memleaf 0.2.93__tar.gz → 0.2.95__tar.gz

{memleaf-0.2.93 → memleaf-0.2.95}/CHANGELOG.md

@@ -5,6 +5,49 @@ All notable changes to memleaf are documented here.
 ## Unreleased
 
 
+## 0.2.95 — 2026-10-05
+
+### Preserve turn identity around internal Hermes continuations
+
+- Exclude ephemeral transcript rows using the installed Hermes persistence
+  contract before checking sources and resolving capture identity. Internal
+  continuation nudges with role=user no longer hide the actual user admission.
+  Keep exact real-user boundaries, unsupported-host fallback, source checks and
+  duplicate-delivery handling; never identify scaffolding by business wording.
+- Accept valid redundant responsibility annotations for unchanged existing
+  assignees in UPDATE. Preserve the original responsibility basis; unknown
+  assignees, conflicting clears, invalid citations and unauthorized changes
+  remain rejected. Bind the adjusted instructions to the protocol digest.
+- Add no model stage, request or retry. Source callback and saved-candidate
+  replays used no external model calls. All 58 affected checks passed in the
+  native Hermes runtime. Three live Hermes acceptance turns updated, completed
+  and reopened the original two memories correctly: six Memleaf requests,
+  24,758 observed tokens, and no unresolved post-fix retention intents.
+
+
+## 0.2.94 — 2026-10-05
+
+### Reopen existing memories without redundant model fields
+
+- Treat an evidence-backed UPDATE with status=active as the reopen intent for a
+  closed memory. Keep legacy reopen=true compatible and reject known older or
+  same-observation updates; missing source ordering stays unknown. Apply the
+  same ordering rule to explicitly restoring retracted facts.
+- Preserve an independently verified user's source time when the final assistant
+  display cannot be matched. Keep assistant metadata unknown and distinguish
+  verified user deliveries while keeping repeated callbacks idempotent. Never
+  infer timestamps, borrow an older matching user message, or weaken target,
+  permission, conflict or semantic-review checks.
+- Remove the redundant reopen flag requirement from extraction instructions.
+  Add no model request, retry or stage; the existing request budget is unchanged.
+- Validation: 267 applicable checks and four isolated frozen-fault replays passed
+  against the current contract. Eight pre-existing obsolete test expectations
+  were recorded separately. An isolated live-model acceptance reopened one
+  existing memory with the correct deadline and no source time or reopen flag:
+  two successful requests, 3,089 observed tokens. Installed Hermes acceptance
+  remains separate from source and package verification.
+
+
 ## 0.2.93 — 2026-10-05
 
 ### Keep useful body compaction when another candidate has no net savings

{memleaf-0.2.93/src/memleaf.egg-info → memleaf-0.2.95}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: memleaf
-Version: 0.2.93
+Version: 0.2.95
 Summary: A local-first Markdown memory core for AI agents
 Author: memleaf contributors
 License-Expression: MIT
@@ -23,7 +23,7 @@ Dynamic: license-file
 
 [English](README.en.md) · [PyPI](https://pypi.org/project/memleaf/) · [GitHub](https://github.com/miffyblueboo/memleaf)
 
-> **版本：0.2.93。**
+> **版本：0.2.95。**
 > 自动提炼和复核共用同一条“未来记忆价值”标准：模型综合未来复用、信息增量、再次读取时的直接可用性和忘记成本，只保留对未来理解、判断或行动有实质影响的最小核心；没有明确价值的信息不提炼，不按具体业务场景硬编码排除。
 > 提炼先比较旧事项中的编号、目标、进展、责任和期限；补全未知信息也应维护原事项，另建独立待办不能替代项目事实更新。确认重复的记忆在现有处理请求中合并，保留历史与来源；项目和独立任务保持各自生命周期，Core 不按标题强行合并。
 > 新请求兼容等价的字段表示；未知执行人保持 `null`，新增或变更负责人和期限需要真实来源；等待条件在正文维护，不再提炼 `waiting_on`。行动及相关字段经过有界语义复核，提炼与复核共用每份工作的最多五次模型请求额度，成功即停止；复核失败不写入。来源时间用于解析相对日期，不补造事实日期。

