fix(LongTermMemory): preserve temporal precision and complete evidence spans

plugins/KnowledgeEngine/LongTermMemory/README.md

@@ -620,7 +620,7 @@ Star the repo if you find it useful!
 | `source_ref` | Declared reference to a source message or document; not proof of Host verification. |
 | `claim_type` | `unknown` (default), `user_statement`, or `inference`. |
 
-Time fields require ISO-8601 values with an explicit timezone and are normalized to UTC with second precision. When both validity bounds are present, `valid_from` must precede `valid_until`. Leave unknown values empty; the plugin does not turn “tomorrow” into an invented date or assume a timezone for these fields. Imports reject malformed temporal/provenance metadata before writing the batch. Old records missing these fields remain readable with unknown evidence and no invented event time.
+Time fields require ISO-8601 values with an explicit timezone and are normalized to UTC while preserving subsecond precision. When both validity bounds are present, `valid_from` must precede `valid_until`. Leave unknown values empty; the plugin does not turn “tomorrow” into an invented date or assume a timezone for these fields. Imports reject malformed temporal/provenance metadata before writing the batch. Old records missing these fields remain readable with unknown evidence and no invented event time.
 
 `remember` accepts these optional fields. A caller can declare `user_statement` only for a direct user statement; this declaration is not a verification or authorization mechanism. Model deductions belong under `inference`. JSON document ingestion supplies a `document:<id>` reference when no explicit source reference was supplied. Exported references remain declarations when imported elsewhere.
 
@@ -630,7 +630,7 @@ Automatic candidate extraction uses actual `user_message_text` only. An assistan
 
 Set `candidate_model_uuid` on the memory knowledge base to select a model for candidate extraction. Leave it empty to keep the existing rule-based path and avoid an extra model call. Extraction still requires `candidate_extraction_enabled=true`.
 
-The optional path selects and classifies complete original sentences or the full current user text; it does not rewrite statements into new facts. `evidence_quote` must match an original span exactly, and the quote itself becomes the episodic content or profile value. This preserves negations and qualifiers rather than accepting a generated fact that could contradict its evidence. Only the current actual user text and the Host's UTC processing timestamp are sent; assistant replies, older conversations, and model-invented source IDs are not used. No automatic relative-date conversion or cross-turn reference resolution is performed.
+The optional path selects and classifies complete original sentences or the full current user text; it does not rewrite statements into new facts. `evidence_quote` must match an original span exactly, and the quote itself becomes the episodic content or profile value. This preserves negations and qualifiers rather than accepting a generated fact that could contradict its evidence. Only the current actual user text and the plugin Runtime's UTC processing timestamp are sent; assistant replies, older conversations, and model-invented source IDs are not used. No automatic relative-date conversion or cross-turn reference resolution is performed.
 
 Model-selected candidates are labeled `inference` because selection/classification remains a model judgment. They always remain pending for manual review, even when `candidate_auto_apply=true`; the original quote and the declared query reference stay inspectable in the candidate record. This is evidence selection, not proof that a statement is true or that a quotation is safe to treat as a current preference.
 

plugins/KnowledgeEngine/LongTermMemory/store/candidate_extraction.py

@@ -14,7 +14,7 @@ async def extract_candidates(
         return []
     spans = {
         part.strip()
-        for part in re.split(r"(?<=[.!?。！？])\s*|\n+", text)
+        for part in re.split(r"(?<=[.!?])\s+|(?<=[。！？])\s*|\n+", text)
         if part.strip()
     }
     spans.add(text.strip())

