feat(memory,conversation): 记忆富化/证据包、时间线幂等字段与对话分段全链路

数据库
- 新增迁移 0003：timeline_events.memory_source_id 外键 → memory_sources，便于按 ingest 源做时间线幂等

后端 - 记忆
- 新增 ingest 后 LLM 富化（摘要/事实/时间线），可配置开关与最大字符数
- 新增证据包组装：合并 chunk、摘要、事实、时间线、故事等检索结果；支持空 query 时是否仍带 rolling 等开关
- repo/retriever/service/router/schemas/summarizer/timeline/extractor 等扩展；文档 memory-retrieval.md 更新

后端 - 对话 WS
- 增加 PING/PONG；分段 ASR 日志与空音频处理；转写失败与「无助手回复」错误提示更明确
- 助手多段回复持久化使用统一分隔符，与分段逻辑一致

后端 - Agent
- reply_limits：按 [SPLIT] 与段落拆段，并保证非空 fallback，供 WS 与 TTS 多段下发

后端 - 回忆录任务
- transcript ingest 记录 source_id；任务成功结?

api/tests/test_classification_fragment.py

@@ -1,4 +1,4 @@
-"""ClassificationAgent：零散档案启发式与分类 none 语义（纯函数/无 LLM）。"""
+"""ClassificationAgent：零散档案启发式与 none→summary 兜底（纯函数/无 LLM）。"""
 
 import pytest
 
@@ -28,9 +28,11 @@ def test_looks_like_fragment_only(text: str, expected_fragment: bool) -> None:
     assert _looks_like_fragment_only(text) is expected_fragment
 
 
-def test_classify_skips_story_for_birth_year_without_llm() -> None:
+def test_classify_maps_birth_year_fragment_to_summary_without_llm() -> None:
     agent = ClassificationAgent()
-    assert agent.classify("1999年出生", fallback_stage="childhood", llm=None) is None
+    assert (
+        agent.classify("1999年出生", fallback_stage="childhood", llm=None) == "summary"
+    )
 
 
 @pytest.mark.parametrize(

api/tests/test_memory_evidence.py

@@ -0,0 +1,30 @@
+"""Memory evidence 组装与检索契约（纯函数 / 无 DB）。"""
+
+from app.features.memory.evidence import (
+    EMPTY_EVIDENCE_BUNDLE,
+    _facts_to_dicts,
+    _stories_to_dicts,
+    _timeline_to_dicts,
+)
+from app.features.memory.schemas import EvidenceBundle
+
+
+def test_empty_evidence_bundle_keys() -> None:
+    assert set(EMPTY_EVIDENCE_BUNDLE.keys()) == {
+        "relevant_chunks",
+        "relevant_summaries",
+        "relevant_facts",
+        "timeline_hints",
+        "relevant_stories",
+    }
+
+
+def test_evidence_bundle_model_accepts_dict() -> None:
+    b = EvidenceBundle.model_validate(EMPTY_EVIDENCE_BUNDLE)
+    assert b.relevant_chunks == []
+
+
+def test_format_helpers_empty() -> None:
+    assert _facts_to_dicts([]) == []
+    assert _timeline_to_dicts([]) == []
+    assert _stories_to_dicts([]) == []

api/tests/test_reply_segments.py

@@ -0,0 +1,25 @@
+"""segments_from_llm_response：与客户端 split 规则对齐的单元校验。"""
+
+from app.agents.chat.reply_limits import (
+    nonempty_segments_or_fallback,
+    segments_from_llm_response,
+)
+
+
+def test_split_marker():
+    assert segments_from_llm_response("a[SPLIT]b", max_segments=3) == ["a", "b"]
+
+
+def test_paragraph_fallback_when_no_marker():
+    a = "太为你高兴了！在上海大剧院的舞台绽放，聚光灯下的你。"
+    b = "说到舞台，我忽然想起你黄浦江边的童年。从看着江水流淌，到在舞台上演绎别人的悲欢。"
+    assert segments_from_llm_response(f"{a}\n\n{b}", max_segments=3) == [a, b]
+
+
+def test_short_paragraphs_not_split():
+    t = "a\n\nb"
+    assert segments_from_llm_response(t, max_segments=3) == [t]
+
+
+def test_nonempty_fallback_when_all_blank():
+    assert nonempty_segments_or_fallback(["", "  "], fallback="ok") == ["ok"]

api/tests/test_whisper_local.py

@@ -0,0 +1,61 @@
+import asyncio
+import sys
+from types import SimpleNamespace
+
+import pytest
+
+from app.adapters.asr.whisper_local import (
+    WhisperASRProvider,
+    _looks_like_subtitle_hallucination,
+)
+
+
+def test_subtitle_watermark_detection() -> None:
+    assert _looks_like_subtitle_hallucination("字幕by索兰娅") is True
+    assert _looks_like_subtitle_hallucination("今天想聊聊童年往事") is False
+
+
+@pytest.mark.asyncio
+async def test_transcribe_retries_decode_audio_after_discarded_pass2(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    class DummyModel:
+        def __init__(self) -> None:
+            self.calls: list[object] = []
+
+        def transcribe(self, audio: object, **_: object):
+            self.calls.append(audio)
+            n = len(self.calls)
+            if n == 1:
+                return iter([]), SimpleNamespace()
+            if n == 2:
+                return iter([SimpleNamespace(text="字幕by索兰娅")]), SimpleNamespace()
+            if n == 3:
+                assert audio == "decoded-audio"
+                return (
+                    iter([SimpleNamespace(text="你好，今天想聊聊童年。")]),
+                    SimpleNamespace(),
+                )
+            raise AssertionError(f"unexpected transcribe call #{n}")
+
+    async def fake_to_thread(fn):
+        return fn()
+
+    def fake_decode_audio(_: str, sampling_rate: int = 16000):
+        assert sampling_rate == 16000
+        return "decoded-audio"
+
+    monkeypatch.setattr(asyncio, "to_thread", fake_to_thread)
+    monkeypatch.setitem(
+        sys.modules,
+        "faster_whisper",
+        SimpleNamespace(decode_audio=fake_decode_audio),
+    )
+
+    provider = WhisperASRProvider()
+    provider._model = DummyModel()
+
+    text = await provider.transcribe(b"fake-audio", format="m4a")
+
+    assert text == "你好，今天想聊聊童年。"
+    assert len(provider._model.calls) == 3