@inerrata-corporation/errata 2.0.22-dev.4641 → 2.0.22-dev.4657

package/consolidate-worker.mjs

@@ -17126,6 +17126,7 @@ var EXPOSURE_SCOPE = "exposure";
 // ../../packages/local-graph/src/change-feed.ts
 var CHANGE_FEED_KEY = "change-feed";
 var ANY_CHANGE = "*";
+var nodeScope = (label) => `node:${label}`;
 var CHANGE_LOG_KEEP = 25e4;
 var PRUNED_TO_KEY = "change-log-pruned-to";
 function installChangeFeed(db) {
@@ -17596,8 +17597,10 @@ function column(c) {
 function compileNodeWhere(q) {
   if (q.labels.length === 0) throw new Error("selectNodes: no labels");
   const params = [...q.labels];
-  const conds = [`label IN (${q.labels.map(() => "?").join(", ")})`];
-  if (!q.includeClosed) conds.push("valid_to IS NULL");
+  const byId = (q.where ?? []).some((w) => "idIn" in w);
+  const unindexed = byId ? "+" : "";
+  const conds = [`${unindexed}label IN (${q.labels.map(() => "?").join(", ")})`];
+  if (!q.includeClosed) conds.push(`${unindexed}valid_to IS NULL`);
   for (const w of q.where ?? []) {
     if ("embedding" in w) {
       conds.push(`${EMBEDDING_DIM_SQL} ${w.embedding === "present" ? ">" : "="} 0`);
@@ -17649,6 +17652,14 @@ function compileNodeWhere(q) {
 function attrValueSql(p) {
   return `CASE json_type(attrs_json, ${p}) WHEN 'integer' THEN CAST(json_extract(attrs_json, ${p}) AS REAL) ELSE json_extract(attrs_json, ${p}) END`;
 }
+function nodeSelectFrom(q) {
+  if (q.indexedBy) return `nodes INDEXED BY ${selectIndex(q.indexedBy)}`;
+  const where = q.where ?? [];
+  if (!q.includeClosed && where.some((w) => "descriptionIn" in w && w.descriptionIn.length > 0) && !where.some((w) => "idIn" in w)) {
+    return "nodes INDEXED BY nodes_live_description";
+  }
+  return "nodes";
+}
 function compileNodeSelect(q) {
   const { where, params } = compileNodeWhere(q);
   const attrs = q.attrs ?? [];
@@ -17666,7 +17677,7 @@ function compileNodeSelect(q) {
   }
   const orderParams = !q.orderBy && q.labels.length > 1 ? [...q.labels] : [];
   const limit = q.limit !== void 0 ? ` LIMIT ${Math.max(0, Math.floor(q.limit))}` : "";
-  const sql = `SELECT id, label, description, created_at, last_updated_at, page_rank, cumulative_hits, extraction_confidence, extraction_source, community, created_at_seq, last_reinforced_at_seq, valid_to, logical_id, version, ${EMBEDDING_DIM_SQL} AS embedding_dim${q.withEmbedding ? ", embedding_blob, embedding_json" : ""}${attrCols} FROM nodes${q.indexedBy ? ` INDEXED BY ${selectIndex(q.indexedBy)}` : ""} WHERE ${where} ORDER BY ${order}${limit}`;
+  const sql = `SELECT id, label, description, created_at, last_updated_at, page_rank, cumulative_hits, extraction_confidence, extraction_source, community, created_at_seq, last_reinforced_at_seq, valid_to, logical_id, version, ${EMBEDDING_DIM_SQL} AS embedding_dim${q.withEmbedding ? ", embedding_blob, embedding_json" : ""}${attrCols} FROM ${nodeSelectFrom(q)} WHERE ${where} ORDER BY ${order}${limit}`;
   return { sql, params: [...params, ...orderParams] };
 }
 function decodeAttr(value, type) {
@@ -17903,6 +17914,18 @@ CREATE INDEX IF NOT EXISTS edges_live_prior_tag ON edges(created_at) WHERE valid
 -- incremental reindex can fetch just the changed files' symbols instead of
 -- scanning every versioned node. Partial: only live rows are ever queried.
 CREATE INDEX IF NOT EXISTS nodes_relpath ON nodes(json_extract(attrs_json, '$.relPath')) WHERE valid_to IS NULL;
+-- Symbols by qname and doc comments by the symbol they attach to: the drain's
+-- symbol lexicon resolves the names in the text it ships one at a time, instead
+-- of reading every Function/Method/Class/Interface/Const and Comment row on each
+-- live push (~300 MB of a 1.1 GB store per push, 2026-10-20).
+CREATE INDEX IF NOT EXISTS nodes_live_qname ON nodes(json_extract(attrs_json, '$.qname')) WHERE valid_to IS NULL;
+CREATE INDEX IF NOT EXISTS nodes_live_attached_to ON nodes(json_extract(attrs_json, '$.attachedTo')) WHERE valid_to IS NULL AND label = 'Comment';
+-- The attrs the cloud drain gates every instance row on, so it reads them from
+-- the index alone: pulled cloud rows are most of these labels (7,105 of 7,708
+-- Solutions on the live store) at ~10 KB of attrs each, and five full scans for
+-- two attrs read ~220 MB per push to find the few local rows that can ship.
+-- created_at leads so the index order is findNodeFieldsByLabel's.
+CREATE INDEX IF NOT EXISTS nodes_live_instance_gate ON nodes(label, created_at, json_extract(attrs_json, '$.source'), json_extract(attrs_json, '$.privateFinding'), json_extract(attrs_json, '$.resolvedAs'), id) WHERE valid_to IS NULL AND label IN ('Problem', 'Vulnerability', 'Solution', 'RootCause');
 -- Closed rows by relPath, for reviveReobserved (every incremental reindex): the
 -- live-only index above cannot serve it, so it scanned the whole table (~275 MB
 -- of a 1.1 GB store per reindex, 2026-10-20).
@@ -18018,6 +18041,7 @@ var SqliteGraphStore = class {
   initialPageRankScale;
   constructor(opts) {
     this.db = openDatabase(opts.path);
+    this.db.exec("PRAGMA cache_size = -32768");
     this.db.exec(SCHEMA_SQL);
     const versionRow = this.db.prepare("SELECT version FROM schema_version LIMIT 1").get();
     if (!versionRow) {
@@ -18082,6 +18106,7 @@ var SqliteGraphStore = class {
           attrs_json = excluded.attrs_json
       `),
       getNode: this.db.prepare("SELECT * FROM nodes WHERE id = ?"),
+      labelOf: this.db.prepare("SELECT label FROM nodes WHERE id = ?"),
       // Single-column fast paths for the nightly pass — writing 100k pageRank
       // values through the 18-column COALESCE updateNode took ~30s; these are ~10x
       // cheaper (one bound column, no COALESCE evaluation per row).
@@ -18269,6 +18294,36 @@ var SqliteGraphStore = class {
     }
     return out;
   }
+  firstLiveByNameOrQname(labels, name) {
+    if (labels.length === 0) return null;
+    const inLabels = labels.map(() => "?").join(",");
+    const probe = (index, match) => this.db.prepare(
+      `SELECT rowid AS rid, created_at AS at, label, description, json_extract(attrs_json, '$.qname') AS qname FROM nodes INDEXED BY ${index}
+          WHERE ${match} = ? AND label IN (${inLabels}) AND valid_to IS NULL`
+    ).all(name, ...labels);
+    const rank = (r) => [labels.indexOf(r.label), r.at, r.rid];
+    const before = (a, b) => a[0] < b[0] || a[0] === b[0] && (a[1] < b[1] || a[1] === b[1] && a[2] < b[2]);
+    let best;
+    for (const r of [...probe("nodes_live_description", "description"), ...probe("nodes_live_qname", "json_extract(attrs_json, '$.qname')")]) {
+      if (!best || before(rank(r), rank(best))) best = r;
+    }
+    return best ? { label: best.label, description: best.description, qname: typeof best.qname === "string" ? best.qname : null } : null;
+  }
+  instanceGateRows(label) {
+    return this.db.prepare(
+      `SELECT id, json_extract(attrs_json, '$.source') AS source, json_extract(attrs_json, '$.privateFinding') AS privateFinding,
+              json_extract(attrs_json, '$.resolvedAs') AS resolvedAs
+         FROM nodes INDEXED BY nodes_live_instance_gate
+        WHERE label IN ('Problem', 'Vulnerability', 'Solution', 'RootCause') AND label = ? AND valid_to IS NULL
+        ORDER BY label, created_at, rowid`
+    ).all(label);
+  }
+  liveCommentsAttachedTo(name) {
+    return this.db.prepare(
+      `SELECT json_extract(attrs_json, '$.kind') AS kind, json_extract(attrs_json, '$.text') AS text, description FROM nodes INDEXED BY nodes_live_attached_to
+        WHERE label = 'Comment' AND valid_to IS NULL AND json_extract(attrs_json, '$.attachedTo') = ? ORDER BY created_at, rowid`
+    ).all(name);
+  }
   strandedEdges() {
     return countStrandedEdges(this.db);
   }
@@ -18447,7 +18502,11 @@ var SqliteGraphStore = class {
     if (!stmt2) {
       const cols = attrKeys.map((k, i) => `, ${attrValueSql(`'$.${k}'`)} AS a${i}, json_type(attrs_json, '$.${k}') AS t${i}`).join("");
       stmt2 = this.db.prepare(
-        `SELECT id, description, community, extraction_confidence${cols} FROM nodes WHERE label = ? AND valid_to IS NULL`
+        // Ordered, not left to the planner: readers take "first row wins" (the
+        // symbol lexicon's name winner and doc pick), and the order otherwise
+        // followed whichever partial index the planner chose. nodes_live_created
+        // serves this order without a sort.