plugins/KnowledgeEngine/LongTermMemory/store/episode_provenance.py

@@ -15,6 +15,7 @@ CLAIM_TYPES = {"unknown", "user_statement", "inference"}
 
 def normalize_provenance(values: dict[str, Any]) -> dict[str, str]:
     result = {}
+    parsed_times = {}
     for field in ("event_at", "valid_from", "valid_until"):
         value = values.get(field, "")
         if not isinstance(value, str):
@@ -33,11 +34,8 @@ def normalize_provenance(values: dict[str, Any]) -> dict[str, str]:
         if parsed.tzinfo is None:
             raise ValueError(f"{field} must include an explicit timezone")
         try:
-            result[field] = (
-                parsed.astimezone(timezone.utc)
-                .isoformat(timespec="seconds")
-                .replace("+00:00", "Z")
-            )
+            parsed_times[field] = parsed.astimezone(timezone.utc)
+            result[field] = parsed_times[field].isoformat().replace("+00:00", "Z")
         except (ValueError, OverflowError) as exc:
             raise ValueError(
                 f"{field} is outside the supported UTC date range"
@@ -45,7 +43,7 @@ def normalize_provenance(values: dict[str, Any]) -> dict[str, str]:
     if (
         result["valid_from"]
         and result["valid_until"]
-        and result["valid_from"] >= result["valid_until"]
+        and parsed_times["valid_from"] >= parsed_times["valid_until"]
     ):
         raise ValueError(
             "valid_from must be earlier than valid_until; validity uses [from, until)"

plugins/KnowledgeEngine/LongTermMemory/tests/store/test_episode_provenance.py

@@ -292,3 +292,33 @@ async def test_document_ingestion_validates_and_preserves_provenance(invalid):
         assert meta["event_at"] == "2026-11-03T04:00:00Z"
         assert meta["claim_type"] == "user_statement"
         assert meta["source_ref"] == FACT["source_ref"]
+
+
+def test_subsecond_validity_and_offset_conversion_preserve_precision():
+    value = normalize_provenance(
+        {
+            "event_at": "2026-11-03T12:00:00.123456+08:00",
+            "valid_from": "2026-11-03T12:00:00.1+08:00",
+            "valid_until": "2026-11-03T12:00:00.2+08:00",
+        }
+    )
+    assert value["event_at"] == "2026-11-03T04:00:00.123456Z"
+    assert value["valid_from"] == "2026-11-03T04:00:00.100000Z"
+    assert value["valid_until"] == "2026-11-03T04:00:00.200000Z"
+
+
+def test_validity_compares_instants_instead_of_variable_precision_strings():
+    value = normalize_provenance(
+        {
+            "valid_from": "2026-11-03T04:00:00Z",
+            "valid_until": "2026-11-03T04:00:00.1Z",
+        }
+    )
+    assert value["valid_from"] == "2026-11-03T04:00:00Z"
+    with pytest.raises(ValueError, match="valid_from"):
+        normalize_provenance(
+            {
+                "valid_from": "2026-11-03T04:00:00.1Z",
+                "valid_until": "2026-11-03T04:00:00Z",
+            }
+        )

plugins/KnowledgeEngine/LongTermMemory/tests/store/test_model_candidate_extraction.py

@@ -185,3 +185,29 @@ async def test_sensitive_user_text_is_not_sent_to_candidate_model(monkeypatch):
         "bot:person_user", include_statuses=["accepted", "pending"]
     )
     assert all("do-not-store-this" not in json.dumps(item) for item in entries)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("quote", ["We do not support v1.", "5."])
+async def test_model_cannot_select_a_span_cut_inside_a_version_number(quote):
+    text = "We do not support v1.5. Use v2.0."
+    plugin = ModelPlugin(
+        json.dumps([{"candidate_type": "l2_episode", "evidence_quote": quote}])
+    )
+    with pytest.raises(ValueError, match="original span"):
+        await extract_candidates(
+            plugin, "model", text, "2026-11-02T00:00:00Z", "alice", 3
+        )
+
+
+@pytest.mark.asyncio
+async def test_full_version_statement_keeps_its_negation_and_decimal():
+    text = "We do not support v1.5. Use v2.0."
+    quote = "We do not support v1.5."
+    plugin = ModelPlugin(
+        json.dumps([{"candidate_type": "l2_episode", "evidence_quote": quote}])
+    )
+    result = await extract_candidates(
+        plugin, "model", text, "2026-11-02T00:00:00Z", "alice", 3
+    )
+    assert result[0]["payload"]["content"] == quote