feat(api): 访谈人格/回复长度策略、口述归一、背景语气与输入净稿全链路

Chat 访谈
- 新增 persona 系统（default / warm_listener / curious_guide）与 background_voice 语气层
- 回复长度由 compute_reply_plan 统一决策（brief / standard / expanded），融合信息密度启发式
- 输入净稿（input_normalize）：编排层可选 rules/llm 归一用户口语后再喂模型与记忆检索
- 记忆证据注入：按用户话检索 memory evidence 并注入 prompt

Memoir 回忆录
- 口述归一（oral_normalize）：segment 原文保留，story 管线取派生净稿作叙事输入
- segment 入队批次门闸：累计字数 + 最长等待秒数，减少零碎提交
- fidelity_check / prompts / narrative_agent 微调
- Alembic 0005：清理跨章节 story 外键

Infra
- Dockerfile 加入 ffmpeg
- pyproject.toml 新增依赖并同步 uv.lock
- .env.example / .env.production 补全新配置项

Tests
- 新增 test_background_voice、test_chat_input_normalize、test_experience_regressions
- 扩展 test_interview_prompts、test_interview_reply_length、test_story_route_oral_invariant

Made-with: Cursor

api/tests/test_interview_reply_length.py

@@ -0,0 +1,203 @@
+"""访谈回复长度策略：分桶与 InterviewAgent 的 max_tokens / 截断联动。"""
+
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from app.agents.chat.interview_reply_length import (
+    ReplyLengthMode,
+    bump_reply_length_strategy_for_background_voice,
+    compute_reply_length_strategy,
+    compute_reply_plan,
+)
+from app.agents.state_schema import MemoirStateSchema
+
+
+def _fake_settings(**overrides: object) -> SimpleNamespace:
+    base = {
+        "chat_interview_max_tokens": 380,
+        "chat_interview_max_segments": 2,
+        "chat_interview_max_chars_per_segment": 260,
+        "chat_interview_brief_max_tokens": 260,
+        "chat_interview_brief_max_chars_per_segment": 200,
+        "chat_interview_expanded_max_tokens": 520,
+        "chat_interview_expanded_max_chars_per_segment": 380,
+    }
+    base.update(overrides)
+    return SimpleNamespace(**base)
+
+
+def test_strategy_brief_when_very_short() -> None:
+    s = compute_reply_length_strategy(
+        5,
+        likely_new_detail=False,
+        likely_chit_chat=False,
+        settings=_fake_settings(),
+    )
+    assert s.mode == ReplyLengthMode.brief
+    assert s.max_tokens == 260
+    assert s.max_chars_per_segment == 200
+
+
+def test_strategy_standard_mid_length() -> None:
+    s = compute_reply_length_strategy(
+        50,
+        likely_new_detail=True,
+        likely_chit_chat=False,
+        settings=_fake_settings(),
+    )
+    assert s.mode == ReplyLengthMode.standard
+    assert s.max_tokens == 380
+    assert s.max_chars_per_segment == 260
+
+
+def test_strategy_long_chit_stays_standard() -> None:
+    s = compute_reply_length_strategy(
+        120,
+        likely_new_detail=False,
+        likely_chit_chat=True,
+        settings=_fake_settings(),
+    )
+    assert s.mode == ReplyLengthMode.standard
+    assert s.max_tokens == 380
+
+
+def test_strategy_long_with_new_detail_expanded() -> None:
+    s = compute_reply_length_strategy(
+        120,
+        likely_new_detail=True,
+        likely_chit_chat=False,
+        settings=_fake_settings(),
+    )
+    assert s.mode == ReplyLengthMode.expanded
+    assert s.max_tokens == 520
+    assert s.max_chars_per_segment == 380
+
+
+def test_strategy_boundary_len_20_brief_len_21_standard() -> None:
+    a = compute_reply_length_strategy(
+        20, likely_new_detail=False, likely_chit_chat=False, settings=_fake_settings()
+    )
+    b = compute_reply_length_strategy(
+        21, likely_new_detail=False, likely_chit_chat=False, settings=_fake_settings()
+    )
+    assert a.mode == ReplyLengthMode.brief
+    assert b.mode == ReplyLengthMode.standard
+
+
+def test_bump_standard_only_for_cadre_military() -> None:
+    s0 = compute_reply_length_strategy(
+        50,
+        likely_new_detail=False,
