fix(local_embedding_store): retry batch computation on vector space changes

- Add up to 3 attempts to recompute embedding batch if vector space changes during processing
- Log warnings when maximum retries reached and discard stale results
- Prevent caching results from outdated vector spaces to maintain consistency
- Add tests to verify retry behavior and abort after continuous vector space churn

fix(daily_paper): update digest search logic and tests

- Change search to query existing memory, not only previous articles in daily_dir
- Allow multiple searches outside daily_dir but limit links to dated markdown in daily_dir before today
- Update test assertions to reflect revised search and linking rules

plugins/daily_paper/src/reme_daily_paper/digest.yaml

@@ -4,9 +4,10 @@ digest_user: |
   内容只能依据输入文档，不得补充文档中没有提供的事实。
   保留技术准确性，同时解释三篇论文为什么值得关注，以及它们之间有什么联系。
 
-  在写作前，先调用 `search` 检索以前的文章：围绕三篇论文的核心问题、方法、关键词和同义表达组织查询。
-  主题跨度较大时可以多次检索。只把 `{daily_dir}/` 下日期早于今天、
-  且与本期内容确实相似或互补的 Markdown 文章作为候选；必要时调用 `read` 核验全文，不要仅凭标题判断。
+  在写作前，先调用 `search` 检索已有记忆：围绕三篇论文的核心问题、方法、关键词和同义表达组织查询。
+  主题跨度较大时可以多次检索，搜索结果不必局限于 `{daily_dir}/`。只有 `{daily_dir}/` 下日期早于今天、
+  且与本期内容确实相似或互补的 Markdown 文章才可作为正文中的历史链接候选；必要时调用 `read` 核验全文，
+  不要仅凭标题判断。
   将确认相关的旧文章以 Wikilink 自然织入正文，并用句子说明关联（延续、对比、补充或方法相似）；
   链接必须采用带 `.md` 的完整 workspace-relative 路径，例如
   `[[{daily_dir}/2026-07-01/旧文章.md|此前的相关解读]]`。不要输出裸链接、独立关系字段，也不要虚构搜索未命中的路径。

plugins/daily_paper/tests/test_daily_paper.py

@@ -643,6 +643,7 @@ def test_digest_prompt_uses_configured_daily_directory(tmp_path: Path):
     )
 
     assert "`memory/`" in prompt
+    assert "搜索结果不必局限于 `memory/`" in prompt
     assert "[[memory/2026-07-01/旧文章.md" in prompt
     assert "[[daily/2026-07-01/" not in prompt
 
@@ -932,7 +933,8 @@ async def test_pipeline_filters_strict_yesterday_and_writes_outputs(
     assert "调用 Read" not in digest_prompt
     assert "daily/2026-07-21" not in digest_prompt
     assert "长期记忆" not in digest_prompt
-    assert "先调用 `search` 检索以前的文章" in digest_prompt
+    assert "先调用 `search` 检索已有记忆" in digest_prompt
+    assert "搜索结果不必局限于 `daily/`" in digest_prompt
     assert "end_date" not in digest_prompt
     assert "limit=" not in digest_prompt
     assert "Wikilink" in digest_prompt

reme/components/embedding_store/local_embedding_store.py

@@ -12,6 +12,7 @@ from ..component_registry import R
 from ..as_embedding import BaseAsEmbedding
 
 Miss = tuple[int, str, str]  # (result_index, text, cache_key)
+_MAX_VECTOR_SPACE_ATTEMPTS = 3
 
