role prompts: emit loop records as schema-validated JSON

[sawy3r]

Part of the records-as-JSON protocol push (sworn ADR-0009). Once the loop records have published schemas (board/spec/proof/journeys + reconciled slice-status), the four role prompts must instruct each role to emit records as schema-validated JSON (structured output), not as hand-written markdown that a parser later scrapes. This is what makes the protocol portable across consumption models (manual slash commands, automation, end-to-end orchestration).

Scope — claude/baton/role-prompts/{planner,implementer,verifier,captain}.md

• Planner: emit board.json + per-slice spec.json against their schemas; intake stays prose (LLM-drafted, human-reviewed).
• Implementer: emit proof.json (+ render proof.md); update status.json against slice-status-v1.
• Verifier: read records as JSON; emit verdict into the status record; never hand-edit prose to certify.
• Captain: design decisions captured as structured fields (option + rationale), not free markdown.
• Across all: a human supplies conversation + a sentence of rationale; the role produces the JSON. No flow asks a human to hand-edit a record's JSON.

Depends on the schema issues landing first (prompts reference the schemas).
Ref: sworn ADR-0009 (emitted-vs-hand-authored framing; rendering is the implementation's job).