@ainyc/canonry 4.110.0 → 4.112.0

package/dist/{chunk-JGRVQJHX.js → chunk-UOMNYP5T.js}

@@ -102,6 +102,7 @@ import {
   contentTargetDismissalDtoSchema,
   contentTargetDismissalsResponseDtoSchema,
   contentTargetsResponseDtoSchema,
+  cosineSimilarity,
   createApiKeyRequestSchema,
   createdApiKeyDtoSchema,
   dedupeReportActions,
@@ -270,7 +271,7 @@ import {
   wordpressSchemaDeployResultDtoSchema,
   wordpressSchemaStatusResultDtoSchema,
   wordpressStatusDtoSchema
-} from "./chunk-YQ4KOC5Z.js";
+} from "./chunk-VICGMHY2.js";
 
 // src/intelligence-service.ts
 import { eq as eq38, desc as desc18, asc as asc6, and as and28, ne as ne5, or as or6, inArray as inArray14, gte as gte7, lte as lte4 } from "drizzle-orm";
@@ -1164,6 +1165,10 @@ var discoverySessions = sqliteTable("discovery_sessions", {
   // session identity for consolidation — and per-provider candidate counts.
   seedProviders: text("seed_providers", { mode: "json" }).$type(),
   seedProviderCounts: text("seed_provider_counts", { mode: "json" }).$type(),
+  // TRUE canonical count after dedup, BEFORE the probe-budget slice. seed_count
+  // is post-truncation (probedCanonicals.length), so a deliberately small
+  // maxProbes deflates it; this column is the honest numerator.
+  canonicalCount: integer("canonical_count"),
   dedupThreshold: real("dedup_threshold"),
   probeCount: integer("probe_count"),
   citedCount: integer("cited_count"),
@@ -3466,6 +3471,15 @@ var MIGRATION_VERSIONS = [
       `ALTER TABLE discovery_sessions ADD COLUMN seed_providers TEXT`,
       `ALTER TABLE discovery_sessions ADD COLUMN seed_provider_counts TEXT`
     ]
+  },
+  {
+    version: 94,
+    name: "discovery-session-canonical-count",
+    statements: [
+      // True post-dedup, pre-truncation canonical count (seed_count is
+      // post-truncation). Additive + nullable (downgrade-safe).
+      `ALTER TABLE discovery_sessions ADD COLUMN canonical_count INTEGER`
+    ]
   }
 ];
 function rebuildBacklinkTableWithSource(tx, table) {
@@ -34179,6 +34193,7 @@ function serializeSession(row) {
     seedRawCandidates: row.seedRawCandidates ?? null,
     seedProviders: row.seedProviders ?? null,
     seedProviderCounts: row.seedProviderCounts ?? null,
+    canonicalCount: row.canonicalCount ?? null,
     dedupClusterMinSims: row.dedupClusterMinSims ?? null,
     dedupBandPairFraction: row.dedupBandPairFraction ?? null,
     dedupPairsTotal: row.dedupPairsTotal ?? null,
@@ -34264,14 +34279,41 @@ async function classifyCompetitorDomains(deps, project, icpDescription, domains)
     return {};
   }
 }
-async function pickCanonicalsWithStats(candidates, deps, dedupThreshold) {
+async function pickCanonicalsWithStats(candidates, deps, dedupThreshold, primaryMask) {
   const emptyStats = { perClusterMinSimilarity: [], bandPairFraction: null, pairsTotal: 0 };
   if (candidates.length === 0) return { canonicals: [], stats: emptyStats };
   if (candidates.length === 1) return { canonicals: candidates, stats: emptyStats };
   const vectors = await deps.embed(candidates);
-  const indexClusters = clusterByCosine(candidates.map((_, i) => i), vectors, dedupThreshold);
-  const canonicals = indexClusters.map((cluster) => pickClusterRepresentative(cluster.map((i) => candidates[i])));
-  return { canonicals, stats: computeDedupSimilarityStats(vectors, indexClusters) };
+  const indices = candidates.map((_, i) => i);
+  const hasBothRoles = primaryMask !== void 0 && primaryMask.length === candidates.length && primaryMask.some((p) => p) && primaryMask.some((p) => !p);
+  if (!hasBothRoles) {
+    const indexClusters = clusterByCosine(indices, vectors, dedupThreshold);
+    const canonicals = indexClusters.map((cluster) => pickClusterRepresentative(cluster.map((i) => candidates[i])));