+        `SELECT id, description, community, extraction_confidence${cols} FROM nodes WHERE label = ? AND valid_to IS NULL ORDER BY created_at, rowid`
       );
       this.fieldStmts.set(cacheKey, stmt2);
     }
@@ -18473,7 +18532,7 @@ var SqliteGraphStore = class {
   }
   countNodes(q) {
     const { where, params } = compileNodeWhere(q);
-    return Number(this.db.prepare(`SELECT COUNT(*) AS n FROM nodes WHERE ${where}`).get(...params).n);
+    return Number(this.db.prepare(`SELECT COUNT(*) AS n FROM ${nodeSelectFrom(q)} WHERE ${where}`).get(...params).n);
   }
   setCommunity(id, community) {
     this.stmts.setCommunity.run({ id, v: community });
@@ -18547,11 +18606,13 @@ var SqliteGraphStore = class {
     }
   }
   edgeRuleViolation(edge) {
-    return localEdgeViolation(
-      this.getNode(edge.from)?.label,
-      edge.type,
-      this.getNode(edge.to)?.label
-    );
+    return localEdgeViolation(this.labelOf(edge.from), edge.type, this.labelOf(edge.to));
+  }
+  /** An endpoint's label, any version, as `getNode(id)?.label` reads it. Every
+   *  edge write checks both endpoints, and hydrating them (attrs + vector) to
+   *  read one leading column was most of an incremental reindex's reads. */
+  labelOf(id) {
+    return this.stmts.labelOf.get(id)?.label;
   }
   updateEdge(id, patch) {
     this.stmts.updateEdge.run({
@@ -18595,7 +18656,21 @@ var SqliteGraphStore = class {
    *  per project per request; the materializing form held the daemon's event
    *  loop for seconds at scale. */
   countNodesByLabel(label) {
-    return Number(this.stmts.countByLabel.get(label).n);
+    return this.countAt(`label:${label}`, [nodeScope(label)], () => Number(this.stmts.countByLabel.get(label).n));
+  }
+  /** Counts asked for on every status poll (/health, `errata status`, the
+   *  dashboards): a live-node COUNT(*) reads the whole liveness index, ~3 MB per
+   *  call on a 1.1 GB store, to recount a set that moves only on a write. Each is
+   *  kept with the change-feed reading of the rows it counts — triggers bump it on
+   *  every write from any connection, so a hit is exact, not a TTL guess. */
+  counts = /* @__PURE__ */ new Map();
+  countAt(key, scopes, count) {
+    const seq = readChangeSeq(this.db, scopes);
+    const hit = this.counts.get(key);
+    if (hit && hit.seq === seq) return hit.n;
+    const n = count();
+    this.counts.set(key, { seq, n });
+    return n;
   }
   /** Live rows for a label with attrs and creation stamps only — for scans that
    *  read attrs and never the vector. `findNodesByLabel` decodes every embedding
@@ -18862,7 +18937,7 @@ var SqliteGraphStore = class {
     return out;
   }
   nodeCount() {
-    return Number(this.stmts.nodeCount.get().n);
+    return this.countAt("nodes", [ANY_CHANGE], () => Number(this.stmts.nodeCount.get().n));
   }
   mutationCount() {
     return this.mutations;
@@ -18906,7 +18981,7 @@ var SqliteGraphStore = class {
     return this.ingestSeq;
   }
   edgeCount() {
-    return Number(this.stmts.edgeCount.get().n);
+    return this.countAt("edges", [ANY_CHANGE], () => Number(this.stmts.edgeCount.get().n));
   }
   // Lightweight whole-graph scans — select only the columns the caller needs so a
   // 100k-node pass skips the per-row attrs JSON.parse + embedding decode rowToNode
@@ -20893,6 +20968,62 @@ function buildSymbolLexicon(store) {
   }
   return lex;
 }
+function lazySymbolLexicon(store) {
+  if (!store.firstLiveByNameOrQname || !store.liveCommentsAttachedTo) return null;
+  const labels = SYMBOL_LABELS.map(([label]) => label);
+  const kindOf = new Map(SYMBOL_LABELS.map(([label, kind]) => [label, kind]));
+  const members = /* @__PURE__ */ new Map();
+  const docs = /* @__PURE__ */ new Map();
+  const entries = /* @__PURE__ */ new Map();
+  const member = (name) => {
+    if (!members.has(name)) {
+      const hit = isDistinctiveIdentifier(name) ? store.firstLiveByNameOrQname(labels, name) : null;
+      members.set(name, hit ? { kind: kindOf.get(hit.label), description: hit.description, qname: hit.qname } : null);
+    }
+    return members.get(name);
+  };
+  const docFor = (name) => {
+    if (!docs.has(name)) {
+      let found;
+      for (const c of store.liveCommentsAttachedTo(name)) {
+        if (typeof c.kind !== "string" || !DOC_COMMENT_KINDS.has(c.kind)) continue;
+        const text = typeof c.text === "string" ? c.text : c.description;
+        if (!text?.trim()) continue;
+        found = text.trim().slice(0, MAX_DOC_CHARS);
+        break;
+      }
+      docs.set(name, found);
+    }
+    return docs.get(name);
+  };
+  const isKnown = (tok) => member(tok) !== null;
+  let anySymbol;
+  return {
+    has: isKnown,
+    get(name) {
+      if (!entries.has(name)) {
+        const m = member(name);
+        let entry = null;
+        if (m) {
+          entry = { kind: m.kind };
+          const doc = (m.description ? docFor(m.description) : void 0) ?? (m.qname ? docFor(m.qname) : void 0);
+          for (const phrase of summaryCandidates(name, m.kind, doc)) {
+            if (summaryRejectionReason(phrase, isKnown) === null) {
+              entry.summary = phrase;
+              break;
+            }
+          }
+        }
+        entries.set(name, entry);
+      }
+      return entries.get(name) ?? void 0;
+    },
+    get size() {
+      anySymbol ??= labels.some((label) => store.countNodesByLabel(label) > 0);
+      return anySymbol ? 1 : 0;
+    }
+  };
+}
 var LEXICON_SYMBOL_LABELS = ["Function", "Method", "Class", "Interface", "Const", "Comment"];
 function codeGraphKey(store) {
   const files = store.labelFingerprint?.(["File"]);
@@ -20904,7 +21035,7 @@ function lexiconsFor(store) {
   const key = codeGraphKey(store);
   const hit = key !== null ? lexiconCache.get(store) : void 0;
   if (hit && hit.key === key) return hit;
-  const built = { symbols: buildSymbolLexicon(store), files: buildFileLexicon(store) };
+  const built = { symbols: lazySymbolLexicon(store) ?? buildSymbolLexicon(store), files: buildFileLexicon(store) };
   if (key !== null) lexiconCache.set(store, { key, ...built });
   return built;
 }
@@ -21143,6 +21274,38 @@ function stripCodebaseScope(scope) {
   delete s["codebase"];
   return s;
 }
+function instanceGateRows(store, label) {
+  return store.instanceGateRows?.(label) ?? store.findNodeFieldsByLabel(label, ["source", "privateFinding", "resolvedAs"]).map((r) => ({
+    id: r.id,
+    source: r.attrs["source"],
+    privateFinding: r.attrs["privateFinding"],
+    resolvedAs: r.attrs["resolvedAs"]
+  }));
+}
+var GATE_VIEW_ATTRS = [
+  "privateFinding",
+  "source",
+  "sources",
+  "resolvedAs",
+  "instanceReceipts",
+  "cloudWithdrawn",
+  "withdrawnAtCorrectionSeq",
+  "contributedAtSeq",
+  "cloudNodeId"
+];
+var GATE_VIEW_CHUNK = 500;
+var gateView = /* @__PURE__ */ new WeakSet();
+function gateViews(store, label, ids) {
+  const byId = /* @__PURE__ */ new Map();
+  for (let i = 0; i < ids.length; i += GATE_VIEW_CHUNK) {
+    for (const r of store.selectNodes({ labels: [label], attrs: GATE_VIEW_ATTRS, where: [{ idIn: ids.slice(i, i + GATE_VIEW_CHUNK) }] })) {
+      const view = { ...r, embedding: [] };
+      gateView.add(view);
+      byId.set(r.id, view);
+    }
+  }
+  return ids.flatMap((id) => byId.get(id) ?? []);
+}
 function contributedWireId(n, destinationKey) {
   const cloudId = instanceReceipt(n, destinationKey)["cloudNodeId"];
   return typeof cloudId === "string" && cloudId ? cloudId : n.id;
@@ -21301,7 +21464,7 @@ function buildInstanceIngest(store, profile, daemonVersion, ignorePatterns = [],
     edgesHeldDropped: 0
   };
   const sharedCauses = /* @__PURE__ */ new Map();