+        likely_chit_chat=False,
+        settings=_fake_settings(),
+    )
+    bumped = bump_reply_length_strategy_for_background_voice(
+        s0,
+        background_voice="cadre",
+        settings=_fake_settings(
+            chat_interview_cadre_military_standard_extra_tokens=40,
+            chat_interview_cadre_military_standard_extra_chars=40,
+        ),
+    )
+    assert bumped.max_tokens == s0.max_tokens + 40
+    assert bumped.max_chars_per_segment == s0.max_chars_per_segment + 40
+
+    brief = compute_reply_length_strategy(
+        5,
+        likely_new_detail=False,
+        likely_chit_chat=False,
+        settings=_fake_settings(
+            chat_interview_cadre_military_standard_extra_tokens=40,
+            chat_interview_cadre_military_standard_extra_chars=40,
+        ),
+    )
+    same = bump_reply_length_strategy_for_background_voice(
+        brief,
+        background_voice="military",
+        settings=_fake_settings(
+            chat_interview_cadre_military_standard_extra_tokens=40,
+            chat_interview_cadre_military_standard_extra_chars=40,
+        ),
+    )
+    assert same.max_tokens == brief.max_tokens
+
+
+def test_plan_short_information_rich_is_standard_not_brief() -> None:
+    """短句但含高密度锚点（如「那年」「我爸」）→ standard，避免误压成 brief。"""
+    p = compute_reply_plan(
+        "那年我爸突然病了",
+        background_voice=None,
+        settings=_fake_settings(),
+    )
+    assert p.mode == ReplyLengthMode.standard
+    assert p.information_rich is True
+
+
+def test_plan_long_chit_stays_standard_not_expanded() -> None:
+    """长段明显闲聊 → standard，不因字数进入 expanded。"""
+    msg = "今天天气真好哈哈" * 11
+    assert len(msg) >= 80
+    p = compute_reply_plan(
+        msg,
+        background_voice=None,
+        settings=_fake_settings(),
+    )
+    assert p.mode == ReplyLengthMode.standard
+    assert p.likely_chit_chat is True
+
+
+def test_strategy_boundary_len_79_standard_len_80_long_branch() -> None:
+    a = compute_reply_length_strategy(
+        79, likely_new_detail=False, likely_chit_chat=False, settings=_fake_settings()
+    )
+    b = compute_reply_length_strategy(
+        80,
+        likely_new_detail=False,
+        likely_chit_chat=False,
+        settings=_fake_settings(),
+    )
+    assert a.mode == ReplyLengthMode.standard
+    assert b.mode == ReplyLengthMode.standard
+
+
+@pytest.mark.asyncio
+async def test_interview_agent_passes_strategy_to_bind_and_truncate() -> None:
+    """同一套 strategy 用于 llm.bind(max_tokens=) 与 truncate_chat_segments。"""
+    from app.agents.chat import interview_agent as ia
+
+    mock_llm = MagicMock()
+    mock_bound = MagicMock()
+    mock_bound.ainvoke = AsyncMock(
+        return_value=MagicMock(content="你好，后来呢？[SPLIT]还有吗？")
+    )
+    mock_llm.bind = MagicMock(return_value=mock_bound)
+
+    agent = ia.InterviewAgent()
+    agent.llm = mock_llm
+
+    state = MemoirStateSchema(
+        stage_order=["childhood"],
+        current_stage="childhood",
+        covered_stages=[],
+        slots={"childhood": {}},
+    )
+
+    with patch(
+        "app.agents.chat.interview_agent.get_history_messages",
+        new=AsyncMock(return_value=[]),
+    ):
+        turn = await agent.generate_response_with_state(
+            conversation_id="c1",
+            user_message="x" * 100 + "第一次认识他",
+            memoir_state=state,
+        )
+
+    mock_llm.bind.assert_called_once()
+    call_kw = mock_llm.bind.call_args[1]
+    assert call_kw["max_tokens"] == 520
+
+    assert len(turn.messages) >= 1
+    for seg in turn.messages:
+        assert len(seg) <= 380