{memleaf-0.2.93 → memleaf-0.2.95}/README.en.md

@@ -4,7 +4,7 @@
 
 [中文](README.md) · [PyPI](https://pypi.org/project/memleaf/) · [GitHub](https://github.com/miffyblueboo/memleaf)
 
-> **Version: 0.2.93.**
+> **Version: 0.2.95.**
 > Automatic extraction and review share one future-value standard: the model weighs likely reuse, information gain, direct usability when read again, and the cost of forgetting. It keeps only the smallest core that can materially help future understanding, decisions, or actions; information without clear value is not extracted, and no business-specific exclusion rule is hard-coded.
 > Extraction compares existing identifiers, objectives, progress, responsibility and deadlines first. Filling an unknown also maintains the original item; a separate task does not replace updating shared project facts. Confirmed duplicates can be merged in the existing processing requests with their sources and history preserved. Independent project and task lifecycles remain separate; Core does not merge matching titles.
 > New requests accept equivalent field representations. Unknown executors remain `null`; new or changed responsibility and deadlines require actual source evidence. Dependencies belong in the body; `waiting_on` is no longer extracted. Proposed actions and related fields undergo bounded semantic verification within a shared five-request limit per work item, stopping on success; failed verification blocks writes. Source time resolves relative dates without inventing factual dates.

{memleaf-0.2.93 → memleaf-0.2.95}/README.md

@@ -4,7 +4,7 @@
 
 [English](README.en.md) · [PyPI](https://pypi.org/project/memleaf/) · [GitHub](https://github.com/miffyblueboo/memleaf)
 
-> **版本：0.2.93。**
+> **版本：0.2.95。**
 > 自动提炼和复核共用同一条“未来记忆价值”标准：模型综合未来复用、信息增量、再次读取时的直接可用性和忘记成本，只保留对未来理解、判断或行动有实质影响的最小核心；没有明确价值的信息不提炼，不按具体业务场景硬编码排除。
 > 提炼先比较旧事项中的编号、目标、进展、责任和期限；补全未知信息也应维护原事项，另建独立待办不能替代项目事实更新。确认重复的记忆在现有处理请求中合并，保留历史与来源；项目和独立任务保持各自生命周期，Core 不按标题强行合并。
 > 新请求兼容等价的字段表示；未知执行人保持 `null`，新增或变更负责人和期限需要真实来源；等待条件在正文维护，不再提炼 `waiting_on`。行动及相关字段经过有界语义复核，提炼与复核共用每份工作的最多五次模型请求额度，成功即停止；复核失败不写入。来源时间用于解析相对日期，不补造事实日期。

{memleaf-0.2.93 → memleaf-0.2.95}/docs/hermes-source-comparison-context.md

@@ -8,6 +8,11 @@ an event does not create a source time.
 
 The legacy callback checks the final assistant and the nearest user. It has no
 32-message lookback limit. It does not borrow a matching user from an older turn.
+Before checking that boundary, it excludes ephemeral rows using the installed
+Hermes persistence predicate. Internal continuation nudges are not user events,
+even when their role is `user`. It never identifies them by wording or arbitrary
+private flags. If the host predicate is unavailable, the strict legacy boundary
+is retained. Source checks and turn association use the same filtered snapshot.
 
 Some Hermes versions persist a reply before adding their file-mutation verifier
 footer. The Provider accepts that display transformation only when the installed

{memleaf-0.2.93 → memleaf-0.2.95}/docs/incremental-planner-preview.md

@@ -150,9 +150,11 @@ scanned and auto-filled as deadlines. This does **not** prove that a model-selec
 ISO date was semantically a deadline: that remains a live-model acceptance item.
 
 `field_basis` is self-contained source metadata, never eN or a list offset.
-Provably older conflicting field changes are rejected. Explicit reopen/restore
-requires a later comparable basis; legacy heads lacking it produce a visible
-unverified-time issue rather than a fabricated business time. Unknown completion
+Provably older conflicting field changes are rejected. Reopening uses UPDATE
+with `patch.status=active`; the legacy `reopen:true` flag remains optional.
+Reopen/restore reject a known older or identical observation. Missing optional
+source time/order remains unknown and does not block the proposed transition;
+the normal source review checks whether the change is supported. Unknown completion
 time is not replaced with observation time. Future effective state changes are
 blocked, while a fact describing a plan can still be proposed. No timer executes
 future plans. This is bounded observation-order checking, not a bitemporal database.