-  const causalQueue = store.findNodeFieldsByLabel("Problem", ["source", "resolvedAs"]).filter((r) => r.attrs["source"] !== "cloud" && r.attrs["resolvedAs"] !== PROBLEM_RESOLUTION.FALSE_POSITIVE).map((r) => r.id).filter((id) => ignorePatterns.length === 0 || !ignored(store.getNode(id)?.description ?? ""));
+  const causalQueue = instanceGateRows(store, "Problem").filter((r) => r.source !== "cloud" && r.resolvedAs !== PROBLEM_RESOLUTION.FALSE_POSITIVE).map((r) => r.id).filter((id) => ignorePatterns.length === 0 || !ignored(store.getNode(id)?.description ?? ""));
   const causalSeen = new Set(causalQueue);
   for (let i = 0; i < causalQueue.length; i++) {
     const id = causalQueue[i];
@@ -21317,21 +21480,21 @@ function buildInstanceIngest(store, profile, daemonVersion, ignorePatterns = [],
   for (const label of INSTANCE_LABELS) {
     const ref = [];
     includedByLabel.set(label, ref);
-    const local = [];
-    for (const row of store.findNodeFieldsByLabel(label, ["source", "privateFinding"])) {
-      if (row.attrs["source"] !== "cloud") {
-        const n = store.getNode(row.id);
-        if (n) local.push(n);
+    const localIds = [];
+    for (const row of instanceGateRows(store, label)) {
+      if (row.source !== "cloud") {
+        localIds.push(row.id);
         continue;
       }
       if (seen.has(row.id) || walked.has(row.id)) continue;
       walked.add(row.id);
       composition.candidates++;
-      if ((row.attrs["privateFinding"] || label === "Vulnerability") && !opts.project) composition.privateTargetMissing++;
+      if ((row.privateFinding || label === "Vulnerability") && !opts.project) composition.privateTargetMissing++;
       else composition.cloudSourced++;
     }
+    const local = gateViews(store, label, localIds);
     const population = label === "RootCause" && opts.sharedStore ? [...local, ...sharedCauses.values()] : local;
-    for (const n of population) {
+    for (let n of population) {
       if (seen.has(n.id)) continue;
       if (walked.has(n.id)) continue;
       walked.add(n.id);
@@ -21367,6 +21530,11 @@ function buildInstanceIngest(store, profile, daemonVersion, ignorePatterns = [],
         semanticEndpoints.set(n.id, contributedWireId(n, opts.destinationKey));
         continue;
       }
+      if (gateView.has(n)) {
+        const full = store.getNode(n.id);
+        if (!full) continue;
+        n = full;
+      }
       if (isEditPlaceholderSolution(n)) {
         composition.autoMint++;
         continue;

package/errata.mjs

@@ -20393,8 +20393,10 @@ function column(c) {
 function compileNodeWhere(q) {
   if (q.labels.length === 0) throw new Error("selectNodes: no labels");
   const params = [...q.labels];
-  const conds = [`label IN (${q.labels.map(() => "?").join(", ")})`];
-  if (!q.includeClosed) conds.push("valid_to IS NULL");
+  const byId = (q.where ?? []).some((w) => "idIn" in w);
+  const unindexed = byId ? "+" : "";
+  const conds = [`${unindexed}label IN (${q.labels.map(() => "?").join(", ")})`];
+  if (!q.includeClosed) conds.push(`${unindexed}valid_to IS NULL`);
   for (const w of q.where ?? []) {
     if ("embedding" in w) {
       conds.push(`${EMBEDDING_DIM_SQL} ${w.embedding === "present" ? ">" : "="} 0`);
@@ -20446,6 +20448,14 @@ function compileNodeWhere(q) {
 function attrValueSql(p) {
   return `CASE json_type(attrs_json, ${p}) WHEN 'integer' THEN CAST(json_extract(attrs_json, ${p}) AS REAL) ELSE json_extract(attrs_json, ${p}) END`;
 }
+function nodeSelectFrom(q) {
+  if (q.indexedBy) return `nodes INDEXED BY ${selectIndex(q.indexedBy)}`;
+  const where = q.where ?? [];
+  if (!q.includeClosed && where.some((w) => "descriptionIn" in w && w.descriptionIn.length > 0) && !where.some((w) => "idIn" in w)) {
+    return "nodes INDEXED BY nodes_live_description";
+  }
+  return "nodes";
+}
 function compileNodeSelect(q) {
   const { where, params } = compileNodeWhere(q);
   const attrs = q.attrs ?? [];
@@ -20463,7 +20473,7 @@ function compileNodeSelect(q) {
   }
   const orderParams = !q.orderBy && q.labels.length > 1 ? [...q.labels] : [];
   const limit = q.limit !== void 0 ? ` LIMIT ${Math.max(0, Math.floor(q.limit))}` : "";
-  const sql = `SELECT id, label, description, created_at, last_updated_at, page_rank, cumulative_hits, extraction_confidence, extraction_source, community, created_at_seq, last_reinforced_at_seq, valid_to, logical_id, version, ${EMBEDDING_DIM_SQL} AS embedding_dim${q.withEmbedding ? ", embedding_blob, embedding_json" : ""}${attrCols} FROM nodes${q.indexedBy ? ` INDEXED BY ${selectIndex(q.indexedBy)}` : ""} WHERE ${where} ORDER BY ${order}${limit}`;
+  const sql = `SELECT id, label, description, created_at, last_updated_at, page_rank, cumulative_hits, extraction_confidence, extraction_source, community, created_at_seq, last_reinforced_at_seq, valid_to, logical_id, version, ${EMBEDDING_DIM_SQL} AS embedding_dim${q.withEmbedding ? ", embedding_blob, embedding_json" : ""}${attrCols} FROM ${nodeSelectFrom(q)} WHERE ${where} ORDER BY ${order}${limit}`;
   return { sql, params: [...params, ...orderParams] };
 }
 function decodeAttr(value, type) {
@@ -20812,6 +20822,18 @@ CREATE INDEX IF NOT EXISTS edges_live_prior_tag ON edges(created_at) WHERE valid
 -- incremental reindex can fetch just the changed files' symbols instead of
 -- scanning every versioned node. Partial: only live rows are ever queried.
 CREATE INDEX IF NOT EXISTS nodes_relpath ON nodes(json_extract(attrs_json, '$.relPath')) WHERE valid_to IS NULL;
+-- Symbols by qname and doc comments by the symbol they attach to: the drain's
+-- symbol lexicon resolves the names in the text it ships one at a time, instead
+-- of reading every Function/Method/Class/Interface/Const and Comment row on each
+-- live push (~300 MB of a 1.1 GB store per push, 2026-10-20).
+CREATE INDEX IF NOT EXISTS nodes_live_qname ON nodes(json_extract(attrs_json, '$.qname')) WHERE valid_to IS NULL;
+CREATE INDEX IF NOT EXISTS nodes_live_attached_to ON nodes(json_extract(attrs_json, '$.attachedTo')) WHERE valid_to IS NULL AND label = 'Comment';
+-- The attrs the cloud drain gates every instance row on, so it reads them from
+-- the index alone: pulled cloud rows are most of these labels (7,105 of 7,708
+-- Solutions on the live store) at ~10 KB of attrs each, and five full scans for
+-- two attrs read ~220 MB per push to find the few local rows that can ship.
+-- created_at leads so the index order is findNodeFieldsByLabel's.
+CREATE INDEX IF NOT EXISTS nodes_live_instance_gate ON nodes(label, created_at, json_extract(attrs_json, '$.source'), json_extract(attrs_json, '$.privateFinding'), json_extract(attrs_json, '$.resolvedAs'), id) WHERE valid_to IS NULL AND label IN ('Problem', 'Vulnerability', 'Solution', 'RootCause');
 -- Closed rows by relPath, for reviveReobserved (every incremental reindex): the
 -- live-only index above cannot serve it, so it scanned the whole table (~275 MB
 -- of a 1.1 GB store per reindex, 2026-10-20).
@@ -20870,6 +20892,7 @@ CREATE INDEX IF NOT EXISTS nodes_revisit_true ON nodes(valid_from) WHERE json_ty
       initialPageRankScale;
       constructor(opts) {
         this.db = openDatabase(opts.path);
+        this.db.exec("PRAGMA cache_size = -32768");
         this.db.exec(SCHEMA_SQL);
         const versionRow = this.db.prepare("SELECT version FROM schema_version LIMIT 1").get();
         if (!versionRow) {
@@ -20934,6 +20957,7 @@ CREATE INDEX IF NOT EXISTS nodes_revisit_true ON nodes(valid_from) WHERE json_ty
           attrs_json = excluded.attrs_json
       `),
           getNode: this.db.prepare("SELECT * FROM nodes WHERE id = ?"),
+          labelOf: this.db.prepare("SELECT label FROM nodes WHERE id = ?"),
           // Single-column fast paths for the nightly pass — writing 100k pageRank
           // values through the 18-column COALESCE updateNode took ~30s; these are ~10x
           // cheaper (one bound column, no COALESCE evaluation per row).
