refactor(chat): decide buy/walk by per-turn LLM judge (not fixed keywords)

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.

backend/app/api/chat_routes.py

@@ -27,39 +27,6 @@ def _sim(group, persona):
     return Simulator(llm)
 
 
-WALK_FRAGMENTS = [
-    "no thanks", "not interested", "forget it", "never mind", "no longer", "out of budget",
-    "too expensive", "can't afford", "cannot afford", "ซื้อไม่ไหว", "ไม่ซื้อ", "ไม่เอาแล้ว",
-    "ไม่สนใจ", "พอแค่นี้", "ไม่ไหว", "แพงไป", "ลืมไปเถอะ", "ตัดสินใจไม่ซื้อ", "ขอไม่ซื้อ",
-    "ยังไม่ซื้อ", "ไว้ก่อน", "ไปก่อนนะ", "ขอบคุณแต่ไม่เอา", "ไม่เอาครับ", "ไม่เอาค่ะ", "ไม่เอา",
-]
-BUY_FRAGMENTS = [
-    "i'll take it", "i'll go with it", "i'll buy", "deal,", "let's do it", "count me in",
-    "how do i pay", "where do i sign", "ok i'll buy", "ซื้อเลย", "ตกลงซื้อ", "เอาครับ", "เอาค่ะ",
-    "ซื้อครับ", "ซื้อค่ะ", "ตัดสินใจซื้อ", "เอาเลย", "ตกลง", "ซื้อ", "รับไว้", "โอเคค่ะ รับ", "โอเคครับ รับ",
-]
-
-
-def _detect_customer_decision(text: str, locale: str = "th") -> str | None:
-    """Detect whether the customer clearly walked away or decided to buy, from their message.
-
-    Returns 'walk' | 'buy' | None. Used as a fallback because real LLMs rarely emit a
-    structured meta.decision — they just state it in the conversation.
-    """
-    if not text:
-        return None
-    low = text.lower()
-    # Walk-away signals can appear anywhere; be conservative to avoid false ends on
-    # phrases like "I can't afford that, what's the discount?" (question) vs "too expensive, no."
-    for frag in WALK_FRAGMENTS:
-        if frag in low:
-            return "walk"
-    for frag in BUY_FRAGMENTS:
-        if frag in low:
-            return "buy"
-    return None
-
-
 def _scenarios(locale: str = "th"):
     """Scenario presets, localized. Returns {id: {label, init, adapt}}."""
     t = locale != "en"
@@ -291,16 +258,6 @@ def send_message(gid: str, pid: str):
     internal.setdefault("turns", 0)
     internal["turns"] = internal.get("turns", 0) + 1
     internal["signals"] = internal.get("signals", [])
-    try:
-        mood = int(meta.get("mood", 0))
-    except (TypeError, ValueError):
-        mood = 0
-    # A clearly annoyed customer is a "miss" against the seller.
-    if mood <= -1:
-        internal["misses"] = internal.get("misses", 0) + 1
-        internal["signals"].append({"turn": internal["turns"], "mood": mood, "type": "annoy"})
-    elif mood >= 1:
-        internal["signals"].append({"turn": internal["turns"], "mood": mood, "type": "warm"})
 
     # Re-contact persona behavior: after enough info is exchanged (turn 2), the customer
     # goes quiet, a time-lapse system note is shown, and the customer re-engages warmer.
@@ -316,13 +273,31 @@ def send_message(gid: str, pid: str):
         # Save the time-lapse note immediately so the UI shows it even if send ends here.
         s["sessions"].update(session["id"], messages=messages, internal=internal)
 
-    # Decision by the persona ends the session (one-shot lock).
-    decision = meta.get("decision")
-    # Real LLMs usually DON'T emit a structured meta.decision — they just say it in the
-    # customer's reply. Fall back to a lightweight text detector so "ซื้อไม่ไหว"/"no thanks"
-    # actually ENDS the chat (otherwise it never reports win/loss).
-    if decision not in ("buy", "walk"):
-        decision = _detect_customer_decision(reply, locale=slocale)
+    # Evaluate this turn via the (judge) LLM: how the persona feels + whether it has decided.
+    # This is context-based (NOT fixed keywords), so e.g. "ซื้อไม่ไหว แต่ว่ามีผ่อนไหม?" stays
+    # pending until the customer truly commits to (or abandons) the decision.
+    turn_eval = sim.evaluate_turn(
+        persona=persona, messages=messages, internal=internal
+    )
+    decision = turn_eval.get("decision", "pending")
+    try:
+        mood = int(turn_eval.get("mood", 0))
+    except (TypeError, ValueError):
+        mood = 0
+    # Apply the judge's score delta to internal score trend.
+    try:
+        sd = int(turn_eval.get("score_delta", 0))
+    except (TypeError, ValueError):
+        sd = 0
+    internal["score"] = max(0, min(100, int(internal.get("score", 50)) + sd))
+    internal["last_reason"] = turn_eval.get("reason", "")
+    # Track mood trend for debrief.
+    if mood <= -1:
+        internal["misses"] = internal.get("misses", 0) + 1
+        internal["signals"].append({"turn": internal["turns"], "mood": mood, "type": "annoy"})
+    elif mood >= 1:
+        internal["signals"].append({"turn": internal["turns"], "mood": mood, "type": "warm"})
+
     if decision in ("buy", "walk"):
         outcome = "won" if decision == "buy" else "lost"
         debrief = _build_abbrev_debrief(outcome, persona, internal, slocale)