{memleaf-0.2.93 → memleaf-0.2.95}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "memleaf"
-version = "0.2.93"
+version = "0.2.95"
 description = "A local-first Markdown memory core for AI agents"
 readme = "README.md"
 requires-python = ">=3.11"

{memleaf-0.2.93 → memleaf-0.2.95}/src/memleaf/__init__.py

@@ -1,6 +1,6 @@
 """Local-first Markdown memory core for AI agents."""
 
-__version__ = "0.2.93"
+__version__ = "0.2.95"
 
 from .config import DEFAULT_CONFIG, default_config, load_config, save_config
 from .frontmatter import FrontmatterError, dump_frontmatter, dump_yaml, load_yaml, parse_frontmatter

{memleaf-0.2.93 → memleaf-0.2.95}/src/memleaf/hermes_provider/_provider.py

@@ -29,6 +29,23 @@ def _capture_assistant_text(value: str) -> str:
                   flags=re.IGNORECASE)
 
 
+def _source_messages(messages: Any) -> Any:
+    """Use the host's durable transcript boundary, not its internal nudges.
+
+    Hermes may retain synthetic user/assistant rows while continuing after an
+    empty response. Its persistence contract excludes those rows. Follow that
+    exact host predicate rather than recognizing wording or scanning past a
+    real user. Hosts without this predicate retain the strict legacy boundary.
+    """
+    if not isinstance(messages, list):
+        return messages
+    try:
+        from agent.session_persistence import _is_ephemeral_scaffolding
+        return [row for row in messages if not _is_ephemeral_scaffolding(row)]
+    except (ImportError, AttributeError, TypeError, ValueError):
+        return messages
+
+
 def _assistant_source_matches(messages: Any, content: str) -> bool:
     """Accept exact text or an exactly reconstructed native verifier footer.
 
@@ -135,6 +152,18 @@ def _sync_delivery_digest(messages: Any, users: Any, user_content: str, assistan
                 identities = []
                 break
             identities.append((uid, timestamp))
+    elif (isinstance(messages, list) and messages
+          and isinstance(messages[-1], Mapping) and messages[-1].get("role") == "assistant"
+          and not messages[-1].get("tool_calls")):
+        user = next((row for row in reversed(messages[:-1])
+                     if isinstance(row, Mapping) and row.get("role") == "user"), None)
+        if user is not None and user.get("content") == user_content:
+            uid = user.get("message_uid") or user.get("message_id")
+            timestamp = _independent_source_time(user, "user").get("user", {}).get("source_time")
+            if uid or timestamp:
+                # Only the verified user identifies this delivery; do not
+                # bind the host's transformed reply to an assistant source.
+                identities.append((uid, timestamp))
     payload = f"{user_content}\x00{_capture_assistant_text(assistant_content)}"
     if identities:
         payload += "\x00" + json.dumps(identities, ensure_ascii=False, separators=(",", ":"))
@@ -151,13 +180,17 @@ def _sync_source_metadata(messages: Any, user_content: str, assistant_content: s
     from datetime import datetime
     from math import isfinite
 
+    messages = _source_messages(messages)
     if not isinstance(messages, list) or not messages:
         logger.info("memleaf source-metadata reason=messages_missing")
         return {}
     assistant = messages[-1]
-    if not _assistant_source_matches(messages, assistant_content):
+    assistant_matches = _assistant_source_matches(messages, assistant_content)
+    if not assistant_matches:
         logger.info("memleaf source-metadata reason=assistant_mismatch")
-        return {}
+        if (not isinstance(assistant, Mapping) or assistant.get("role") != "assistant"
+                or assistant.get("tool_calls")):
+            return {}
     # Tool rounds do not define a turn boundary. Stop at the nearest user,
     # even if its text mismatches; never borrow an older matching message.
     preceding = reversed(messages)
@@ -166,10 +199,15 @@ def _sync_source_metadata(messages: Any, user_content: str, assistant_content: s
                  if isinstance(item, Mapping) and item.get("role") == "user"), None)
     if user is None:
         logger.info("memleaf source-metadata reason=user_missing")
-        return _independent_assistant_time(assistant, assistant_content)
+        return _independent_assistant_time(assistant, assistant_content) if assistant_matches else {}
     if user.get("content") != user_content:
         logger.info("memleaf source-metadata reason=user_mismatch")
-        return _independent_assistant_time(assistant, assistant_content)
+        return _independent_assistant_time(assistant, assistant_content) if assistant_matches else {}
+    if not assistant_matches:
+        # Host warnings/display transformations cannot invalidate an exact
+        # current user source. Keep its time independently, without binding
+        # the unverified assistant or inventing paired identities/order.
+        return _independent_source_time(user, "user")
     result: dict[str, dict[str, Any]] = {}
     for role, item in (("user", user), ("assistant", assistant)):
         row: dict[str, Any] = {}
@@ -213,11 +251,15 @@ def _sync_source_metadata(messages: Any, user_content: str, assistant_content: s
 