@@ -21121,6 +21145,36 @@ CREATE INDEX IF NOT EXISTS nodes_revisit_true ON nodes(valid_from) WHERE json_ty
         }
         return out2;
       }
+      firstLiveByNameOrQname(labels, name2) {
+        if (labels.length === 0) return null;
+        const inLabels = labels.map(() => "?").join(",");
+        const probe = (index, match2) => this.db.prepare(
+          `SELECT rowid AS rid, created_at AS at, label, description, json_extract(attrs_json, '$.qname') AS qname FROM nodes INDEXED BY ${index}
+          WHERE ${match2} = ? AND label IN (${inLabels}) AND valid_to IS NULL`
+        ).all(name2, ...labels);
+        const rank = (r) => [labels.indexOf(r.label), r.at, r.rid];
+        const before = (a, b) => a[0] < b[0] || a[0] === b[0] && (a[1] < b[1] || a[1] === b[1] && a[2] < b[2]);
+        let best;
+        for (const r of [...probe("nodes_live_description", "description"), ...probe("nodes_live_qname", "json_extract(attrs_json, '$.qname')")]) {
+          if (!best || before(rank(r), rank(best))) best = r;
+        }
+        return best ? { label: best.label, description: best.description, qname: typeof best.qname === "string" ? best.qname : null } : null;
+      }
+      instanceGateRows(label) {
+        return this.db.prepare(
+          `SELECT id, json_extract(attrs_json, '$.source') AS source, json_extract(attrs_json, '$.privateFinding') AS privateFinding,
+              json_extract(attrs_json, '$.resolvedAs') AS resolvedAs
+         FROM nodes INDEXED BY nodes_live_instance_gate
+        WHERE label IN ('Problem', 'Vulnerability', 'Solution', 'RootCause') AND label = ? AND valid_to IS NULL
+        ORDER BY label, created_at, rowid`
+        ).all(label);
+      }
+      liveCommentsAttachedTo(name2) {
+        return this.db.prepare(
+          `SELECT json_extract(attrs_json, '$.kind') AS kind, json_extract(attrs_json, '$.text') AS text, description FROM nodes INDEXED BY nodes_live_attached_to
+        WHERE label = 'Comment' AND valid_to IS NULL AND json_extract(attrs_json, '$.attachedTo') = ? ORDER BY created_at, rowid`
+        ).all(name2);
+      }
       strandedEdges() {
         return countStrandedEdges(this.db);
       }
@@ -21299,7 +21353,11 @@ CREATE INDEX IF NOT EXISTS nodes_revisit_true ON nodes(valid_from) WHERE json_ty
         if (!stmt2) {
           const cols = attrKeys.map((k, i2) => `, ${attrValueSql(`'$.${k}'`)} AS a${i2}, json_type(attrs_json, '$.${k}') AS t${i2}`).join("");
           stmt2 = this.db.prepare(
-            `SELECT id, description, community, extraction_confidence${cols} FROM nodes WHERE label = ? AND valid_to IS NULL`
+            // Ordered, not left to the planner: readers take "first row wins" (the
+            // symbol lexicon's name winner and doc pick), and the order otherwise
+            // followed whichever partial index the planner chose. nodes_live_created
+            // serves this order without a sort.
+            `SELECT id, description, community, extraction_confidence${cols} FROM nodes WHERE label = ? AND valid_to IS NULL ORDER BY created_at, rowid`
           );
           this.fieldStmts.set(cacheKey2, stmt2);
         }
@@ -21325,7 +21383,7 @@ CREATE INDEX IF NOT EXISTS nodes_revisit_true ON nodes(valid_from) WHERE json_ty
       }
       countNodes(q) {
         const { where, params } = compileNodeWhere(q);
-        return Number(this.db.prepare(`SELECT COUNT(*) AS n FROM nodes WHERE ${where}`).get(...params).n);
+        return Number(this.db.prepare(`SELECT COUNT(*) AS n FROM ${nodeSelectFrom(q)} WHERE ${where}`).get(...params).n);
       }
       setCommunity(id, community) {
         this.stmts.setCommunity.run({ id, v: community });
@@ -21399,11 +21457,13 @@ CREATE INDEX IF NOT EXISTS nodes_revisit_true ON nodes(valid_from) WHERE json_ty
         }
       }
       edgeRuleViolation(edge2) {
-        return localEdgeViolation(
-          this.getNode(edge2.from)?.label,
-          edge2.type,
-          this.getNode(edge2.to)?.label
-        );
+        return localEdgeViolation(this.labelOf(edge2.from), edge2.type, this.labelOf(edge2.to));
+      }
+      /** An endpoint's label, any version, as `getNode(id)?.label` reads it. Every
+       *  edge write checks both endpoints, and hydrating them (attrs + vector) to
+       *  read one leading column was most of an incremental reindex's reads. */
+      labelOf(id) {
+        return this.stmts.labelOf.get(id)?.label;
       }
       updateEdge(id, patch) {
         this.stmts.updateEdge.run({
@@ -21447,7 +21507,21 @@ CREATE INDEX IF NOT EXISTS nodes_revisit_true ON nodes(valid_from) WHERE json_ty
        *  per project per request; the materializing form held the daemon's event
        *  loop for seconds at scale. */
       countNodesByLabel(label) {
-        return Number(this.stmts.countByLabel.get(label).n);
+        return this.countAt(`label:${label}`, [nodeScope(label)], () => Number(this.stmts.countByLabel.get(label).n));
+      }
+      /** Counts asked for on every status poll (/health, `errata status`, the
+       *  dashboards): a live-node COUNT(*) reads the whole liveness index, ~3 MB per
+       *  call on a 1.1 GB store, to recount a set that moves only on a write. Each is
+       *  kept with the change-feed reading of the rows it counts — triggers bump it on
+       *  every write from any connection, so a hit is exact, not a TTL guess. */
+      counts = /* @__PURE__ */ new Map();
+      countAt(key, scopes, count) {
+        const seq = readChangeSeq(this.db, scopes);
+        const hit = this.counts.get(key);
+        if (hit && hit.seq === seq) return hit.n;
+        const n = count();
+        this.counts.set(key, { seq, n });
+        return n;
       }
       /** Live rows for a label with attrs and creation stamps only — for scans that
        *  read attrs and never the vector. `findNodesByLabel` decodes every embedding
@@ -21714,7 +21788,7 @@ CREATE INDEX IF NOT EXISTS nodes_revisit_true ON nodes(valid_from) WHERE json_ty
         return out2;
       }
       nodeCount() {
-        return Number(this.stmts.nodeCount.get().n);
+        return this.countAt("nodes", [ANY_CHANGE], () => Number(this.stmts.nodeCount.get().n));
       }
       mutationCount() {
         return this.mutations;
@@ -21758,7 +21832,7 @@ CREATE INDEX IF NOT EXISTS nodes_revisit_true ON nodes(valid_from) WHERE json_ty
         return this.ingestSeq;
       }
       edgeCount() {
-        return Number(this.stmts.edgeCount.get().n);
+        return this.countAt("edges", [ANY_CHANGE], () => Number(this.stmts.edgeCount.get().n));
       }
       // Lightweight whole-graph scans — select only the columns the caller needs so a
       // 100k-node pass skips the per-row attrs JSON.parse + embedding decode rowToNode
@@ -37302,6 +37376,62 @@ function buildSymbolLexicon(store) {
   }
   return lex;
 }
+function lazySymbolLexicon(store) {
+  if (!store.firstLiveByNameOrQname || !store.liveCommentsAttachedTo) return null;
+  const labels = SYMBOL_LABELS2.map(([label]) => label);
+  const kindOf = new Map(SYMBOL_LABELS2.map(([label, kind]) => [label, kind]));
+  const members = /* @__PURE__ */ new Map();
+  const docs = /* @__PURE__ */ new Map();
+  const entries = /* @__PURE__ */ new Map();
+  const member = (name2) => {
+    if (!members.has(name2)) {
+      const hit = isDistinctiveIdentifier(name2) ? store.firstLiveByNameOrQname(labels, name2) : null;
+      members.set(name2, hit ? { kind: kindOf.get(hit.label), description: hit.description, qname: hit.qname } : null);
+    }
+    return members.get(name2);
+  };
+  const docFor = (name2) => {
+    if (!docs.has(name2)) {
+      let found;
+      for (const c of store.liveCommentsAttachedTo(name2)) {
+        if (typeof c.kind !== "string" || !DOC_COMMENT_KINDS.has(c.kind)) continue;
+        const text = typeof c.text === "string" ? c.text : c.description;
+        if (!text?.trim()) continue;
+        found = text.trim().slice(0, MAX_DOC_CHARS);
+        break;
+      }
+      docs.set(name2, found);
+    }
+    return docs.get(name2);
+  };
+  const isKnown = (tok) => member(tok) !== null;
+  let anySymbol;
+  return {
+    has: isKnown,
+    get(name2) {
+      if (!entries.has(name2)) {
+        const m = member(name2);
+        let entry = null;
+        if (m) {
+          entry = { kind: m.kind };
+          const doc = (m.description ? docFor(m.description) : void 0) ?? (m.qname ? docFor(m.qname) : void 0);
+          for (const phrase2 of summaryCandidates(name2, m.kind, doc)) {
+            if (summaryRejectionReason(phrase2, isKnown) === null) {
+              entry.summary = phrase2;
+              break;
+            }
+          }
+        }
+        entries.set(name2, entry);
+      }
+      return entries.get(name2) ?? void 0;
+    },
+    get size() {
+      anySymbol ??= labels.some((label) => store.countNodesByLabel(label) > 0);
+      return anySymbol ? 1 : 0;
+    }
+  };
+}
 function codeGraphKey(store) {
