fix(debrief): raise public coaching budget 1200→6000 so loss coaching stops failing closed

backend/app/services/simulator.py

@@ -412,7 +412,12 @@ class Simulator:
             system,
             user_prompt,
             temperature=0.2,
-            max_tokens=1200,
+            # 6000 (was 1200): same hidden-reasoning truncation class as the final
+            # judge (500 #4). At 1200 the Qwen GGUF's internal chain consumed the
+            # shared cap on a full seller transcript, truncated the JSON, and the
+            # coaching silently failed closed to {} — the loss debrief then showed
+            # only the outcome banner + persona, with no "why" and no coaching.
+            max_tokens=6000,
         )
         return result if isinstance(result, dict) else {}
 

backend/tests/test_llm_budgets.py

@@ -59,6 +59,22 @@ def test_judge_budget_covers_hidden_reasoning():
     assert stub.json_kwargs[0]["max_tokens"] >= 1600
 
 
+def test_public_debrief_budget_covers_hidden_reasoning():
+    # The public coaching pass (why / failurePoints / coaching) analyzes the full
+    # seller transcript. Same hidden-reasoning truncation class as the final judge
+    # (500 #4): at 1200 the internal chain ate the budget, the JSON got truncated,
+    # complete_json threw LLMError, and _public_debrief failed closed to {} — so the
+    # loss debrief showed only the outcome banner + persona, no "why" and no coaching.
+    # 6000 matches the final judge.
+    stub = _CapturingLLM("")
+    Simulator(stub).public_debrief(
+        messages=[{"role": "seller", "text": "สวัสดี"}],
+        outcome="lost",
+        score=30,
+    )
+    assert stub.json_kwargs[0]["max_tokens"] >= 6000
+
+
 def test_final_judge_budget_covers_hidden_reasoning():
     # The FINAL judge (end-of-session verdict) is the fatal path — a truncation
     # there 500s the whole chat send (the 2026-10-02 live incident: JSON cut off