@anuma/sdk 6.10.0 → 6.11.0

package/dist/server/index.d.mts

@@ -5641,6 +5641,17 @@ interface RecallDiagnostics {
      * never the cost.
      */
     vaultRowsDecrypted?: number;
+    /**
+     * Rows the fact lane had to re-embed through the portal because their stored
+     * vector was unusable — stale `embedding_model`, wrong dimension, or
+     * unparseable. Absent when the fact lane didn't run.
+     *
+     * Expected to be 0 on a healthy vault: rows are embedded at write time and the
+     * re-embed writes the current model back, so a persistently non-zero value
+     * means the writeback is not sticking and every turn is paying for it. On the
+     * LEGACY read path this batch is uncapped, so it can be the whole vault.
+     */
+    vaultRowsEmbedded?: number;
     /** Facts the fact lane returned (post-dedupe, pre-fusion). */
     factCount: number;
     /** Chunks the chunk lane returned (post-dedupe, pre-fusion). */
@@ -5653,6 +5664,37 @@ interface RecallDiagnostics {
         prep: number;
         /** Vault fact-lane search (`searchVaultMemoriesWithSize`). */
         factLane: number;
+        /**
+         * The cross-encoder's share of {@link factLane} — wall-clock spent inside
+         * `rerankPairs`. Billed even when the CE threw partway through, because a
+         * rerank that burned three seconds and then failed still cost them.
+         *
+         * A SUBSET of `factLane`, not a sibling: read `factLane - rerank` for
+         * everything else the lane did (query embed, vault read, fused ranking).
+         * 0 when the CE did not run; `reranked` is what distinguishes that from a
+         * rerank that cost nothing.
+         *
+         * Exists because #845 spent three rounds arguing about which stage inside
+         * the fact lane dominated — whole-vault read, then admission window, then
+         * the CE — with one aggregate number for all of them. Every hypothesis was
+         * an inference; this makes the question a query.
+         */
+        rerank: number;
+        /**
+         * The query embed's share of {@link factLane} — one portal round trip.
+         *
+         * Also a SUBSET of `factLane`. Together with `rerank` this makes the lane
+         * decomposable: `factLane - queryEmbed - rerank` is the local work (vault
+         * read, decrypt, BM25/cosine fusion), and each of the three implies a
+         * completely different fix.
+         *
+         * 0 when the lane returned before embedding, which includes every empty
+         * vault — `prepareVaultCandidates` short-circuits first. That is why the
+         * fast `vault_size = 0` population never established a baseline for this
+         * cost, and why the ~850ms floor on the smallest NON-empty vaults had no
+         * attributable owner.
+         */
+        queryEmbed: number;
         /** Chunk-lane search (`searchChunksOp`). */
         chunkLane: number;
         /** Cross-lane RRF fusion + provenance dedup after both lanes. */

package/dist/server/index.d.ts

@@ -5641,6 +5641,17 @@ interface RecallDiagnostics {
      * never the cost.
      */
     vaultRowsDecrypted?: number;
+    /**
+     * Rows the fact lane had to re-embed through the portal because their stored
+     * vector was unusable — stale `embedding_model`, wrong dimension, or
+     * unparseable. Absent when the fact lane didn't run.
+     *
+     * Expected to be 0 on a healthy vault: rows are embedded at write time and the
+     * re-embed writes the current model back, so a persistently non-zero value
+     * means the writeback is not sticking and every turn is paying for it. On the
+     * LEGACY read path this batch is uncapped, so it can be the whole vault.
+     */
+    vaultRowsEmbedded?: number;
     /** Facts the fact lane returned (post-dedupe, pre-fusion). */
     factCount: number;
     /** Chunks the chunk lane returned (post-dedupe, pre-fusion). */
@@ -5653,6 +5664,37 @@ interface RecallDiagnostics {
         prep: number;
         /** Vault fact-lane search (`searchVaultMemoriesWithSize`). */
         factLane: number;
+        /**
+         * The cross-encoder's share of {@link factLane} — wall-clock spent inside
+         * `rerankPairs`. Billed even when the CE threw partway through, because a
+         * rerank that burned three seconds and then failed still cost them.
+         *
+         * A SUBSET of `factLane`, not a sibling: read `factLane - rerank` for
+         * everything else the lane did (query embed, vault read, fused ranking).
+         * 0 when the CE did not run; `reranked` is what distinguishes that from a
+         * rerank that cost nothing.
+         *
+         * Exists because #845 spent three rounds arguing about which stage inside
+         * the fact lane dominated — whole-vault read, then admission window, then
+         * the CE — with one aggregate number for all of them. Every hypothesis was
+         * an inference; this makes the question a query.
+         */
+        rerank: number;
+        /**
+         * The query embed's share of {@link factLane} — one portal round trip.
+         *
+         * Also a SUBSET of `factLane`. Together with `rerank` this makes the lane
+         * decomposable: `factLane - queryEmbed - rerank` is the local work (vault
+         * read, decrypt, BM25/cosine fusion), and each of the three implies a
+         * completely different fix.
+         *
+         * 0 when the lane returned before embedding, which includes every empty
+         * vault — `prepareVaultCandidates` short-circuits first. That is why the
+         * fast `vault_size = 0` population never established a baseline for this
+         * cost, and why the ~850ms floor on the smallest NON-empty vaults had no
+         * attributable owner.
+         */
+        queryEmbed: number;
         /** Chunk-lane search (`searchChunksOp`). */
         chunkLane: number;
         /** Cross-lane RRF fusion + provenance dedup after both lanes. */

package/dist/server/index.mjs

@@ -141,7 +141,7 @@ import {
   updateVaultMemoryOp,
   upsertMessageOp,
   vaultMemoryToStored
-} from "./chunk-I56324FR.mjs";
+} from "./chunk-BKB6AT7Z.mjs";
 import {
   createStreamingDeAnonymizer,
   notifyConsolidationFallback,

package/dist/server/{recall-T7CNXTQ3.mjs → recall-HVIEAH7J.mjs}

@@ -1,6 +1,6 @@
 import {
   recall
-} from "./chunk-I56324FR.mjs";
+} from "./chunk-BKB6AT7Z.mjs";
 import "./chunk-KF5OU4IN.mjs";
 import "./chunk-3RPGXXPZ.mjs";
 export {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@anuma/sdk",
-  "version": "6.10.0",
+  "version": "6.11.0",
   "description": "",
   "main": "./dist/index.cjs",
   "module": "./dist/index.mjs",