   const files = store.labelFingerprint?.(["File"]);
   if (files === void 0) return null;
@@ -37311,7 +37441,7 @@ function lexiconsFor(store) {
   const key = codeGraphKey(store);
   const hit = key !== null ? lexiconCache.get(store) : void 0;
   if (hit && hit.key === key) return hit;
-  const built = { symbols: buildSymbolLexicon(store), files: buildFileLexicon(store) };
+  const built = { symbols: lazySymbolLexicon(store) ?? buildSymbolLexicon(store), files: buildFileLexicon(store) };
   if (key !== null) lexiconCache.set(store, { key, ...built });
   return built;
 }
@@ -39491,9 +39621,10 @@ async function runIndexer(store, opts) {
       if (ei > 0) await new Promise((r) => setImmediate(r));
       const slice = pendingEdges.slice(ei, ei + EDGE_BATCH);
       store.transaction(() => {
+        const present = store.nodeLabels([...new Set(slice.flatMap((e) => [e.from, e.to]))]);
         for (const e of slice) {
           if (e.from === e.to) continue;
-          if (!store.getNode(e.from) || !store.getNode(e.to)) continue;
+          if (!present.has(e.from) || !present.has(e.to)) continue;
           upsertEdge(
             store,
             e.from,
@@ -86894,7 +87025,7 @@ function createWatchBreaker(deps) {
 }
 
 // src/engine.ts
-var DAEMON_VERSION = true ? "2.0.22-dev.4641" : "2.0.0-alpha.0";
+var DAEMON_VERSION = true ? "2.0.22-dev.4657" : "2.0.0-alpha.0";
 var IGNORED_PATH = /[\\/](?:\.git|node_modules|\.errata|\.claude|\.codex|\.cursor|\.turbo|\.next|dist|coverage|test-results|playwright-report|__pycache__)(?:[\\/]|$)/;
 var IGNORED_NOISE = /castalia\.db|eventlog\.sqlite|turn-cursors/;
 var GIT_OP_MUTE_MS = 4e3;
@@ -90732,6 +90863,38 @@ function compositionCounters(c) {
     ...Object.fromEntries(buckets.map((b) => [`sidecarDrop_${b}`, sidecarDropReasons[b] ?? 0]))
   };
 }
+function instanceGateRows(store, label) {
+  return store.instanceGateRows?.(label) ?? store.findNodeFieldsByLabel(label, ["source", "privateFinding", "resolvedAs"]).map((r) => ({
+    id: r.id,
+    source: r.attrs["source"],
+    privateFinding: r.attrs["privateFinding"],
+    resolvedAs: r.attrs["resolvedAs"]
+  }));
+}
+var GATE_VIEW_ATTRS = [
+  "privateFinding",
+  "source",
+  "sources",
+  "resolvedAs",
+  "instanceReceipts",
+  "cloudWithdrawn",
+  "withdrawnAtCorrectionSeq",
+  "contributedAtSeq",
+  "cloudNodeId"
+];
+var GATE_VIEW_CHUNK = 500;
+var gateView = /* @__PURE__ */ new WeakSet();
+function gateViews(store, label, ids) {
+  const byId = /* @__PURE__ */ new Map();
+  for (let i2 = 0; i2 < ids.length; i2 += GATE_VIEW_CHUNK) {
+    for (const r of store.selectNodes({ labels: [label], attrs: GATE_VIEW_ATTRS, where: [{ idIn: ids.slice(i2, i2 + GATE_VIEW_CHUNK) }] })) {
+      const view = { ...r, embedding: [] };
+      gateView.add(view);
+      byId.set(r.id, view);
+    }
+  }
+  return ids.flatMap((id) => byId.get(id) ?? []);
+}
 function contributedWireId(n, destinationKey) {
   const cloudId = instanceReceipt(n, destinationKey)["cloudNodeId"];
   return typeof cloudId === "string" && cloudId ? cloudId : n.id;
@@ -90890,7 +91053,7 @@ function buildInstanceIngest(store, profile, daemonVersion, ignorePatterns = [],
     edgesHeldDropped: 0
   };
   const sharedCauses = /* @__PURE__ */ new Map();
-  const causalQueue = store.findNodeFieldsByLabel("Problem", ["source", "resolvedAs"]).filter((r) => r.attrs["source"] !== "cloud" && r.attrs["resolvedAs"] !== PROBLEM_RESOLUTION.FALSE_POSITIVE).map((r) => r.id).filter((id) => ignorePatterns.length === 0 || !ignored(store.getNode(id)?.description ?? ""));
+  const causalQueue = instanceGateRows(store, "Problem").filter((r) => r.source !== "cloud" && r.resolvedAs !== PROBLEM_RESOLUTION.FALSE_POSITIVE).map((r) => r.id).filter((id) => ignorePatterns.length === 0 || !ignored(store.getNode(id)?.description ?? ""));
   const causalSeen = new Set(causalQueue);
   for (let i2 = 0; i2 < causalQueue.length; i2++) {
     const id = causalQueue[i2];
@@ -90906,21 +91069,21 @@ function buildInstanceIngest(store, profile, daemonVersion, ignorePatterns = [],
   for (const label of INSTANCE_LABELS) {
     const ref = [];
     includedByLabel.set(label, ref);
-    const local = [];
-    for (const row of store.findNodeFieldsByLabel(label, ["source", "privateFinding"])) {
-      if (row.attrs["source"] !== "cloud") {
-        const n = store.getNode(row.id);
-        if (n) local.push(n);
+    const localIds = [];
+    for (const row of instanceGateRows(store, label)) {
+      if (row.source !== "cloud") {
+        localIds.push(row.id);
         continue;
       }
       if (seen.has(row.id) || walked.has(row.id)) continue;
       walked.add(row.id);
       composition.candidates++;
-      if ((row.attrs["privateFinding"] || label === "Vulnerability") && !opts.project) composition.privateTargetMissing++;
+      if ((row.privateFinding || label === "Vulnerability") && !opts.project) composition.privateTargetMissing++;
       else composition.cloudSourced++;
     }
+    const local = gateViews(store, label, localIds);
     const population = label === "RootCause" && opts.sharedStore ? [...local, ...sharedCauses.values()] : local;
-    for (const n of population) {
+    for (let n of population) {
       if (seen.has(n.id)) continue;
       if (walked.has(n.id)) continue;
       walked.add(n.id);
@@ -90956,6 +91119,11 @@ function buildInstanceIngest(store, profile, daemonVersion, ignorePatterns = [],
         semanticEndpoints.set(n.id, contributedWireId(n, opts.destinationKey));
         continue;
       }
+      if (gateView.has(n)) {
+        const full = store.getNode(n.id);
+        if (!full) continue;
+        n = full;
+      }
       if (isEditPlaceholderSolution(n)) {
         composition.autoMint++;
         continue;

package/generalize-worker.mjs

@@ -15700,6 +15700,7 @@ var EXPOSURE_SCOPE = "exposure";
 // ../../packages/local-graph/src/change-feed.ts
 var CHANGE_FEED_KEY = "change-feed";
 var ANY_CHANGE = "*";
+var nodeScope = (label) => `node:${label}`;
 var CHANGE_LOG_KEEP = 25e4;
 var PRUNED_TO_KEY = "change-log-pruned-to";
 function installChangeFeed(db) {
@@ -15993,8 +15994,10 @@ function column(c) {
 function compileNodeWhere(q) {
   if (q.labels.length === 0) throw new Error("selectNodes: no labels");
   const params = [...q.labels];
-  const conds = [`label IN (${q.labels.map(() => "?").join(", ")})`];
-  if (!q.includeClosed) conds.push("valid_to IS NULL");
+  const byId = (q.where ?? []).some((w) => "idIn" in w);
+  const unindexed = byId ? "+" : "";
+  const conds = [`${unindexed}label IN (${q.labels.map(() => "?").join(", ")})`];
+  if (!q.includeClosed) conds.push(`${unindexed}valid_to IS NULL`);
   for (const w of q.where ?? []) {
     if ("embedding" in w) {
       conds.push(`${EMBEDDING_DIM_SQL} ${w.embedding === "present" ? ">" : "="} 0`);
@@ -16046,6 +16049,14 @@ function compileNodeWhere(q) {
 function attrValueSql(p) {
   return `CASE json_type(attrs_json, ${p}) WHEN 'integer' THEN CAST(json_extract(attrs_json, ${p}) AS REAL) ELSE json_extract(attrs_json, ${p}) END`;
 }
+function nodeSelectFrom(q) {
+  if (q.indexedBy) return `nodes INDEXED BY ${selectIndex(q.indexedBy)}`;
+  const where = q.where ?? [];
+  if (!q.includeClosed && where.some((w) => "descriptionIn" in w && w.descriptionIn.length > 0) && !where.some((w) => "idIn" in w)) {
+    return "nodes INDEXED BY nodes_live_description";
+  }
+  return "nodes";
+}
 function compileNodeSelect(q) {
   const { where, params } = compileNodeWhere(q);
   const attrs = q.attrs ?? [];
@@ -16063,7 +16074,7 @@ function compileNodeSelect(q) {
   }
   const orderParams = !q.orderBy && q.labels.length > 1 ? [...q.labels] : [];
   const limit = q.limit !== void 0 ? ` LIMIT ${Math.max(0, Math.floor(q.limit))}` : "";
-  const sql = `SELECT id, label, description, created_at, last_updated_at, page_rank, cumulative_hits, extraction_confidence, extraction_source, community, created_at_seq, last_reinforced_at_seq, valid_to, logical_id, version, ${EMBEDDING_DIM_SQL} AS embedding_dim${q.withEmbedding ? ", embedding_blob, embedding_json" : ""}${attrCols} FROM nodes${q.indexedBy ? ` INDEXED BY ${selectIndex(q.indexedBy)}` : ""} WHERE ${where} ORDER BY ${order}${limit}`;
+  const sql = `SELECT id, label, description, created_at, last_updated_at, page_rank, cumulative_hits, extraction_confidence, extraction_source, community, created_at_seq, last_reinforced_at_seq, valid_to, logical_id, version, ${EMBEDDING_DIM_SQL} AS embedding_dim${q.withEmbedding ? ", embedding_blob, embedding_json" : ""}${attrCols} FROM ${nodeSelectFrom(q)} WHERE ${where} ORDER BY ${order}${limit}`;
   return { sql, params: [...params, ...orderParams] };
 }
 function decodeAttr(value, type) {
@@ -16300,6 +16311,18 @@ CREATE INDEX IF NOT EXISTS edges_live_prior_tag ON edges(created_at) WHERE valid
 -- incremental reindex can fetch just the changed files' symbols instead of
 -- scanning every versioned node. Partial: only live rows are ever queried.
