fix(orchestration): record durable worker handles + correct in-session SendMessage revival

[DumoeDss]

Why

A real autopilot run failed to resume its workers because of two coupled defects in the orchestration-worker-lifecycle area.

1. The LEAD recorded each stage worker with a fabricated name field instead of the durable agentId + transcript the Agent tool actually returned, so collectStageWorkers silently dropped the record — the worker became invisible to rasen pipeline resume, forcing a cold reconstruction and losing the warm-seed path.
2. The generated playbook asserts that SendMessage-ing a completed worker within a live session revives it (by name) — but the transcript proves a completed Agent-tool subagent is NOT reliably name-addressable even ~27 messages later in the same un-compacted session (the harness replied "No agent named 'implementer' is currently addressable … use the agent ID"). The playbook's revival claim is wrong, and the run-state gap that made the failure silent had no validation to catch it.

What Changes

• Durable handles in run-state. The generated playbook's Step B dispatch instructions now tell the LEAD to capture agentId + transcript from the Agent (Task) tool's spawn RESULT and write them into the stage's worker record (Claude), or threadId + transcript/turnId (Codex) — never a made-up name. The Worker schema fields stay optional and .passthrough() so archived auto-run.json still parses.
• Validation, not silence. rasen pipeline resume surfaces a non-fatal warning (json + text) for every stage whose worker record lacks any durable handle (agentId / transcript / threadId) — i.e. a name-only or role-only record — naming the offending stage and the non-durable keys present. A new duplicate-key detector likewise warns (non-fatal) when auto-run.json carries duplicate JSON keys, which JSON.parse silently collapses.
• Correct the revival claim. Rewrote the playbook's Step A, Step F.1 (incl. the "two are the SAME mechanism" note), and Step H.4a(b)/H.4b revival paths to prescribe agentId-first re-engagement: name is a non-durable dispatch label, NOT a resume handle; a completed Agent-tool subagent may be unreachable even within the same session; fall back to the transcript warm-seed (Tier-B) when agentId is absent or does not resolve. Code comments in run-state.ts (~L34–43) and pipeline.ts (~L423–425) aligned with this reality.
• Hidden bugs (in-scope). collectStageWorkers silent drop → caught by the resume warning; the unknown name field → flagged by name in the warning (passthrough preserved); Tier-A honesty → Step A text and the claude-settings.ts header comment stop overclaiming re-addressability.
• Tests. Extended run-state.test.ts (worker-handle validation helper, duplicate-key detector, collectStageWorkers behavior unchanged for durable handles) and pipeline.test.ts (resume warns on a name-only worker and on duplicate keys, in json + text; a durable-handle worker warns nothing).

No breaking CLI changes. No new required schema fields. Archived run-state continues to parse.

Impact

• Generated playbook (src/core/templates/workflows/_orchestration.ts): Step A (tier honesty), Step B (capture agentId+transcript from the spawn result), Step F.1 (agentId-first revival + the "same mechanism" note), Step H.4a(b)/H.4b (agentId-first revival). Embeds via ORCHESTRATION_PLAYBOOK into auto.ts, review-cycle.ts, goal-command.ts — their parity hashes in test/core/templates/skill-templates-parity.test.ts move.
• src/core/templates/workflows/auto.ts: the Resume-section comment that restated the (wrong) within-session revival claim is aligned.
• src/core/pipeline-registry/run-state.ts: doc comment (~L34–43) aligned; new non-fatal helpers — a duplicate-key detector over the raw JSON text and a worker-handle-validation helper. RunStateWorkerSchema stays optional + .passthrough().
• src/commands/pipeline.ts: resume surfaces workerHandleWarnings + duplicate-key warnings (json + text); the "dead SendMessage handles" comment (~L423–425) is aligned.
• src/core/claude-settings.ts: header doc comment (~L1–9) stops overclaiming Tier-A re-review guarantees.
• Tests: test/core/pipeline-registry/run-state.test.ts, test/commands/pipeline.test.ts, and the parity hash list in test/core/templates/skill-templates-parity.test.ts.

No runtime tier detection is added (tier is LEAD-self-reported from the playbook); adding a runtime env-probe is explicitly out of scope. No auto-run.json is written or edited by this change.

Test plan

• npx tsc --noEmit exit 0
• pnpm lint 0 errors
• pnpm test 2526 passed / 0 failed
• pnpm run build ok
• affected test/commands/pipeline.test.ts 46 passed / 0 failed (re-review)
• rasen validate clean; isComplete: true

🤖 Generated with Claude Code