Implement the approved narrow coaching pipeline from /home/cube/projects/richard/traning coach/.gjc/_session-019f8455-334a-7000-99ca-318dfd0e06b1/plans/ralplan/5d6f0f32-0de0-4f65-9a78-c57de328be71/pending-approval.md (sha256 088e8b91d8966cc9638c0216221bcb04557e51715df390a9ee3c9f311128b1c2): approved Choi-coach principle retrieval and verified customer memory/playbook inputs, typed Choi-coach draft decisions, deterministic fact/safety validation, separate Korean expression polishing, final invariant validation, and privacy-safe evidence receipts. Preserve code-authoritative facts, decisions, actions, timing, safety, approval/delivery state, current Telegram routing, and canonical fallback. Modify only the existing Korean humanizer/coaching grounding integration and directly related tests/docs. Use fallback OpenAI agents only, never Kimi/Kimi3. Do not send Telegram messages, call live providers, deploy, restart services, modify synthetic data, credentials, or unrelated code.