 CREATE INDEX IF NOT EXISTS nodes_relpath ON nodes(json_extract(attrs_json, '$.relPath')) WHERE valid_to IS NULL;
+-- Symbols by qname and doc comments by the symbol they attach to: the drain's
+-- symbol lexicon resolves the names in the text it ships one at a time, instead
+-- of reading every Function/Method/Class/Interface/Const and Comment row on each
+-- live push (~300 MB of a 1.1 GB store per push, 2026-10-20).
+CREATE INDEX IF NOT EXISTS nodes_live_qname ON nodes(json_extract(attrs_json, '$.qname')) WHERE valid_to IS NULL;
+CREATE INDEX IF NOT EXISTS nodes_live_attached_to ON nodes(json_extract(attrs_json, '$.attachedTo')) WHERE valid_to IS NULL AND label = 'Comment';
+-- The attrs the cloud drain gates every instance row on, so it reads them from
+-- the index alone: pulled cloud rows are most of these labels (7,105 of 7,708
+-- Solutions on the live store) at ~10 KB of attrs each, and five full scans for
+-- two attrs read ~220 MB per push to find the few local rows that can ship.
+-- created_at leads so the index order is findNodeFieldsByLabel's.
+CREATE INDEX IF NOT EXISTS nodes_live_instance_gate ON nodes(label, created_at, json_extract(attrs_json, '$.source'), json_extract(attrs_json, '$.privateFinding'), json_extract(attrs_json, '$.resolvedAs'), id) WHERE valid_to IS NULL AND label IN ('Problem', 'Vulnerability', 'Solution', 'RootCause');
 -- Closed rows by relPath, for reviveReobserved (every incremental reindex): the
 -- live-only index above cannot serve it, so it scanned the whole table (~275 MB
 -- of a 1.1 GB store per reindex, 2026-10-20).
@@ -16415,6 +16438,7 @@ var SqliteGraphStore = class {
   initialPageRankScale;
   constructor(opts) {
     this.db = openDatabase(opts.path);
+    this.db.exec("PRAGMA cache_size = -32768");
     this.db.exec(SCHEMA_SQL);
     const versionRow = this.db.prepare("SELECT version FROM schema_version LIMIT 1").get();
     if (!versionRow) {
@@ -16479,6 +16503,7 @@ var SqliteGraphStore = class {
           attrs_json = excluded.attrs_json
       `),
       getNode: this.db.prepare("SELECT * FROM nodes WHERE id = ?"),
+      labelOf: this.db.prepare("SELECT label FROM nodes WHERE id = ?"),
       // Single-column fast paths for the nightly pass — writing 100k pageRank
       // values through the 18-column COALESCE updateNode took ~30s; these are ~10x
       // cheaper (one bound column, no COALESCE evaluation per row).
@@ -16666,6 +16691,36 @@ var SqliteGraphStore = class {
     }
     return out;
   }
+  firstLiveByNameOrQname(labels, name) {
+    if (labels.length === 0) return null;
+    const inLabels = labels.map(() => "?").join(",");
+    const probe = (index, match) => this.db.prepare(
+      `SELECT rowid AS rid, created_at AS at, label, description, json_extract(attrs_json, '$.qname') AS qname FROM nodes INDEXED BY ${index}
+          WHERE ${match} = ? AND label IN (${inLabels}) AND valid_to IS NULL`
+    ).all(name, ...labels);
+    const rank = (r) => [labels.indexOf(r.label), r.at, r.rid];
+    const before = (a, b) => a[0] < b[0] || a[0] === b[0] && (a[1] < b[1] || a[1] === b[1] && a[2] < b[2]);
+    let best;
+    for (const r of [...probe("nodes_live_description", "description"), ...probe("nodes_live_qname", "json_extract(attrs_json, '$.qname')")]) {
+      if (!best || before(rank(r), rank(best))) best = r;
+    }
+    return best ? { label: best.label, description: best.description, qname: typeof best.qname === "string" ? best.qname : null } : null;
+  }
+  instanceGateRows(label) {
+    return this.db.prepare(
+      `SELECT id, json_extract(attrs_json, '$.source') AS source, json_extract(attrs_json, '$.privateFinding') AS privateFinding,
+              json_extract(attrs_json, '$.resolvedAs') AS resolvedAs
+         FROM nodes INDEXED BY nodes_live_instance_gate
+        WHERE label IN ('Problem', 'Vulnerability', 'Solution', 'RootCause') AND label = ? AND valid_to IS NULL
+        ORDER BY label, created_at, rowid`
+    ).all(label);
+  }
+  liveCommentsAttachedTo(name) {
+    return this.db.prepare(
+      `SELECT json_extract(attrs_json, '$.kind') AS kind, json_extract(attrs_json, '$.text') AS text, description FROM nodes INDEXED BY nodes_live_attached_to
+        WHERE label = 'Comment' AND valid_to IS NULL AND json_extract(attrs_json, '$.attachedTo') = ? ORDER BY created_at, rowid`
+    ).all(name);
+  }
   strandedEdges() {
     return countStrandedEdges(this.db);
   }
@@ -16844,7 +16899,11 @@ var SqliteGraphStore = class {
     if (!stmt2) {
       const cols = attrKeys.map((k, i) => `, ${attrValueSql(`'$.${k}'`)} AS a${i}, json_type(attrs_json, '$.${k}') AS t${i}`).join("");
       stmt2 = this.db.prepare(
-        `SELECT id, description, community, extraction_confidence${cols} FROM nodes WHERE label = ? AND valid_to IS NULL`
+        // Ordered, not left to the planner: readers take "first row wins" (the
+        // symbol lexicon's name winner and doc pick), and the order otherwise
+        // followed whichever partial index the planner chose. nodes_live_created
+        // serves this order without a sort.
+        `SELECT id, description, community, extraction_confidence${cols} FROM nodes WHERE label = ? AND valid_to IS NULL ORDER BY created_at, rowid`
       );
       this.fieldStmts.set(cacheKey, stmt2);
     }
@@ -16870,7 +16929,7 @@ var SqliteGraphStore = class {
   }
   countNodes(q) {
     const { where, params } = compileNodeWhere(q);
-    return Number(this.db.prepare(`SELECT COUNT(*) AS n FROM nodes WHERE ${where}`).get(...params).n);
+    return Number(this.db.prepare(`SELECT COUNT(*) AS n FROM ${nodeSelectFrom(q)} WHERE ${where}`).get(...params).n);
   }
   setCommunity(id, community) {
     this.stmts.setCommunity.run({ id, v: community });
@@ -16944,11 +17003,13 @@ var SqliteGraphStore = class {
     }
   }
   edgeRuleViolation(edge) {
-    return localEdgeViolation(
-      this.getNode(edge.from)?.label,
-      edge.type,
-      this.getNode(edge.to)?.label
-    );
+    return localEdgeViolation(this.labelOf(edge.from), edge.type, this.labelOf(edge.to));
+  }
+  /** An endpoint's label, any version, as `getNode(id)?.label` reads it. Every
+   *  edge write checks both endpoints, and hydrating them (attrs + vector) to
+   *  read one leading column was most of an incremental reindex's reads. */
+  labelOf(id) {
+    return this.stmts.labelOf.get(id)?.label;
   }
   updateEdge(id, patch) {
     this.stmts.updateEdge.run({
@@ -16992,7 +17053,21 @@ var SqliteGraphStore = class {
    *  per project per request; the materializing form held the daemon's event
    *  loop for seconds at scale. */
   countNodesByLabel(label) {
-    return Number(this.stmts.countByLabel.get(label).n);
+    return this.countAt(`label:${label}`, [nodeScope(label)], () => Number(this.stmts.countByLabel.get(label).n));
+  }
+  /** Counts asked for on every status poll (/health, `errata status`, the
+   *  dashboards): a live-node COUNT(*) reads the whole liveness index, ~3 MB per
+   *  call on a 1.1 GB store, to recount a set that moves only on a write. Each is
+   *  kept with the change-feed reading of the rows it counts — triggers bump it on
+   *  every write from any connection, so a hit is exact, not a TTL guess. */
+  counts = /* @__PURE__ */ new Map();
+  countAt(key, scopes, count) {
+    const seq = readChangeSeq(this.db, scopes);
+    const hit = this.counts.get(key);
+    if (hit && hit.seq === seq) return hit.n;
+    const n = count();
+    this.counts.set(key, { seq, n });
+    return n;
   }
   /** Live rows for a label with attrs and creation stamps only — for scans that
    *  read attrs and never the vector. `findNodesByLabel` decodes every embedding
@@ -17259,7 +17334,7 @@ var SqliteGraphStore = class {
     return out;
   }
   nodeCount() {
-    return Number(this.stmts.nodeCount.get().n);
+    return this.countAt("nodes", [ANY_CHANGE], () => Number(this.stmts.nodeCount.get().n));
   }
   mutationCount() {
     return this.mutations;
@@ -17303,7 +17378,7 @@ var SqliteGraphStore = class {
     return this.ingestSeq;
   }
   edgeCount() {
-    return Number(this.stmts.edgeCount.get().n);
