not output X" sections. Opus 4.7's literal instruction following caused the model to emit those tokens as text content instead of invoking the tool API — totalToolUseCount: 0 across the PR feat(control-plane): compose pg.Pool-backed Querier (Plan-001 T5.5) #36 / feat(sidecar-rust-pty): bootstrap crate + protocol + framing (Plan-024 P1) #42 / feat(contracts): add PtyHost interface (Plan-024 T-024-2-1) #45 / feat(daemon): NodePtyHost + PtyHostSelector (Plan-024 Phase 2) #51 dispatches. Replaced negative-example warnings with positive-only action contracts; defense-in-depth from the orchestrator's existing validator (narration_mode_detected gap, check docs(repo): resolve Tier 1-3 plan governance audit findings #13) and auto-deviation fallback is preserved.
• Housekeeper heading cleanup: Consolidated duplicate ## First action (mandatory) into ## Manifest contents so each heading has a single, distinct purpose (the "first action is Read" rule lives in ## Action contract alone now).

Builds on the guardrails landed in #53.

Test plan

• Post-restart smoke tests against both agents — filesystem-confirmed Edit/Write effects (housekeeper rewrote target.md + manifest preserving _script_stage; contract-author wrote a Zod schema file)
• All 393 plan-execution skill tests pass (including the buildHousekeeperPrompt snapshot test that pins the orchestrator's runtime brief)
• All 411 package tests pass (runtime-daemon 295, control-plane 81, client-sdk 35)
• pnpm typecheck 8/8 successful, pnpm lint 5/5 clean
• Zero orphan refs to removed headings / anti-pattern strings across the repo
• Memory captured at feedback_subagent_authoring_pitfalls documenting both rules + the unknown-root-cause caveat for the YAML parser

🤖 Generated with Claude Code