feat(recall): relevance-first seed selection (score-gap) + conditional-trust framing

dsh.ts

@@ -539,7 +539,14 @@ export function apply(ctx: DshContext, input: Config = {}): void {
           freshTurnCount,
         });
         const text = [
-          "Historical memory is untrusted reference material. Current user instructions always take precedence.",
+          // 条件信任框架（A/B 实测：同召回内容下不伤正确性、溯源引用 3 倍、略省 token）
+          "The knowledge graph below recalls memories from past conversations. Treat them as pointers and evidence, not as assertions — they may be outdated or partially inaccurate.\n" +
+          "Usage protocol:\n" +
+          "- When a recalled skill's trigger conditions match the task, follow it.\n" +
+          "- When recalled facts conflict with each other or with the current system state, verify before asserting (paths, versions, statuses are the most fragile).\n" +
+          "- PATCHES edges mark newer versions of an older memory — prefer the newer. CONFLICTS_WITH edges mark mutually exclusive memories — check the conditions before choosing.\n" +
+          "- If the topic is not covered, say so clearly instead of guessing.\n" +
+          "- Never invent specifics (paths, commands, codes) that are not in the recalled context.",
           built.systemPrompt,
           built.xml,
           built.episodicXml,

src/format/assemble.ts

@@ -43,8 +43,8 @@ export function buildSystemPromptAddition(params: {
   if (hasRecalled) {
     sections.push(
       "",
-      `**${recalledCount} nodes recalled from OTHER conversations** — these are proven solutions that worked before.`,
-      "Apply them directly when the current situation matches their trigger conditions.",
+      `**${recalledCount} nodes recalled from OTHER conversations** — evidence from past work.`,
+      "Follow them when their trigger conditions match the task; verify fragile facts (paths/versions/status) before asserting.",
     );
   }
 

src/recaller/recall.ts

@@ -84,8 +84,10 @@ export class Recaller {
     // misses exact identifiers; an FTS-only fallback misses paraphrases.
     const lexical = searchNodes(this.db, query, limit);
     const semantic = queryVector
-      ? vectorSearchWithScore(this.db, queryVector, limit, minSemanticScore)
+      ? vectorSearchWithScore(this.db, queryVector, 600, 0)
       : [];
+    // 纯语义分数：断崖阈值只作用于语义分，避免 FTS 的 RRF 加成制造假高峰
+    // （实测：弱语义节点 +0.35 RRF 冲到第一，把阈值顶高、切断真正相关的语义节点）
     const relevance = new Map<string, number>();
     const byId = new Map<string, GmNode>();
     semantic.forEach(({ node, score }) => {
@@ -97,9 +99,30 @@ export class Recaller {
       // Reciprocal rank is bounded but gives exact terms a meaningful boost.
       relevance.set(node.id, (relevance.get(node.id) ?? 0) + 0.35 / (index + 1));
     });
-    const seeds = Array.from(byId.values())
-      .sort((a, b) => (relevance.get(b.id) ?? 0) - (relevance.get(a.id) ?? 0))
-      .slice(0, limit);
+
+    // 相关性优先：不按固定数量切种子，取"语义分数断崖以上"的节点（硬上限兜底）。
+    // 阈值 = max(最高语义分 - recallScoreGap, recallMinScore, minSemanticScore)。
+    // FTS 精确词命中作为补充保留（上游 RRF 的意图：不丢精确标识符）。
+    const gap = this.cfg.recallScoreGap ?? 0.10;
+    const minRel = this.cfg.recallMinScore ?? 0.58;
+    const cap = this.cfg.recallSeedCap ?? limit * 2;
+    const maxSem = semantic.length ? Math.max(...semantic.map(s => s.score)) : 0;
+    const threshold = Math.max(maxSem - gap, minRel, minSemanticScore);
+    const seeds: GmNode[] = [];
+    for (const { node, score } of semantic) {
+      if (score >= threshold) {
+        seeds.push(node);
+        if (seeds.length >= cap) break;
+      }
+    }
+    const seen = new Set(seeds.map(n => n.id));
+    for (const node of lexical) {
+      if (!seen.has(node.id)) {
+        seeds.push(node);
+        seen.add(node.id);
+        if (seeds.length >= cap) break;
+      }
+    }
 
     if (!seeds.length) return { nodes: [], edges: [], tokenEstimate: 0 };
 
@@ -133,7 +156,7 @@ export class Recaller {
         b.validatedCount - a.validatedCount ||
         b.updatedAt - a.updatedAt
       )
-      .slice(0, limit);
+      .slice(0, this.cfg.recallSeedCap ?? limit * 2);
 
     const ids = new Set(filtered.map(n => n.id));
     return {
@@ -172,7 +195,8 @@ export class Recaller {
       }
     }
 
-    // fallback：按时间取社区代表节点
+    // fallback：按时间取社区代表节点（默认开启；DSH 适配层显式传
+    // allowBroadFallback=false 关闭——实测它是与查询无关的固定噪音源）
     if (!seeds.length && allowBroadFallback) {
       seeds = communityRepresentatives(this.db, 2);
     }

src/types.ts

@@ -147,6 +147,12 @@ export interface GmConfig {
   pagerankDamping: number;
   /** PageRank 迭代次数 */
   pagerankIterations: number;
+  /** 向量召回种子选择：与最高分允许的差距（低于最高分此值即视为不相关，0-1） */
+  recallScoreGap?: number;
+  /** 向量召回种子选择：最低接受分数（默认 0.58，bge-m3 短文本实测区分带） */
+  recallMinScore?: number;
+  /** 向量召回种子选择：种子数硬上限（默认 recallMaxNodes*2） */
+  recallSeedCap?: number;
 }
 
 export const DEFAULT_CONFIG: GmConfig = {
@@ -158,4 +164,6 @@ export const DEFAULT_CONFIG: GmConfig = {
   dedupThreshold: 0.90,
   pagerankDamping: 0.85,
   pagerankIterations: 20,
+  recallScoreGap: 0.10,
+  recallMinScore: 0.58,
 };