Improve UMS candidate scoring and snapshot materialization for sparse data.

Lower term threshold, tune promotion scores, project candidate fallback in snapshot, and add rebuild-ums-snapshot script for production backfill.

Co-authored-by: Cursor <cursoragent@cursor.com>

scripts/rebuild-ums-snapshot.mjs

@@ -0,0 +1,38 @@
+#!/usr/bin/env node
+/**
+ * Rebuild UMS candidates + snapshot from existing signals (no new evidence ingest).
+ *
+ *   UMS_DATABASE_URL=... node scripts/rebuild-ums-snapshot.mjs [user_id]
+ */
+import { createUmsPool, isUmsDatabaseConfigured } from '../user-model-service/db.mjs';
+import { mergeCandidatesFromSignals } from '../user-model-service/candidates.mjs';
+import { materializeProfileAndSnapshot } from '../user-model-service/snapshot.mjs';
+import { resolveCanonicalUserId } from '../user-model-service/canonical-user.mjs';
+
+async function main() {
+  if (!isUmsDatabaseConfigured()) {
+    console.error('Set UMS_DATABASE_URL');
+    process.exit(1);
+  }
+  const userId = resolveCanonicalUserId(
+    process.argv[2] ?? process.env.UMS_REBUILD_USER_ID ?? 'a70ff537-8908-486e-9b6c-042e07cc25db',
+  );
+  const pool = createUmsPool();
+  try {
+    const candidates = await mergeCandidatesFromSignals(pool, userId);
+    const snapshot = await materializeProfileAndSnapshot(pool, userId, { reason: 'rebuild_script' });
+    console.log('rebuild OK:', {
+      user_id: userId,
+      candidates_touched: candidates,
+      profile_version: snapshot.profile_version,
+      byte_size: snapshot.byte_size,
+    });
+  } finally {
+    await pool.end();
+  }
+}
+
+main().catch((err) => {
+  console.error(err);
+  process.exit(1);
+});

user-model-service/candidates.mjs

@@ -12,6 +12,14 @@ function slugEntityName(name) {
   return String(name).trim().slice(0, 128);
 }
 
+const MIN_TERM_COUNT = Number(process.env.UMS_CANDIDATE_MIN_TERM_COUNT ?? 3);
+
+function isProjectLikeTerm(term) {
+  if (/^[A-Z][a-zA-Z0-9]+$/.test(term)) return true;
+  if (/meinput|memind|input|agent|rime|fcitx|tkmind|goosed|pgvector/i.test(term)) return true;
+  return term.includes('Input') || term.includes('Mind');
+}
+
 /**
  * V0.1：从 term_frequency signals 推断 project/focus candidates
  * @param {import('mysql2/promise').Pool} pool
@@ -43,13 +51,13 @@ export async function mergeCandidatesFromSignals(pool, userId) {
   let touched = 0;
   const ts = nowMs();
   for (const [term, stats] of termCounts) {
-    if (stats.count < 5) continue;
+    if (stats.count < MIN_TERM_COUNT) continue;
     if (term.length < 2) continue;
 
-    const isProjectLike = /^[A-Z][a-zA-Z0-9]+$/.test(term) || term.includes('Input') || term.includes('Mind');
+    const isProjectLike = isProjectLikeTerm(term);
     const candidateType = isProjectLike ? 'project' : 'focus';
-    const confidence = Math.min(0.99, 0.4 + stats.count * 0.03);
+    const confidence = Math.min(0.99, 0.4 + stats.count * 0.05);
-    const promotionScore = Math.min(0.99, confidence * Math.min(1, stats.count / 20));
+    const promotionScore = Math.min(0.99, 0.3 + stats.count * 0.1);
     const hypothesis = isProjectLike
       ? { type: 'project', name: term, status: 'active' }
       : { type: 'focus', topic: term };
@@ -166,3 +174,5 @@ async function upsertProjectGraph(pool, userId, name, confidence, evidenceIds, w
     );
   }
 }
+
+export { isProjectLikeTerm, MIN_TERM_COUNT };

user-model-service/snapshot.mjs

@@ -41,6 +41,15 @@ export async function materializeProfileAndSnapshot(pool, userId, options = {})
     [userId],
   );
 
+  const [projectCandidates] = await pool.query(
+    `SELECT hypothesis_json, confidence, promotion_score, last_seen_at
+     FROM um_candidates
+     WHERE user_id = ? AND candidate_type = 'project' AND status IN ('open', 'accepted', 'observed')
+     ORDER BY promotion_score DESC
+     LIMIT 8`,
+    [userId],
+  );
+
   const [versionRows] = await pool.query(
     `SELECT COALESCE(MAX(profile_version), 0) AS v FROM um_profile_versions WHERE user_id = ?`,
     [userId],
@@ -55,6 +64,20 @@ export async function materializeProfileAndSnapshot(pool, userId, options = {})
     last_seen: row.last_seen_at,
   }));
 
+  if (activeProjects.length === 0 && projectCandidates.length > 0) {
+    for (const row of projectCandidates) {
+      const h = typeof row.hypothesis_json === 'string' ? JSON.parse(row.hypothesis_json) : row.hypothesis_json;
+      const name = h.name ?? h.topic ?? 'unknown';
+      activeProjects.push({
+        id: `project_${String(name).toLowerCase().replace(/[^a-z0-9]+/g, '_')}`,
+        name,
+        status: h.status ?? 'active',
+        confidence: Number(row.confidence),
+        last_seen: row.last_seen_at,
+      });
+    }
+  }
+
   const recentFocus = focusCandidates.map((row) => {
     const h = typeof row.hypothesis_json === 'string' ? JSON.parse(row.hypothesis_json) : row.hypothesis_json;
     return {