+    return { canonicals, stats: computeDedupSimilarityStats(vectors, indexClusters) };
+  }
+  const primaryIdx = indices.filter((i) => primaryMask[i]);
+  const primaryClusters = clusterByCosine(
+    primaryIdx,
+    primaryIdx.map((i) => vectors[i]),
+    dedupThreshold
+  );
+  const repIndexOf = (cluster) => {
+    const rep = pickClusterRepresentative(cluster.map((i) => candidates[i]));
+    return cluster.find((i) => candidates[i] === rep);
+  };
+  const keptIdx = primaryClusters.map(repIndexOf);
+  for (const s of indices.filter((i) => !primaryMask[i])) {
+    const novel = keptIdx.every((k) => cosineSimilarity(vectors[s], vectors[k]) < dedupThreshold);
+    if (novel) keptIdx.push(s);
+  }
+  return {
+    canonicals: keptIdx.map((i) => candidates[i]),
+    // Band fraction describes the candidate SPACE (not the clustering), so it is
+    // computed over all vectors as before; per-cluster cohesion reports the
+    // primary clusters.
+    stats: computeDedupSimilarityStats(vectors, primaryClusters),
+    primaryCanonicalCount: primaryClusters.length
+  };
 }
 async function executeDiscovery(opts) {
   const dedupThreshold = opts.dedupThreshold ?? DISCOVERY_DEFAULT_DEDUP_THRESHOLD;
@@ -34301,10 +34343,22 @@ async function executeDiscovery(opts) {
   });
   const rawCandidates = dedupeStrings(unbrandedCandidates);
   const seedCountRaw = rawCandidates.length;
+  const primaryProvider = opts.seedProviders?.[0];
+  const multiProvider = (opts.seedProviders?.length ?? 0) >= 2 && seedResult.candidateProviders != null;
+  let primaryMask;
+  if (multiProvider && primaryProvider) {
+    const providerOf = /* @__PURE__ */ new Map();
+    seedResult.candidates.forEach((candidate, i) => {
+      const key = candidate.trim().toLowerCase();
+      if (!providerOf.has(key)) providerOf.set(key, seedResult.candidateProviders[i]);
+    });
+    primaryMask = rawCandidates.map((c) => providerOf.get(c.trim().toLowerCase()) === primaryProvider);
+  }
   const { canonicals, stats: dedupStats } = await pickCanonicalsWithStats(
     rawCandidates,
     { embed: opts.deps.embed },
-    dedupThreshold
+    dedupThreshold,
+    primaryMask
   );
   const warning = seedCollapseWarning({
     seedCountRaw,
@@ -34323,6 +34377,7 @@ async function executeDiscovery(opts) {
     seedFromAnswerCount: seedResult.fromAnswerCount ?? null,
     seedFromGroundingCount: seedResult.fromGroundingCount ?? null,
     seedBrandFilteredCount: droppedBranded.length,
+    canonicalCount: canonicals.length,
     // Full seed provenance + dedup calibration diagnostics: the raw
     // candidate list (as the seed dep returned it, pre-filter) makes every
     // live session a replayable fixture; the similarity stats are the data

package/dist/{chunk-YQ4KOC5Z.js → chunk-VICGMHY2.js}

@@ -2246,6 +2246,9 @@ var discoverySessionDtoSchema = z17.object({
   seedProviders: z17.array(z17.string()).nullable().optional(),
   /** Raw candidate count contributed per seed provider. */
   seedProviderCounts: z17.record(z17.string(), z17.number().int()).nullable().optional(),
+  /** True post-dedup canonical count BEFORE the probe-budget slice (seedCount
+   *  is post-truncation). Null on legacy sessions. */
+  canonicalCount: z17.number().int().nullable().optional(),
   dedupThreshold: z17.number().nullable().optional(),
   probeCount: z17.number().int().nullable().optional(),
   citedCount: z17.number().int().nullable().default(null),
@@ -5760,6 +5763,7 @@ export {
   categorizeSource,
   categorizeSourceWithCompetitors,
   categoryLabel,
+  cosineSimilarity,
   clusterByCosine,
   pickClusterRepresentative,
   computeDedupSimilarityStats,