Ship floating launcher, automatic discovery and bounded media search improvements

tools/desktop-capture/README.md

@@ -22,6 +22,7 @@ then warms its retained CPU worker before listening on 127.0.0.1:1431.
 
 ```powershell
 $env:SEARCHI_FORCE_CPU = '1'
+$env:SEARCHI_EXTRACT_WORKER = (Resolve-Path target/release/searchi-extract-worker.exe).Path
 cargo run --manifest-path tools/desktop-capture/backend/Cargo.toml -- --root .searchi/readme-capture/files --db .searchi/readme-capture/index.sqlite --models .searchi/models/embeddinggemma2 --worker .searchi/runtime/cpu/Release/searchi-gemma-worker.exe --dist apps/desktop/dist --ffmpeg .searchi/media-runtime/bin/ffmpeg.exe --ffprobe .searchi/media-runtime/bin/ffprobe.exe
 ```
 
@@ -46,12 +47,14 @@ loaded fonts, image pixels, and genuine worker inference counters in debug mode.
 
 ## Artifacts and limits
 
-Outputs under docs/screenshots are search-light.png, search-dark.png,
+Outputs under docs/screenshots are launcher-empty.png, search-light.png, search-dark.png,
 media-search.png and performance-details.png, plus capture-manifest.json. The
 manifest contains actual fixture queries, counts/selection, measured response
 metrics, device/model, viewport, PNG SHA256, runner/browser versions, Git base
 commit and dirty-state flag. The base commit does not identify uncommitted
-source exactly. The ephemeral API token is never saved in the manifest.
+source exactly. It also records the actual frontend package version and built
+index SHA-256. The compact empty-bar capture has no manufactured results and
+uses the genuine ready backend. The ephemeral API token is never saved in the manifest.
 
 Captures are **browser-rendered production UI with the actual CPU engine**.
 They omit native window chrome. They do not establish native WebView visual QA,

tools/desktop-capture/backend/Cargo.lock

@@ -1129,11 +1129,16 @@ name = "searchi-core"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "libc",
  "lopdf",
  "quick-xml",
  "rusqlite",
+ "searchi-media",
  "serde",
+ "serde_json",
+ "sha2 0.10.9",
  "walkdir",
+ "windows-sys",
  "zip",
 ]
 
@@ -1192,11 +1197,22 @@ dependencies = [
  "searchi-core",
  "searchi-media",
  "searchi-semantic",
+ "searchi-speech",
  "serde",
  "sha2 0.10.9",
  "usearch",
 ]
 
+[[package]]
+name = "searchi-speech"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "serde",
+ "serde_json",
+ "sha2 0.10.9",
+]
+
 [[package]]
 name = "serde"
 version = "1.0.229"

