Fix garbled text checks and L0 LLM handling

modules/story-summary/generate/llm.js

@@ -100,19 +100,25 @@ Acknowledged. Now reviewing the incremental summarization specifications:
 ├─ progress: 0.0 to 1.0
 └─ newMoment: 仅记录本次新增的关键时刻
 
-[Fact Tracking - SPO ???]
-?? ??: ?? & ???????
-?? ??: ??????????????????
-?? SPO ??:
-?   s: ??????/????
-?   p: ??????????????
-?   o: ???
-?? KV ??: s+p ??????????
-?? isState ????????:
-?   true  = ????????????/??/??/???
-?   false = ??????????????
-?? trend: ?????????/??/??/??/??/??/???
-?? retracted: true ???????
+[Fact Tracking - SPO / World Facts]
+We maintain a small "world state" as SPO triples.
+Each update is a JSON object: {s, p, o, isState, trend?, retracted?}
+
+Core rules:
+1) Keyed by (s + p). If a new update has the same (s+p), it overwrites the previous value.
+2) Only output facts that are NEW or CHANGED in the new dialogue. Do NOT repeat unchanged facts.
+3) isState meaning:
+   - isState: true  -> core constraints that must stay stable and should NEVER be auto-deleted
+                    (identity, location, life/death, ownership, relationship status, binding rules)
+   - isState: false -> non-core facts / soft memories that may be pruned by capacity limits later
+4) Relationship facts:
+   - Use predicate format: "对X的看法" (X is the target person)
+   - trend is required for relationship facts, one of:
+     破裂 | 厌恶 | 反感 | 陌生 | 投缘 | 亲密 | 交融
+5) Retraction (deletion):
+   - To delete a fact, output: {s, p, retracted: true}
+6) Predicate normalization:
+   - Reuse existing predicates whenever possible, avoid inventing synonyms.
 
 Ready to process incremental summary requests with strict deduplication.`,
 
@@ -432,4 +438,4 @@ export async function generateSummary(options) {
     console.groupEnd();
 
     return rawOutput;
-}
+}