v0.4.7

v0.4.7 — agent-facing text tuned for the Claude 5 family

Prompt-engineering release, grounded in Anthropic's official Prompting Claude Fable 5 guidance and instruction-density research. No behavior changes to the gate, ledger, or CLI surface.

• De-prescribed the subagent briefing (export-contract): direct role statement, five-item Do-not list folded to two (the dropped items restated the anti-drift rule already in the primacy zone). Anthropic: skills written for prior models are "often too prescriptive and can degrade output quality" on Fable 5 — a brief instruction is as effective as enumerating each behavior.
• Tightened SKILL.md's subagent note — three negative instructions collapsed into one direct paragraph. 87/100 lines.
• Audited for the new Fable 5 reasoning_extraction refusal trigger (instructions to echo/transcribe internal reasoning): clean — ritalin asks for evidence checked (file:line cites), never reasoning transcripts.

Run ritalin skill install after updating to deploy the new SKILL.md.