 
 @R.register("local")
@@ -125,14 +126,30 @@ class LocalEmbeddingStore(BaseEmbeddingStore):
         return results, misses
 
     async def _fill_misses(self, misses: list[Miss], results: list[np.ndarray | None], **kwargs) -> None:
-        vector_space_id = self._cache_space
         size = self.max_batch_size
         for start in range(0, len(misses), size):
             batch = misses[start : start + size]
-            for idx, key, emb in await self._compute_batch(batch, **kwargs):
-                results[idx] = emb
-                if vector_space_id == self.vector_space_id == self._cache_space:
+            for attempt in range(1, _MAX_VECTOR_SPACE_ATTEMPTS + 1):
+                await self._sync_cache_space()
+                vector_space_id = self._cache_space
+                computed = await self._compute_batch(batch, **kwargs)
+                if vector_space_id != self.vector_space_id or vector_space_id != self._cache_space:
+                    if attempt == _MAX_VECTOR_SPACE_ATTEMPTS:
+                        self.logger.warning(
+                            f"Embedding vector space kept changing while computing a batch; "
+                            f"discarding {len(computed)} stale result(s) after {attempt} attempts",
+                        )
+                    else:
+                        self.logger.info(
+                            f"Embedding vector space changed while computing a batch; "
+                            f"discarding {len(computed)} stale result(s) and retrying "
+                            f"({attempt}/{_MAX_VECTOR_SPACE_ATTEMPTS})",
+                        )
+                    continue
+                for idx, key, emb in computed:
+                    results[idx] = emb
                     self._cache_put(key, emb)
+                break
 
     async def _compute_batch(self, batch: list[Miss], **kwargs) -> list[tuple[int, str, np.ndarray]]:
         texts = [text for _, text, _ in batch]

tests/unit/test_local_embedding_store.py

@@ -509,24 +509,58 @@ def test_cache_space_is_rechecked_after_async_load(monkeypatch, tmp_path):
     run(go())
 
 
-def test_completed_request_only_writes_to_its_active_cache_space():
-    """A v3 request must not populate v4 after the provider switches back to v3."""
+def test_completed_request_retries_after_vector_space_changes():
+    """A request completed by the old provider must not escape into the new vector space."""
 
     async def go():
         embedding = OpenAIAsEmbedding(name="t_space_write_race", backend="openai", model="v3", dimensions=2)
         store = LocalEmbeddingStore(name="t_local_write_race")
         store.as_embedding = embedding
         store._cache_space = embedding.vector_space_id
+        calls = 0
 
         async def compute_after_round_trip(_batch, **_kwargs):
-            embedding.model = FakeProviderModel("v4")
-            store._cache_space = embedding.vector_space_id
-            embedding.model = FakeProviderModel("v3")
-            return [(0, "key", np.array([3.0, 0.0], dtype=np.float16))]
+            nonlocal calls
+            calls += 1
+            if calls == 1:
+                embedding.model = FakeProviderModel("v4")
+                return [(0, "key", np.array([3.0, 0.0], dtype=np.float16))]
+            return [(0, "key", np.array([4.0, 0.0], dtype=np.float16))]
 
         store._compute_batch = compute_after_round_trip
-        await store._fill_misses([(0, "text", "key")], [None])
+        results = [None]
+        await store._fill_misses([(0, "text", "key")], results)
 
+        assert calls == 2
+        assert store._cache_space == embedding.vector_space_id
+        np.testing.assert_array_equal(results[0], np.array([4.0, 0.0], dtype=np.float16))
+        np.testing.assert_array_equal(store._cache["key"], np.array([4.0, 0.0], dtype=np.float16))
+
+    run(go())
+
+
+def test_completed_request_stops_retrying_when_vector_space_keeps_changing():
+    """Continuous configuration churn must leave the batch empty instead of blocking forever."""
+
+    async def go():
+        embedding = OpenAIAsEmbedding(name="t_space_write_churn", backend="openai", model="v3", dimensions=2)
+        store = LocalEmbeddingStore(name="t_local_write_churn")
+        store.as_embedding = embedding
+        store._cache_space = embedding.vector_space_id
+        calls = 0
+
+        async def change_space_every_time(_batch, **_kwargs):
+            nonlocal calls
+            calls += 1
+            embedding.model = FakeProviderModel(f"v{calls + 3}")
+            return [(0, "key", np.array([float(calls), 0.0], dtype=np.float16))]
+
+        store._compute_batch = change_space_every_time
+        results = [None]
+        await store._fill_misses([(0, "text", "key")], results)
+
+        assert calls == 3
+        assert results == [None]
         assert "key" not in store._cache
 
     run(go())