fix(brain): load vector index from disk on the search path

src/index/brain.rs

@@ -385,18 +385,56 @@ pub fn vector_index_needs_rebuild() -> bool {
         .unwrap_or(false)
 }
 
+/// Return a read guard over the vector index cache, loading the on-disk
+/// index once per process if needed (#545). VECTOR_CACHE used to be
+/// populated only by `embed_project` (i.e. `cora index`), so a fresh search
+/// process (`cora brain`, MCP tools) always saw an empty cache and the
+/// vector signal never fired. Load failures degrade gracefully: the cache
+/// stays `None` and search falls back to FTS + graph signals.
+///
+/// `RwLockWriteGuard::downgrade` is still unstable, so this drops the write
+/// guard and re-acquires a read lock; a concurrent reader may load twice,
+/// which is idempotent (same file, same result).
+fn ensure_vector_cache() -> std::sync::RwLockReadGuard<'static, Option<CodeVectorIndex>> {
+    {
+        let cache = VECTOR_CACHE.read().unwrap();
+        if cache.is_some() {
+            return cache;
+        }
+    }
+
+    let mut cache = VECTOR_CACHE.write().unwrap();
+    if cache.is_none() {
+        let dims = active_dims();
+        match CodeVectorIndex::load_or_create(&vector_index_path(), dims) {
+            Ok(vi) => *cache = Some(vi),
+            Err(e) => tracing::warn!("vector index load failed (search degrades to FTS): {e}"),
+        }
+    }
+    drop(cache);
+
+    VECTOR_CACHE.read().unwrap()
+}
+
 /// usearch vector search → (symbol_id, cosine_similarity) pairs, filtered to project.
 /// Uses cached vector index and cached project ID set — no disk I/O per query.
 fn vector_search(conn: &Connection, project_id: i64, query: &str, limit: usize) -> Vec<(i64, f32)> {
-    // Read-lock the cached vector index — no disk load
-    let cache = VECTOR_CACHE.read().unwrap();
+    // Read-lock the cached vector index — lazy-loads from disk on the first
+    // search in this process (#545).
+    let cache = ensure_vector_cache();
     let vi = match cache.as_ref() {
         Some(v) if !v.is_empty() => v,
         _ => return Vec::new(),
     };
 
     let vec = embed_code_dispatch(query);
 
+    // Dimension mismatch (e.g. embedding provider switched since the index
+    // was built) would panic inside the backend — degrade to other signals.
+    if vi.dims() != vec.len() {
+        return Vec::new();
+    }
+
     // Over-fetch to compensate for post-filter by project_id.
     let over_fetch = (limit * 5).max(50);
     let raw = vi.search(&vec, over_fetch);

tests/brain_search_process.rs

@@ -0,0 +1,88 @@
+//! Regression tests for #545: `cora brain` in a fresh process must load
+//! the on-disk vector index and emit the `vector` signal. Before the fix,
+//! `VECTOR_CACHE` was only populated by `embed_project` (`cora index`), so
+//! every search-only process silently degraded to FTS-only results.
+
+use assert_cmd::prelude::*;
+use std::path::PathBuf;
+use std::process::Command;
+
+fn cora_cmd() -> Command {
+    Command::cargo_bin("cora").unwrap()
+}
+
+/// Isolated CODECORA_HOME + tiny project, so tests never touch real data.
+fn sandbox(name: &str) -> (PathBuf, PathBuf) {
+    let root = std::env::temp_dir().join(format!("cora-545-{name}"));
+    let _ = std::fs::remove_dir_all(&root);
+    let proj = root.join("proj");
+    std::fs::create_dir_all(&proj).unwrap();
+    std::fs::write(
+        proj.join("Cargo.toml"),
+        "[package]\nname = \"p545\"\nversion = \"0.1.0\"\nedition = \"2021\"\n",
+    )
+    .unwrap();
+    std::fs::write(
+        proj.join("lib.rs"),
+        "pub fn alpha_widget() {}\npub fn beta_gadget() {}\n",
+    )
+    .unwrap();
+    let home = root.join("home");
+    std::fs::create_dir_all(&home).unwrap();
+    (proj, home)
+}
+
+fn run(cora: &mut Command) -> String {
+    let out = cora.assert().success().get_output().stdout.clone();
+    String::from_utf8(out).unwrap()
+}
+
+#[test]
+fn brain_fresh_process_emits_vector_signal() {
+    let (proj, home) = sandbox("usearch");
+
+    // 1. Index in one process (default usearch backend).
+    run(cora_cmd()
+        .args(["index"])
+        .current_dir(&proj)
+        .env("CODECORA_HOME", &home));
+
+    let db_dir = home.join("cora-code");
+    assert!(
+        db_dir.join("cora_index.usearch").exists(),
+        "usearch index file should exist after `cora index`"
+    );
+
+    // 2. Search in a FRESH process — vector signal must fire.
+    let out = run(cora_cmd()
+        .args(["brain", "alpha"])
+        .current_dir(&proj)
+        .env("CODECORA_HOME", &home));
+
+    assert!(
+        out.contains("vector"),
+        "fresh-process brain search must emit the vector signal, got:\n{out}"
+    );
+}
+
+#[test]
+fn brain_vecq_backend_uses_own_extension() {
+    let (proj, home) = sandbox("vecq");
+
+    std::fs::write(proj.join(".cora.yaml"), "brain:\n  vector_store: vecq\n").unwrap();
+
+    run(cora_cmd()
+        .args(["index"])
+        .current_dir(&proj)
+        .env("CODECORA_HOME", &home));
+
+    let db_dir = home.join("cora-code");
+    assert!(
+        db_dir.join("cora_index.vecq").exists(),
+        "vecq config must produce a .vecq index file (config wiring #543)"
+    );
+    assert!(
+        !db_dir.join("cora_index.usearch").exists(),
+        "vecq config must not create a usearch file"
+    );
+}