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>

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 {