+    return this.countAt("edges", [ANY_CHANGE], () => Number(this.stmts.edgeCount.get().n));
   }
   // Lightweight whole-graph scans — select only the columns the caller needs so a
   // 100k-node pass skips the per-row attrs JSON.parse + embedding decode rowToNode
@@ -18350,6 +18425,62 @@ function buildSymbolLexicon(store) {
   }
   return lex;
 }
+function lazySymbolLexicon(store) {
+  if (!store.firstLiveByNameOrQname || !store.liveCommentsAttachedTo) return null;
+  const labels = SYMBOL_LABELS.map(([label]) => label);
+  const kindOf = new Map(SYMBOL_LABELS.map(([label, kind]) => [label, kind]));
+  const members = /* @__PURE__ */ new Map();
+  const docs = /* @__PURE__ */ new Map();
+  const entries = /* @__PURE__ */ new Map();
+  const member = (name) => {
+    if (!members.has(name)) {
+      const hit = isDistinctiveIdentifier(name) ? store.firstLiveByNameOrQname(labels, name) : null;
+      members.set(name, hit ? { kind: kindOf.get(hit.label), description: hit.description, qname: hit.qname } : null);
+    }
+    return members.get(name);
+  };
+  const docFor = (name) => {
+    if (!docs.has(name)) {
+      let found;
+      for (const c of store.liveCommentsAttachedTo(name)) {
+        if (typeof c.kind !== "string" || !DOC_COMMENT_KINDS.has(c.kind)) continue;
+        const text = typeof c.text === "string" ? c.text : c.description;
+        if (!text?.trim()) continue;
+        found = text.trim().slice(0, MAX_DOC_CHARS);
+        break;
+      }
+      docs.set(name, found);
+    }
+    return docs.get(name);
+  };
+  const isKnown = (tok) => member(tok) !== null;
+  let anySymbol;
+  return {
+    has: isKnown,
+    get(name) {
+      if (!entries.has(name)) {
+        const m = member(name);
+        let entry = null;
+        if (m) {
+          entry = { kind: m.kind };
+          const doc = (m.description ? docFor(m.description) : void 0) ?? (m.qname ? docFor(m.qname) : void 0);
+          for (const phrase of summaryCandidates(name, m.kind, doc)) {
+            if (summaryRejectionReason(phrase, isKnown) === null) {
+              entry.summary = phrase;
+              break;
+            }
+          }
+        }
+        entries.set(name, entry);
+      }
+      return entries.get(name) ?? void 0;
+    },
+    get size() {
+      anySymbol ??= labels.some((label) => store.countNodesByLabel(label) > 0);
+      return anySymbol ? 1 : 0;
+    }
+  };
+}
 var LEXICON_SYMBOL_LABELS = ["Function", "Method", "Class", "Interface", "Const", "Comment"];
 function codeGraphKey(store) {
   const files = store.labelFingerprint?.(["File"]);
@@ -18361,7 +18492,7 @@ function lexiconsFor(store) {
   const key = codeGraphKey(store);
   const hit = key !== null ? lexiconCache.get(store) : void 0;
   if (hit && hit.key === key) return hit;
-  const built = { symbols: buildSymbolLexicon(store), files: buildFileLexicon(store) };
+  const built = { symbols: lazySymbolLexicon(store) ?? buildSymbolLexicon(store), files: buildFileLexicon(store) };
   if (key !== null) lexiconCache.set(store, { key, ...built });
   return built;
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@inerrata-corporation/errata",
-  "version": "2.0.22-dev.4641",
+  "version": "2.0.22-dev.4657",
   "description": "errata - local-first observation engine for AI coding agents",
   "type": "module",
   "bin": {

package/pass-worker.mjs

@@ -18280,9 +18280,10 @@ async function runIndexer(store2, opts) {
       if (ei > 0) await new Promise((r) => setImmediate(r));
       const slice = pendingEdges.slice(ei, ei + EDGE_BATCH);
       store2.transaction(() => {
+        const present = store2.nodeLabels([...new Set(slice.flatMap((e) => [e.from, e.to]))]);
         for (const e of slice) {
           if (e.from === e.to) continue;
-          if (!store2.getNode(e.from) || !store2.getNode(e.to)) continue;
+          if (!present.has(e.from) || !present.has(e.to)) continue;
           upsertEdge(
             store2,
             e.from,
@@ -26142,6 +26143,7 @@ var EXPOSURE_SCOPE = "exposure";
 // ../../packages/local-graph/src/change-feed.ts
 var CHANGE_FEED_KEY = "change-feed";
 var ANY_CHANGE = "*";
+var nodeScope = (label) => `node:${label}`;
 var CHANGE_LOG_KEEP = 25e4;
 var PRUNED_TO_KEY = "change-log-pruned-to";
 function installChangeFeed(db) {
@@ -26871,8 +26873,10 @@ function column(c) {
 function compileNodeWhere(q) {
   if (q.labels.length === 0) throw new Error("selectNodes: no labels");
   const params = [...q.labels];
-  const conds = [`label IN (${q.labels.map(() => "?").join(", ")})`];
-  if (!q.includeClosed) conds.push("valid_to IS NULL");
+  const byId = (q.where ?? []).some((w) => "idIn" in w);
+  const unindexed = byId ? "+" : "";
+  const conds = [`${unindexed}label IN (${q.labels.map(() => "?").join(", ")})`];
+  if (!q.includeClosed) conds.push(`${unindexed}valid_to IS NULL`);
   for (const w of q.where ?? []) {
     if ("embedding" in w) {
       conds.push(`${EMBEDDING_DIM_SQL} ${w.embedding === "present" ? ">" : "="} 0`);
@@ -26924,6 +26928,14 @@ function compileNodeWhere(q) {
 function attrValueSql(p) {
   return `CASE json_type(attrs_json, ${p}) WHEN 'integer' THEN CAST(json_extract(attrs_json, ${p}) AS REAL) ELSE json_extract(attrs_json, ${p}) END`;
 }
+function nodeSelectFrom(q) {
+  if (q.indexedBy) return `nodes INDEXED BY ${selectIndex(q.indexedBy)}`;
+  const where = q.where ?? [];
+  if (!q.includeClosed && where.some((w) => "descriptionIn" in w && w.descriptionIn.length > 0) && !where.some((w) => "idIn" in w)) {
+    return "nodes INDEXED BY nodes_live_description";
+  }
+  return "nodes";
+}
 function compileNodeSelect(q) {
   const { where, params } = compileNodeWhere(q);
   const attrs = q.attrs ?? [];
@@ -26941,7 +26953,7 @@ function compileNodeSelect(q) {
   }
   const orderParams = !q.orderBy && q.labels.length > 1 ? [...q.labels] : [];
   const limit = q.limit !== void 0 ? ` LIMIT ${Math.max(0, Math.floor(q.limit))}` : "";
-  const sql = `SELECT id, label, description, created_at, last_updated_at, page_rank, cumulative_hits, extraction_confidence, extraction_source, community, created_at_seq, last_reinforced_at_seq, valid_to, logical_id, version, ${EMBEDDING_DIM_SQL} AS embedding_dim${q.withEmbedding ? ", embedding_blob, embedding_json" : ""}${attrCols} FROM nodes${q.indexedBy ? ` INDEXED BY ${selectIndex(q.indexedBy)}` : ""} WHERE ${where} ORDER BY ${order}${limit}`;
+  const sql = `SELECT id, label, description, created_at, last_updated_at, page_rank, cumulative_hits, extraction_confidence, extraction_source, community, created_at_seq, last_reinforced_at_seq, valid_to, logical_id, version, ${EMBEDDING_DIM_SQL} AS embedding_dim${q.withEmbedding ? ", embedding_blob, embedding_json" : ""}${attrCols} FROM ${nodeSelectFrom(q)} WHERE ${where} ORDER BY ${order}${limit}`;
   return { sql, params: [...params, ...orderParams] };
 }
 function decodeAttr(value, type) {
@@ -27178,6 +27190,18 @@ CREATE INDEX IF NOT EXISTS edges_live_prior_tag ON edges(created_at) WHERE valid
 -- incremental reindex can fetch just the changed files' symbols instead of
 -- scanning every versioned node. Partial: only live rows are ever queried.
 CREATE INDEX IF NOT EXISTS nodes_relpath ON nodes(json_extract(attrs_json, '$.relPath')) WHERE valid_to IS NULL;
+-- Symbols by qname and doc comments by the symbol they attach to: the drain's
+-- symbol lexicon resolves the names in the text it ships one at a time, instead
+-- of reading every Function/Method/Class/Interface/Const and Comment row on each
+-- live push (~300 MB of a 1.1 GB store per push, 2026-10-20).
+CREATE INDEX IF NOT EXISTS nodes_live_qname ON nodes(json_extract(attrs_json, '$.qname')) WHERE valid_to IS NULL;
+CREATE INDEX IF NOT EXISTS nodes_live_attached_to ON nodes(json_extract(attrs_json, '$.attachedTo')) WHERE valid_to IS NULL AND label = 'Comment';
+-- The attrs the cloud drain gates every instance row on, so it reads them from
+-- the index alone: pulled cloud rows are most of these labels (7,105 of 7,708
+-- Solutions on the live store) at ~10 KB of attrs each, and five full scans for
+-- two attrs read ~220 MB per push to find the few local rows that can ship.