 
 def _independent_assistant_time(assistant: Mapping[str, Any], content: str) -> dict[str, dict[str, Any]]:
+    return _independent_source_time(assistant, "assistant")
+
+
+def _independent_source_time(message: Mapping[str, Any], role: str) -> dict[str, dict[str, Any]]:
     from datetime import datetime
     from math import isfinite
-    value = assistant.get("source_time")
+    value = message.get("source_time")
     if value is None:
-        value = assistant.get("timestamp")
+        value = message.get("timestamp")
     parsed = None
     try:
         if type(value) in (int, float) and isfinite(value) and 0 <= value < 253402300800:
@@ -225,7 +267,7 @@ def _independent_assistant_time(assistant: Mapping[str, Any], content: str) -> d
         elif isinstance(value, str) and len(value) <= 80:
             parsed = datetime.fromisoformat(value.replace("Z", "+00:00"))
         if parsed is not None and parsed.tzinfo is not None and parsed.utcoffset() is not None:
-            return {"assistant": {"source_time": parsed.isoformat()}}
+            return {role: {"source_time": parsed.isoformat()}}
     except (ValueError, OSError, OverflowError):
         pass
     return {}
@@ -1152,6 +1194,7 @@ class MemleafMemoryProvider(MemoryProvider):
         user_content: str,
         assistant_content: str,
         delivery_digest: Optional[str] = None,
+        source_fallback: bool = False,
     ) -> str:
         pair_digest = delivery_digest or sha256(f"{user_content}\x00{assistant_content}".encode("utf-8")).hexdigest()[:16]
         pair_key = (session_id, pair_digest)
@@ -1160,6 +1203,8 @@ class MemleafMemoryProvider(MemoryProvider):
             self._turn_ids_by_pair.move_to_end(pair_key)
             return existing
         resolved = _turn_id(turn_number, user_content, assistant_content)
+        if source_fallback and delivery_digest and resolved.startswith("turn-fallback-"):
+            resolved = "turn-fallback-" + delivery_digest
         if existing is not None and existing == resolved:
             self._turn_ids_by_pair.move_to_end(pair_key)
             return existing
@@ -1912,7 +1957,7 @@ class MemleafMemoryProvider(MemoryProvider):
         # Snapshot the list at invocation; subsequent tool/turn appends cannot
         # change which source rows this callback verifies.
         import copy
-        messages = copy.deepcopy(messages)
+        messages = _source_messages(copy.deepcopy(messages))
         # Hermes serializes provider sync work, but this lock also protects
         # direct/plugin-level concurrent calls and makes process one-shot per
         # captured turn within this provider instance.
@@ -1965,6 +2010,7 @@ class MemleafMemoryProvider(MemoryProvider):
                     user_content,
                     captured_assistant_content,
                     delivery_digest=delivery_digest,
+                    source_fallback=not _assistant_source_matches(messages, raw_assistant_content),
                 )
                 source_metadata = ({} if host_events is not None or correction_users else
                                    _sync_source_metadata(messages, user_content, raw_assistant_content))

{memleaf-0.2.93 → memleaf-0.2.95}/src/memleaf/hermes_provider/plugin.yaml

@@ -1,5 +1,5 @@
 name: memleaf
-version: 0.2.93
+version: 0.2.95
 description: "Hermes-native external MemoryProvider for local-first Markdown memory shared with the memleaf MCP server."
 hooks:
   - prefetch