tools/desktop-capture/backend/src/main.rs

@@ -44,6 +44,8 @@ struct Filters {
     max_duration_ms: Option<u32>,
     modified_after_unix_ms: Option<i64>,
     modified_before_unix_ms: Option<i64>,
+    media_artist: Option<String>,
+    media_title: Option<String>,
 }
 impl Filters {
     fn options(self, root: &Path) -> Result<SearchOptions> {
@@ -63,6 +65,8 @@ impl Filters {
             max_duration_ms: self.max_duration_ms,
             modified_after_unix_ms: self.modified_after_unix_ms,
             modified_before_unix_ms: self.modified_before_unix_ms,
+            media_artist: self.media_artist,
+            media_title: self.media_title,
         })
     }
 }
@@ -87,6 +91,10 @@ fn main() -> Result<()> {
     let mut lexical = SearchIndex::open(&args.db)?;
     validate_database(&args.db, &root)?;
     lexical.index_folder(&root)?;
+    ensure!(
+        lexical.extracted_source_hash(&root.join("Supplier Agreement.pdf").to_string_lossy())?.is_some(),
+        "Capture PDF was not genuinely extracted; set SEARCHI_EXTRACT_WORKER to the verified document helper"
+    );
     lexical.prewarm()?;
     let tools = FfmpegTools::new(args.ffmpeg, args.ffprobe)?;
     let mut hybrid = HybridSearch::open(&args.db, &args.models, &args.worker, 1)?;
@@ -287,7 +295,7 @@ impl App {
                     json!({"model":"Google EmbeddingGemma 2","ready":true,"model_directory":self.models,"worker_ready":true,"downloading":false,"embedding":false,"device_policy":"cpu","error":null,"downloaded_bytes":864676480u64,"total_bytes":864676480u64,"generation":status.generation,"chunks":status.chunks,"stale":status.stale,"pending":status.pending,"warming":false,"worker_info":self.hybrid.worker_info(),"media_tools_ready":true,"media_setup_error":null,"media_failures":status.failures}),
                 )
             }
-            "search_measured" | "hybrid_search_measured" => {
+            "search_measured" | "hybrid_search_measured" | "semantic_search_raw" => {
                 let query = args["query"].as_str().context("Query required")?;
                 ensure!(query.len() <= 4096, "Query too large");
                 let filters: Filters = if args["filters"].is_null() {
@@ -297,7 +305,9 @@ impl App {
                 };
                 let options = filters.options(&self.root)?;
                 let debug = args["debugPerformance"].as_bool().unwrap_or(false);
-                if name == "search_measured" {
+                if name == "semantic_search_raw" {
+                    Ok(serde_json::to_value(self.hybrid.search_semantic(query, 50, &options)?)?)
+                } else if name == "search_measured" {
                     Ok(serde_json::to_value(
                         self.lexical
                             .search_with_options_measured(query, 50, &options, debug)?,

tools/desktop-capture/capture.mjs

@@ -92,6 +92,28 @@ try {
       .evaluate((image) => image.complete && image.naturalWidth > 0),
     "Brand asset failed to load",
   );
+  // Capture the actual ready/empty production layout before typing. This is
+  // browser-rendered UI, not native window chrome or synthesized app results.
+  await page.waitForFunction(() =>
+    document.documentElement.dataset.launcherLayout === "compact" &&
+    document.querySelector("#query")?.value === "",
+  );
+  const compactPath = resolve(out, "launcher-empty.png");
+  const compactBounds = await page.locator("#app").boundingBox();
+  assert(compactBounds && compactBounds.width <= 720 && compactBounds.height <= 112,
+    "Ready empty launcher must remain a compact bar");
+  await page.locator("#app").screenshot({ path: compactPath, animations: "disabled" });
+  evidence.push({
+    file: "launcher-empty.png",
+    sha256: createHash("sha256").update(await readFile(compactPath)).digest("hex"),
+    query: "",
+    mode: "empty-ready",
+    result_count: 0,
+    selected_filename: null,
+    theme: "light",
+    capture_bounds: compactBounds,
+    state_assertions: { semantic_ready: true, compact_empty_layout: true, native_pixels: false },
+  });
   async function query(text, debug = false) {
     const awaited = page.waitForResponse(
       (response) => {
@@ -151,6 +173,10 @@ try {
       () => !document.querySelector("#preview")?.hidden,
     );
     if (text === "cancel the supplier partnership") {
+      assert(
+        !body.results.some((result) => ["Mountain Lake.jpg", "Night Pier.jpg"].includes(result.filename)),
+        "Contract results must not contain the unrelated fixture photos",
+      );
       assert.equal(
         latestMeasured.first_filename,
         "Supplier Agreement.pdf",
@@ -225,9 +251,12 @@ try {
         literal_query_marks: true,
         actual_page_explanation: true,
         why_disclosure_collapsed_by_default: true,
+        unrelated_photos_removed: true,
       };
     }
     if (text === "mountains reflected in calm water") {
+      assert(body.results.every((result) => result.match_kind === "semantic_image"),
+        "Visual search shortlist must not contain the unrelated fixture documents");
       assert.equal(
         latestMeasured.first_filename,
         "Mountain Lake.jpg",
@@ -238,7 +267,7 @@ try {
         0,
         "Meaning-only photo result must not invent word-level alignment",
       );
-      latestMeasured.ui_checks = { meaning_only_image_has_no_text_marks: true };
+      latestMeasured.ui_checks = { meaning_only_image_has_no_text_marks: true, unrelated_documents_removed: true };
     }
     await page.evaluate(
       () =>
@@ -249,8 +278,12 @@ try {
   }
   async function screenshot(name, queryText) {
     assert.deepEqual(failures, [], "Frontend page errors occurred");
+    const pane = await page.locator(".results-pane").boundingBox();
+    const preview = await page.locator("#preview").boundingBox();
+    assert(pane && preview && preview.x >= pane.x + pane.width - 1,
+      "Wide result and preview panes must remain side by side");
     const path = resolve(out, `${name}.png`);
-    await page.screenshot({ path, animations: "disabled", fullPage: false });
+    await page.locator("#app").screenshot({ path, animations: "disabled" });
     const selected = await page.locator("#preview-title").innerText();
     assert.equal(
       selected,
@@ -272,7 +305,9 @@ try {
         matchMedia("(prefers-color-scheme: dark)").matches ? "dark" : "light",
       ),
       summary: await page.locator("#result-summary").innerText(),
+      capture_bounds: await page.locator("#app").boundingBox(),
       state_assertions: {
+        wide_split_preview: true,
         semantic_ready: true,
         hybrid_final: true,
         image_visible: name === "media-search",
@@ -366,6 +401,10 @@ try {
       {
         kind: "browser-rendered-production-ui-real-cpu-engine",
         source,
+        frontend_build: {
+          version: JSON.parse(await readFile(resolve(repository, "apps/desktop/package.json"), "utf8")).version,
+          index_sha256: createHash("sha256").update(await readFile(resolve(repository, "apps/desktop/dist/index.html"))).digest("hex"),
+        },
         viewport: { width: 1440, height: 1000 },
         runner: { playwright: "1.64.0", chromium: browser.version() },
         backend: {