+-- created_at leads so the index order is findNodeFieldsByLabel's.
+CREATE INDEX IF NOT EXISTS nodes_live_instance_gate ON nodes(label, created_at, json_extract(attrs_json, '$.source'), json_extract(attrs_json, '$.privateFinding'), json_extract(attrs_json, '$.resolvedAs'), id) WHERE valid_to IS NULL AND label IN ('Problem', 'Vulnerability', 'Solution', 'RootCause');
 -- Closed rows by relPath, for reviveReobserved (every incremental reindex): the
 -- live-only index above cannot serve it, so it scanned the whole table (~275 MB
 -- of a 1.1 GB store per reindex, 2026-10-20).
@@ -27293,6 +27317,7 @@ var SqliteGraphStore = class {
   initialPageRankScale;
   constructor(opts) {
     this.db = openDatabase(opts.path);
+    this.db.exec("PRAGMA cache_size = -32768");
     this.db.exec(SCHEMA_SQL);
     const versionRow = this.db.prepare("SELECT version FROM schema_version LIMIT 1").get();
     if (!versionRow) {
@@ -27357,6 +27382,7 @@ var SqliteGraphStore = class {
           attrs_json = excluded.attrs_json
       `),
       getNode: this.db.prepare("SELECT * FROM nodes WHERE id = ?"),
+      labelOf: this.db.prepare("SELECT label FROM nodes WHERE id = ?"),
       // Single-column fast paths for the nightly pass — writing 100k pageRank
       // values through the 18-column COALESCE updateNode took ~30s; these are ~10x
       // cheaper (one bound column, no COALESCE evaluation per row).
@@ -27544,6 +27570,36 @@ var SqliteGraphStore = class {
     }
     return out2;
   }
+  firstLiveByNameOrQname(labels, name2) {
+    if (labels.length === 0) return null;
+    const inLabels = labels.map(() => "?").join(",");
+    const probe = (index, match) => this.db.prepare(
+      `SELECT rowid AS rid, created_at AS at, label, description, json_extract(attrs_json, '$.qname') AS qname FROM nodes INDEXED BY ${index}
+          WHERE ${match} = ? AND label IN (${inLabels}) AND valid_to IS NULL`
+    ).all(name2, ...labels);
+    const rank = (r) => [labels.indexOf(r.label), r.at, r.rid];
+    const before = (a, b) => a[0] < b[0] || a[0] === b[0] && (a[1] < b[1] || a[1] === b[1] && a[2] < b[2]);
+    let best;
+    for (const r of [...probe("nodes_live_description", "description"), ...probe("nodes_live_qname", "json_extract(attrs_json, '$.qname')")]) {
+      if (!best || before(rank(r), rank(best))) best = r;
+    }
+    return best ? { label: best.label, description: best.description, qname: typeof best.qname === "string" ? best.qname : null } : null;
+  }
+  instanceGateRows(label) {
+    return this.db.prepare(
+      `SELECT id, json_extract(attrs_json, '$.source') AS source, json_extract(attrs_json, '$.privateFinding') AS privateFinding,
+              json_extract(attrs_json, '$.resolvedAs') AS resolvedAs
+         FROM nodes INDEXED BY nodes_live_instance_gate
+        WHERE label IN ('Problem', 'Vulnerability', 'Solution', 'RootCause') AND label = ? AND valid_to IS NULL
+        ORDER BY label, created_at, rowid`
+    ).all(label);
+  }
+  liveCommentsAttachedTo(name2) {
+    return this.db.prepare(
+      `SELECT json_extract(attrs_json, '$.kind') AS kind, json_extract(attrs_json, '$.text') AS text, description FROM nodes INDEXED BY nodes_live_attached_to
+        WHERE label = 'Comment' AND valid_to IS NULL AND json_extract(attrs_json, '$.attachedTo') = ? ORDER BY created_at, rowid`
+    ).all(name2);
+  }
   strandedEdges() {
     return countStrandedEdges(this.db);
   }
@@ -27722,7 +27778,11 @@ var SqliteGraphStore = class {
     if (!stmt2) {
       const cols = attrKeys.map((k, i2) => `, ${attrValueSql(`'$.${k}'`)} AS a${i2}, json_type(attrs_json, '$.${k}') AS t${i2}`).join("");
       stmt2 = this.db.prepare(
-        `SELECT id, description, community, extraction_confidence${cols} FROM nodes WHERE label = ? AND valid_to IS NULL`
+        // Ordered, not left to the planner: readers take "first row wins" (the
+        // symbol lexicon's name winner and doc pick), and the order otherwise
+        // followed whichever partial index the planner chose. nodes_live_created
+        // serves this order without a sort.
+        `SELECT id, description, community, extraction_confidence${cols} FROM nodes WHERE label = ? AND valid_to IS NULL ORDER BY created_at, rowid`
       );
       this.fieldStmts.set(cacheKey, stmt2);
     }
@@ -27748,7 +27808,7 @@ var SqliteGraphStore = class {
   }
   countNodes(q) {
     const { where, params } = compileNodeWhere(q);
-    return Number(this.db.prepare(`SELECT COUNT(*) AS n FROM nodes WHERE ${where}`).get(...params).n);
+    return Number(this.db.prepare(`SELECT COUNT(*) AS n FROM ${nodeSelectFrom(q)} WHERE ${where}`).get(...params).n);
   }
   setCommunity(id, community) {
     this.stmts.setCommunity.run({ id, v: community });
@@ -27822,11 +27882,13 @@ var SqliteGraphStore = class {
     }
   }
   edgeRuleViolation(edge) {
-    return localEdgeViolation(
-      this.getNode(edge.from)?.label,
-      edge.type,
-      this.getNode(edge.to)?.label
-    );
+    return localEdgeViolation(this.labelOf(edge.from), edge.type, this.labelOf(edge.to));
+  }
+  /** An endpoint's label, any version, as `getNode(id)?.label` reads it. Every
+   *  edge write checks both endpoints, and hydrating them (attrs + vector) to
+   *  read one leading column was most of an incremental reindex's reads. */
+  labelOf(id) {
+    return this.stmts.labelOf.get(id)?.label;
   }
   updateEdge(id, patch) {
     this.stmts.updateEdge.run({
@@ -27870,7 +27932,21 @@ var SqliteGraphStore = class {
    *  per project per request; the materializing form held the daemon's event
    *  loop for seconds at scale. */
   countNodesByLabel(label) {
-    return Number(this.stmts.countByLabel.get(label).n);
+    return this.countAt(`label:${label}`, [nodeScope(label)], () => Number(this.stmts.countByLabel.get(label).n));
+  }
+  /** Counts asked for on every status poll (/health, `errata status`, the
+   *  dashboards): a live-node COUNT(*) reads the whole liveness index, ~3 MB per
+   *  call on a 1.1 GB store, to recount a set that moves only on a write. Each is
+   *  kept with the change-feed reading of the rows it counts — triggers bump it on
+   *  every write from any connection, so a hit is exact, not a TTL guess. */
+  counts = /* @__PURE__ */ new Map();
+  countAt(key, scopes, count) {
+    const seq = readChangeSeq(this.db, scopes);
+    const hit = this.counts.get(key);
+    if (hit && hit.seq === seq) return hit.n;
+    const n = count();
+    this.counts.set(key, { seq, n });
+    return n;
   }
   /** Live rows for a label with attrs and creation stamps only — for scans that
    *  read attrs and never the vector. `findNodesByLabel` decodes every embedding
@@ -28137,7 +28213,7 @@ var SqliteGraphStore = class {
     return out2;
   }
   nodeCount() {
-    return Number(this.stmts.nodeCount.get().n);
+    return this.countAt("nodes", [ANY_CHANGE], () => Number(this.stmts.nodeCount.get().n));
   }
   mutationCount() {
     return this.mutations;
@@ -28181,7 +28257,7 @@ var SqliteGraphStore = class {
     return this.ingestSeq;
   }
   edgeCount() {
-    return Number(this.stmts.edgeCount.get().n);
+    return this.countAt("edges", [ANY_CHANGE], () => Number(this.stmts.edgeCount.get().n));
   }
   // Lightweight whole-graph scans — select only the columns the caller needs so a
   // 100k-node pass skips the per-row attrs JSON.parse + embedding decode rowToNode

package/resources/context-plugin/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "errata-context",
-  "version": "2.0.22-dev.4641",
+  "version": "2.0.22-dev.4657",
   "description": "A minimal 3D view of loaded Errata priors and recent graph bursts.",
   "author": {
     "name": "inErrata"

package/resources/context-plugin/SHA256SUMS

@@ -1,4 +1,4 @@
-ad67194b8165c74c1b5c411cac4f968637fe21c28e678627b77597b0b54c2339  .codex-plugin/plugin.json
+e7f2b11ac106aa4f1ad1c492a48adfed53759ae7c7d5cba5f087ff791f67f48d  .codex-plugin/plugin.json
 a7bebae9fb01c838d6c75fc784129e28be571233fa66ff5ee8039b574f35ccf7  assets/icon.png
 fc38b64e4df1e0b72c35764bef6c4c05430d97c9009ad0ecc8a9e75b4deab232  configure-local.mjs
 73dc6028b3b070257c7c31f252d11e55a9e096f4aed0f187f10cb21716e1e6d7  dist/client.js