{memleaf-0.2.93 → memleaf-0.2.95}/src/memleaf/incremental_prompts.py

@@ -27,11 +27,11 @@ NO_MEMORY：仅 action；自动模式整轮无值得维护/新增的内容时独
 字段：
 title 标识事项，不放交期/发生日/进度，改存正文与结构字段；旧标题交期一并清理。名称固有年度等身份日期保留。
 type=fact|todo|preference|project|event|identity|other。有独立完成条件的持续行动设 actionable:true（todo 隐含）；事实类别与行动属性可分开表达。
-status=active|completed|cancelled；完成/取消/交接同时纠正冲突的 title/body；独立未完行动不能被兄弟项完成吞掉。重新开启 UPDATE 使用 reopen:true 和 patch.status=active，需要较新明确依据。
+status=active|completed|cancelled；完成/取消/交接同时纠正冲突的 title/body；独立未完行动不能被兄弟项完成吞掉。重开原 target：UPDATE patch.status=active，须有 new 依据。
 assignee 是实际执行人，只有明确由用户本人执行才 user，其他执行人用明确名称，未知为 null。用户记录/转发/协调不证明本人执行。等待依赖、卡点仅在正文维护，按最新事实替换过时内容，不生成等待字段。
 scope 采用输入引用；通用为 global，归属未知为 unscoped；只有 allow_new_scopes 才可提出 project:新名称。不把 scopes 当搜索提示，不扩大 write_scopes。
 patch 只含 title/body/scope/status/actionable/assignee/deadline/validity；type 沿用目标。
-新增/变更非空 assignee 必须提供 responsibility_basis:{"assignee":{"ref":"e1","text":"确立执行人的原文"}}；放行级或对应 memory/patch 内均可，不能冲突。ref 必须在 evidence，text 是实际原文；新增用户执行责任必须引用 user 原文，assistant 不能赋予用户新义务。原文必须确立这个实际行动的责任，引用另一件事的请求不算。沿用不需重复引用。
+新增/变更非空 assignee 必须提供 responsibility_basis:{"assignee":{"ref":"e1","text":"确立执行人的原文"}}；放行级或对应 memory/patch 内均可，不能冲突。ref 必须在 evidence，text 是实际原文；新增用户执行责任必须引用 user 原文，assistant 不能赋予用户新义务。原文必须确立这个实际行动的责任，引用另一件事的请求不算。沿用无需重复引用；冗余引用仍须有效，但不刷新旧依据。
 
 日期：有明确期限时放 memory.deadline 或 patch.deadline:{"ref":"e1","text":"期限原文"}，不能只留正文。没有期限变更时省略 deadline。负责人交接、等待验收、尚未完成不能取消旧期限。取消期限用 {"ref":"e1","clear":true,"text":"取消期限原文"}；new 未取消时保留原期限。source_time 只用于解释相对日，不补造发生日；每个相对日按其自己的消息时间解析。无来源时间保持未知，不借处理时间。正文日期与期限、材料日期、发生日期区分；无日期也能记完成事实。CREATE/UPDATE 可给 effective:{"ref":"e1","text":"生效时间原文"}。
 

{memleaf-0.2.93 → memleaf-0.2.95}/src/memleaf/incremental_protocol.py

@@ -559,7 +559,12 @@ def _review_fields(row: Mapping[str, Any], fields: Mapping[str, Any], state: Map
     if needed - set(offered):
         raise ValueError("missing_responsibility_basis")
     for key, selection in offered.items():
-        if key not in fields or fields[key] is None:
+        # An unchanged owner may be omitted from UPDATE while its annotation
+        # is repeated. Validate that annotation without changing the owner or
+        # advancing its field basis. CREATE and explicit clears cannot borrow
+        # an old owner to satisfy a supplied annotation.
+        value = fields.get(key) if key in fields else old.get(key)
+        if value is None:
             raise ValueError("invalid_responsibility_basis")
         if not isinstance(selection, dict) or set(selection) != {"ref", "text"}:
             raise ValueError("invalid_responsibility_basis")
@@ -568,9 +573,10 @@ def _review_fields(row: Mapping[str, Any], fields: Mapping[str, Any], state: Map
         if (ref not in refs or not reading_text(text).strip()
                 or reading_text(text) not in reading_text(evidence[ref]["text"])):
             raise ValueError("unproven_responsibility_basis")
-        if key == "assignee" and fields[key] == "user" and evidence[ref]["role"] != "user":
+        if key == "assignee" and value == "user" and evidence[ref]["role"] != "user":
             raise ValueError("unconfirmed_user_responsibility")
-        bases[key] = source_basis(evidence[ref])
+        if key in fields:
+            bases[key] = source_basis(evidence[ref])
     return bases
 