tools/package_windows.py

@@ -203,7 +203,8 @@ def package(root, executable, output, cli=None):
                 "Windows needs WebView2. The Microsoft C++ runtime ships with these helpers; Visual Studio is not required.\n"
                 "First launch downloads approximately 865 MB of verified Google model files.\n"
                 "Local spoken-passage recognition adds approximately 149 MB of verified speech models.\n"
-                "Choose Add folder to index your own files; ordinary search stays local.\n"
+                "Use Ctrl+Shift+Space for the floating search bar. Local file discovery runs automatically.\n"
+                "Change automatic indexing or add a folder in Settings; ordinary search stays local.\n"
                 "This is an unsigned development build, not a certified production release.\n")
             if cli is not None:
                 archive.writestr("AGENT-SETUP.txt",

tools/relevance-eval/policies/hybrid-tail-v1.json

@@ -0,0 +1,12 @@
+{
+  "policy_id": "hybrid-semantic-tail-gap070-v1",
+  "scope": "hybrid semantic candidates before reciprocal-rank fusion",
+  "minimum_adjacent_gap_exclusive": 0.07,
+  "derivation": "Rounded upward from maximum adjacent gap 0.06504970788955688 among eight calibration All-files absence requests",
+  "calibration_report_sha256": "56f42c523527dd748918bb8f17be6bd7818158227e692fccaf387d915f9d0a94",
+  "boundary": "largest adjacent raw cosine gap in existing filtered file-deduplicated candidate list",
+  "equal_largest_gaps": "retain all; numeric tie tolerance is twice f32 EPSILON",
+  "preserve": ["strongest semantic candidate", "all independent lexical candidates", "exact filename and identifier priority", "raw semantic diagnostic API"],
+  "claims": ["prunes clearly separated weak semantic tails"],
+  "not_claimed": ["no-answer detection", "calibrated confidence", "universal embedding threshold", "new held-out or human quality certification"]
+}