Fix Metal backend crash by properly disposing llama resources

The Metal crash was caused by not disposing resources in the right order
at the right time. The fix:

1. Restore proper dispose() that disposes contexts → models → llama in order
2. Move disposeDefaultLlamaCpp() to global afterAll (after all tests complete)
3. Keep disposed flag to prevent double-dispose

The issue was that disposing per-suite broke tests that share llama,
and not disposing at all left orphaned Metal resources at process exit.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

src/eval.test.ts

@@ -34,7 +34,7 @@ import {
   DEFAULT_EMBED_MODEL,
   type RankedResult,
 } from "./store";
-import { getDefaultLlamaCpp, formatDocForEmbedding } from "./llm";
+import { getDefaultLlamaCpp, disposeDefaultLlamaCpp, formatDocForEmbedding } from "./llm";
 
 // Eval queries with expected documents
 const evalQueries: {
@@ -201,8 +201,8 @@ describe("Vector Search", () => {
     hasEmbeddings = true;
   }, 120000); // 2 minute timeout for embedding generation
 
-  // Note: Don't call disposeDefaultLlamaCpp() here - it causes Metal backend
-  // assertion failures during process exit. Let the process exit handle cleanup.
+  // Note: Don't dispose here - Hybrid tests also use llama.
+  // Dispose happens in the global afterAll.
 
   test("easy queries: ≥60% Hit@3 (vector should match keywords too)", async () => {
     if (!hasEmbeddings) return; // Skip if embedding failed
@@ -392,6 +392,8 @@ describe("Hybrid Search (RRF)", () => {
 // Cleanup
 // =============================================================================
 
-afterAll(() => {
+afterAll(async () => {
+  // Dispose llama before process exit to properly free Metal resources
+  await disposeDefaultLlamaCpp();
   rmSync(tempDir, { recursive: true, force: true });
 });

src/llm.ts

@@ -704,19 +704,38 @@ Generate the structured expansion:`;
       this.inactivityTimer = null;
     }
 
-    // Don't explicitly dispose llama resources - it causes Metal backend
-    // assertion failures during process cleanup. The Metal device cleanup
-    // in ggml-metal expects resources to be freed in a specific order that
-    // we can't control. Just clear references and let the process exit
-    // handle cleanup naturally.
-    // See: https://github.com/ggml-org/llama.cpp/pull/17869
-    this.embedContext = null;
-    this.generateContext = null;
-    this.rerankContext = null;
-    this.embedModel = null;
-    this.generateModel = null;
-    this.rerankModel = null;
-    this.llama = null;
+    // Dispose in order: contexts -> models -> llama
+    // Contexts depend on models, models depend on llama
+    if (this.embedContext) {
+      await this.embedContext.dispose();
+      this.embedContext = null;
+    }
+    if (this.generateContext) {
+      await this.generateContext.dispose();
+      this.generateContext = null;
+    }
+    if (this.rerankContext) {
+      await this.rerankContext.dispose();
+      this.rerankContext = null;
+    }
+
+    if (this.embedModel) {
+      await this.embedModel.dispose();
+      this.embedModel = null;
+    }
+    if (this.generateModel) {
+      await this.generateModel.dispose();
+      this.generateModel = null;
+    }
+    if (this.rerankModel) {
+      await this.rerankModel.dispose();
+      this.rerankModel = null;
+    }
+
+    if (this.llama) {
+      await this.llama.dispose();
+      this.llama = null;
+    }
   }
 }