 
@@ -717,6 +723,8 @@ def _parse_row(row: Any, state: Mapping[str, Any]) -> dict[str, Any]:
 def _order(new: Mapping[str, Any], old: Mapping[str, Any]) -> int | None:
     # These fields prove observation order, not inferred business effectiveness.
     same_stream = all(new.get(k) is not None and new.get(k) == old.get(k) for k in ("source", "session_id"))
+    if same_stream and new.get("event_key") and new.get("event_key") == old.get("event_key"):
+        return 0  # The same observation cannot revive its own closed state.
     if same_stream and type(new.get("source_sequence")) is int and type(old.get("source_sequence")) is int:
         a, b = new["source_sequence"], old["source_sequence"]
         return (a > b) - (a < b)
@@ -768,15 +776,16 @@ def _compile_group(rows: list[dict[str, Any]], state: Mapping[str, Any]) -> dict
     if boundary is not None and (set(chosen) - set(boundary) or (target and set(old["scopes"]) - set(boundary))):
         raise ValueError("blocked_scope")
     if old.get("status") in {"completed", "cancelled"} and fields.get("status") == "active":
-        if not any(row.get("reopen") for row in rows):
-            raise ValueError("explicit_reopen_required")
-        if _order(bases.get("status", {}), old.get("field_basis", {}).get("status", {})) != 1:
-            raise ValueError("unverified_reopen_time")
+        # The explicit patch already expresses reopening. A model flag cannot
+        # prove source support; the existing semantic review checks the change.
+        # Missing optional time/order stays unknown, not an automatic veto.
+        if _order(bases.get("status", {}), old.get("field_basis", {}).get("status", {})) in (-1, 0):
+            raise ValueError("stale_observation")
     if old.get("validity") == "retracted":
         if fields.get("validity") != "valid" or not fields.get("body"):
             raise ValueError("explicit_restore_required")
-        if _order(bases.get("validity", {}), old.get("field_basis", {}).get("validity", {})) != 1:
-            raise ValueError("unverified_restore_time")
+        if _order(bases.get("validity", {}), old.get("field_basis", {}).get("validity", {})) in (-1, 0):
+            raise ValueError("stale_observation")
     if fields.get("validity") == "retracted":
         # Retraction is an explicit lifecycle operation. An explanatory body
         # belongs to the audit metadata, never to the active assertion. Retain

{memleaf-0.2.93 → memleaf-0.2.95/src/memleaf.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: memleaf
-Version: 0.2.93
+Version: 0.2.95
 Summary: A local-first Markdown memory core for AI agents
 Author: memleaf contributors
 License-Expression: MIT
@@ -23,7 +23,7 @@ Dynamic: license-file
 
 [English](README.en.md) · [PyPI](https://pypi.org/project/memleaf/) · [GitHub](https://github.com/miffyblueboo/memleaf)
 
-> **版本：0.2.93。**
+> **版本：0.2.95。**
 > 自动提炼和复核共用同一条“未来记忆价值”标准：模型综合未来复用、信息增量、再次读取时的直接可用性和忘记成本，只保留对未来理解、判断或行动有实质影响的最小核心；没有明确价值的信息不提炼，不按具体业务场景硬编码排除。
 > 提炼先比较旧事项中的编号、目标、进展、责任和期限；补全未知信息也应维护原事项，另建独立待办不能替代项目事实更新。确认重复的记忆在现有处理请求中合并，保留历史与来源；项目和独立任务保持各自生命周期，Core 不按标题强行合并。
 > 新请求兼容等价的字段表示；未知执行人保持 `null`，新增或变更负责人和期限需要真实来源；等待条件在正文维护，不再提炼 `waiting_on`。行动及相关字段经过有界语义复核，提炼与复核共用每份工作的最多五次模型请求额度，成功即停止；复核失败不写入。来源时间用于解析相对日期，不补造事实日期。