refactor(chat): AI-native prompts, remove interview heuristics

- Drop interview_reply_length and utterance_substance; always run stage LLM
  and memory retrieval when enabled; trim Settings fields and .env.example.
- Replace guided/opening prompts with compact fact blocks plus unified
  behavior guidance; slim background_voice and persona to tone hints.
- InterviewAgent uses fixed chat_interview max_tokens/chars/segments.

Also includes stacked work: profile followup/extract path, evaluation rubric
and judge schema updates, transcript SPLIT handling in execution service,
user export markdown split tests, and golden case fixture.

api/app/features/evaluation/execution_service.py

@@ -51,6 +51,11 @@ def _utterances_for_case(case: EvalCase) -> list[str]:
     return [str(u).strip() for u in raw if str(u).strip()]
 
 
+def _assistant_text_for_eval_display(raw: str) -> str:
+    """评审与 transcript 展示：避免字面量 [SPLIT] 干扰 judge 阅读。"""
+    return (raw or "").replace("[SPLIT]", "\n")
+
+
 async def execute_eval_run(
     db: AsyncSession,
     *,
@@ -127,12 +132,14 @@ async def execute_eval_run(
     for i, u in enumerate(utterances):
         if i >= len(replies):
             break
-        transcript_parts.append(f"用户: {u}\nAI: {replies[i]}")
+        transcript_parts.append(
+            f"用户: {u}\nAI: {_assistant_text_for_eval_display(replies[i])}"
+        )
     prior = ""
     for idx, u in enumerate(utterances):
         if idx >= len(replies):
             break
-        reply = replies[idx]
+        reply = _assistant_text_for_eval_display(replies[idx])
         lat = latencies[idx] if idx < len(latencies) else None
         tj = await judge.judge_turn(
             prior_transcript=prior,
@@ -146,7 +153,7 @@ async def execute_eval_run(
             run_id=run.id,
             turn_index=idx,
             user_utterance=u,
-            assistant_reply=reply,
+            assistant_reply=replies[idx],
             duration_ms=lat,
             judge_scores_json=scores,
             judge_rationale=rationale,