Improve rerank failure handling and tokenizer JP support

modules/story-summary/vector/retrieval/metrics.js

@@ -86,6 +86,7 @@ export function createMetrics() {
             l0Candidates: 0,       // W-RRF 融合后的 L0 候选数
             l0Selected: 0,         // rerank 后选中的 L0 数
             rerankApplied: false,
+            rerankFailed: false,
             beforeRerank: 0,
             afterRerank: 0,
             rerankTime: 0,
@@ -283,6 +284,9 @@ export function formatMetricsLog(metrics) {
 
     if (m.evidence.rerankApplied) {
         lines.push(`│   ├─ rerank_applied: true`);
+        if (m.evidence.rerankFailed) {
+            lines.push(`│   ├─ rerank_failed: ⚠ YES (using fusion order)`);
+        }
         lines.push(`│   │   ├─ before: ${m.evidence.beforeRerank}`);
         lines.push(`│   │   ├─ after: ${m.evidence.afterRerank}`);
         lines.push(`│   │   └─ time: ${m.evidence.rerankTime}ms`);
@@ -489,6 +493,10 @@ export function detectIssues(metrics) {
     // L1 挂载问题
     // ─────────────────────────────────────────────────────────────────
 
+    if (m.evidence.rerankFailed) {
+        issues.push('Rerank API failed — using fusion rank order as fallback, relevance scores are zero');
+    }
+
     if (m.evidence.l0Selected > 0 && m.evidence.l1Pulled === 0) {
         issues.push('Zero L1 chunks pulled - L1 vectors may not exist or DB read failed');
     }

modules/story-summary/vector/retrieval/recall.js

@@ -631,6 +631,7 @@ async function locateAndPullEvidence(anchorHits, anchorFloors, queryVector, rera
         metrics.evidence.rerankApplied = true;
         metrics.evidence.beforeRerank = rerankCandidates.length;
         metrics.evidence.afterRerank = rerankedL0.length;
+        metrics.evidence.rerankFailed = rerankedL0.some(c => c._rerankFailed);
         metrics.evidence.l0Selected = rerankedL0.length;
         metrics.evidence.rerankTime = rerankTime;
         metrics.timing.evidenceRerank = rerankTime;