perf: audit EmbeddingGemma 2 and reuse measured query embeddings

apps/desktop/src/main.ts

@@ -11,6 +11,8 @@ import {
   AutomaticIndexGate,
   emptySearchState,
   LatestLane,
+  semanticAvailable,
+  SEMANTIC_SETTLE_MS,
 } from "./search-state";
 import type { SemanticPhase } from "./search-state";
 import { fileIcon, icon } from "./icons";
@@ -272,12 +274,7 @@ function refreshEmptySearchState() {
 }
 
 function canSearchSemantically() {
-  return Boolean(
-    semanticStatus?.ready &&
-      semanticStatus.generation !== null &&
-      !semanticStatus.embedding &&
-      !semanticJobActive,
-  );
+  return semanticAvailable(semanticStatus);
 }
 
 function queueSemanticRefresh(delay = 1000) {
@@ -954,7 +951,7 @@ function refreshPerformance(request: SearchRequest) {
   );
   for (const [lane, delay] of Object.entries(report.lane_schedule_ms))
     append(
-      `${lane}: scheduled ${delay.toFixed(1)} ms after debounce${lane === "hybrid" ? " (includes intentional 350 ms semantic delay or waiting for readiness)" : ""}`,
+      `${lane}: scheduled ${delay.toFixed(1)} ms after debounce${lane === "hybrid" ? ` (includes intentional ${SEMANTIC_SETTLE_MS} ms semantic delay or waiting for readiness)` : ""}`,
     );
   for (const attempt of report.attempts) {
     append(
@@ -1123,7 +1120,7 @@ async function search(inputAt = performance.now()) {
     hybridTimer = setTimeout(() => {
       if (isCurrent(request) && canSearchSemantically())
         submitSearch(request, "hybrid");
-    }, 350);
+    }, SEMANTIC_SETTLE_MS);
   } catch (error) {
     showError(error);
   }

apps/desktop/src/search-state.ts

@@ -1,3 +1,14 @@
+/** Let rapid typing settle before starting the model lane. */
+export const SEMANTIC_SETTLE_MS = 120;
+
+/** A rebuild retains the previous committed generation. Backend readiness is
+ * withdrawn only while a replacement query graph is being prepared. */
+export function semanticAvailable(
+  status: { ready: boolean; generation: number | null } | null,
+) {
+  return Boolean(status?.ready && status.generation !== null);
+}
+
 /** One active request per lane; intermediate keystrokes collapse to the latest.
  * Native work already running cannot be cancelled, so revisions guard every
  * result and retry. Busy responses receive at most one delayed retry. */

apps/desktop/tests/search-state.test.mjs

@@ -1,6 +1,6 @@
 import test from 'node:test';
 import assert from 'node:assert/strict';
-import { LatestLane, AutomaticIndexGate, emptySearchState } from '../src/search-state.ts';
+import { LatestLane, AutomaticIndexGate, emptySearchState, semanticAvailable, SEMANTIC_SETTLE_MS } from '../src/search-state.ts';
 const settle = () => new Promise(resolve => setImmediate(resolve));
 
 test('coalesces queued keystrokes and discards stale responses', async () => {
@@ -67,3 +67,14 @@ test('no document content and unavailable semantics have actionable distinct gui
   assert.match(emptySearchState('lexical', 'no-content').detail, /readable documents/);
   assert.match(emptySearchState('lexical', 'unavailable').detail, /retry the search/);
 });
+
+
+test('old committed semantic generation stays available during rebuild, activation readiness gates replacement', () => {
+  assert.equal(semanticAvailable(null), false);
+  assert.equal(semanticAvailable({ready: true, generation: null, embedding: true}), false);
+  assert.equal(semanticAvailable({ready: true, generation: 1, embedding: false}), true);
+  assert.equal(semanticAvailable({ready: true, generation: 1, embedding: true}), true);
+  assert.equal(semanticAvailable({ready: false, generation: 2, embedding: true}), false);
+  assert.equal(semanticAvailable({ready: true, generation: 2, embedding: false}), true);
+  assert.equal(SEMANTIC_SETTLE_MS, 120);
+});