feat(api): 回忆录管线简化、路由延迟池与相关加固

- Phase1/2：移除 MemoirOrchestrator.run 与 process_memoir_segments 别名；文档改为 process_memoir_phase1。
- 槽位校验集中到 stage_constants（filter_stage_slots），批处理与顺序路径及 state_service 写库一致。
- StoryRoute：no_llm/parse_error/invalid_target 保守 new_story；短篇护栏不覆盖这些 fallback。
- Phase2 低置信单路径可选延迟（StoryPipelineResult.deferred）：不写 Chapter/Story，Segment 记录 defer 元数据，冷却内不重复消费；上限后停自动重试，Phase1 同类目新段唤醒池内段。
- Alembic 0017：segments 表 narrative_defer_* 列。
- ProfileAgent：经 LlmGateway/注入 Provider 统一聊天与 JSON，新增测试。
- ImagePromptOrchestrator：LLM 初始化失败可依配置降级或硬失败；补充策略测试。
- 配套单测与 README/本地开发文档表述更新。

Co-authored-by: Cursor <cursoragent@cursor.com>

api/tests/test_memoir_pipeline_optimization.py

@@ -106,7 +106,10 @@ def test_orchestrator_fallback_to_sequential(monkeypatch: pytest.MonkeyPatch) ->
 
     orch._prepare_batches_via_batch_llm = fail_batch
     orch.extraction_agent.extract = MagicMock(
-        return_value=ExtractionResult(detected_stage="childhood", slots={"toy": "ball"})
+        return_value=ExtractionResult(
+            detected_stage="childhood",
+            slots={"place": "潍坊"},
+        )
     )
     orch.classification_agent.classify = MagicMock(
         return_value=ChapterClassifyResult(category="childhood", llm_said_none=False)
@@ -134,6 +137,52 @@ def test_orchestrator_fallback_to_sequential(monkeypatch: pytest.MonkeyPatch) ->
     assert "s1" in result.segment_chapter_category
 
 
+def test_orchestrator_sequential_filters_invalid_slots(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Sequential fallback should match batch path slot validation."""
+    monkeypatch.setattr(
+        "app.agents.memoir.orchestrator.settings.memoir_phase1_batch_llm_enabled",
+        False,
+    )
+
+    orch = MemoirOrchestrator()
+    orch.extraction_agent.extract = MagicMock(
+        return_value=ExtractionResult(
+            detected_stage="childhood",
+            slots={"place": "潍坊", "hallucinated": "bad"},
+        )
+    )
+    orch.classification_agent.classify = MagicMock(
+        return_value=ChapterClassifyResult(category="childhood", llm_said_none=False)
+    )
+
+    st = MemoirStateSchema(
+        stage_order=["childhood"],
+        current_stage="childhood",
+        covered_stages=[],
+        slots={},
+    )
+    calls: list[tuple] = []
+
+    class _Seg:
+        id = "s1"
+        user_input_text = "我小时候在潍坊。"
+
+    def update_slot(*args):
+        calls.append(args)
+        return st
+
+    orch.prepare_batches(
+        segments=[_Seg()],
+        llm=MagicMock(),
+        get_or_create_state=lambda: st,
+        update_slot=update_slot,
+    )
+
+    assert calls == [("childhood", "place", "潍坊", ["s1"])]
+
+
 # ---------------------------------------------------------------------------
 # Memory enrichment decoupled from ingest
 # ---------------------------------------------------------------------------
@@ -216,6 +265,33 @@ def test_resolve_append_target_forced_new_on_overflow() -> None:
     assert dsrc == "forced_new_due_to_append_limit"
 
 
+def test_resolve_append_target_does_not_guardrail_route_fallback() -> None:
+    """No-LLM / parse fallback new_story decisions must not append by recency."""
+    from app.features.memoir.story_pipeline_sync import _resolve_append_target
+
+    session = MagicMock()
+    candidate = MagicMock()
+    candidate.id = "story-1"
+
+    tid, existing, dsrc = _resolve_append_target(
+        session,
+        route_decision="new_story",
+        route_target_story_id=None,
+        user_id="u1",
+        chapter_category="childhood",
+        oral_norm="short text",
+        candidate_stories=[candidate],
+        story_meta={"story-1": {"char_count": 10, "version_count": 1}},
+        decision_source="no_llm",
+        memoir_correlation_id=None,
+    )
+
+    assert tid is None
+    assert existing == ""
+    assert dsrc == "no_llm"
+    session.get.assert_not_called()
+
+
 # ---------------------------------------------------------------------------
 # _run_post_pipeline_commit helper
 # ---------------------------------------------------------------------------