Removed the fixed-value text detector. persona_reply no longer forces JSON meta; instead a
per-turn evaluate_turn() calls the judge LLM after every customer reply to read the persona's
current mood + whether it has decided (buy/walk/pending) + score_delta + reason. send_message
consumes that context-based decision to (a) end the chat as won/lost and (b) move the score.
This is what the user asked: the system evaluates EVERY turn and decides at the moment it's
truly committed — not keyword matching (so 'ซื้อไม่ไหว แต่ว่ามีผ่อนไหม?' stays pending).
Mock updated: judge returns buy on first send (keeps E2E deterministic). 11/11 suites pass.
The conversation now ENDS when the persona makes a decision (option C), not when the
trainee clicks a button:
- persona_id replies carry {reply, decision(none/buy/walk), mood}; when decision is
buy/walk the session auto-finishes (won/lost) with a debrief that reveals latent
details + per-turn 'turning points'.
- personas have a tolerance (1-5, 'temper'): impatient personas walk away fast after
poor answers (fed via internal.misses on mood<=-1); tough 'wrong text' cases can
still be won by a strong, gentle response (judge realism).
- trainee 'Finish' button removed; if they leave mid-chat an active session is resumed
via /chat/resume (continue, not restart). One-shot lock still enforced once decided.
- mock/tests updated: persona deciding buy -> send auto-finishes won.
Rebuilt dist.
