Fix macOS integration test: copy mlx.metallib next to binary

The MLX runtime looks for mlx.metallib in the same directory as the
executable. Without it, Metal GPU operations crash with exit code 255.
Also update aarch64 and macOS server tests to match the verbose x86_64
test script (health check, JSON, text, verbose_json responses).

Signed-off-by: Michael Yuan <michael@secondstate.io>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: juntao

.github/workflows/integration-test.yml

@@ -280,21 +280,67 @@ jobs:
           echo "Transcript: $result"
           echo "$result" | grep -qi "fox\|lazy\|dog"
 
-      - name: Server — health + transcription
+      - name: Server — start, health check, transcription, stop
         if: steps.sve.outputs.available == 'true'
         run: |
+          # Start server in background
           ./target/release/transcribe-server \
-            --model-dir "$MODEL_DIR" --port 18080 &
+            --model-dir "$MODEL_DIR" \
+            --port 18080 \
+            --verbose &
           SERVER_PID=$!
+          echo "Server PID: $SERVER_PID"
+
+          # Wait for server ready (up to 120 s — model loading takes time)
           for i in $(seq 1 120); do
-            curl -sf http://localhost:18080/health > /dev/null 2>&1 && break; sleep 1
+            if curl -sf http://localhost:18080/health > /dev/null 2>&1; then
+              echo "Server ready after ${i}s"; break
+            fi
+            sleep 1
           done
+
+          # Health endpoint
           curl -sf http://localhost:18080/health | grep -q '"ok"'
+          echo "Health OK"
+
+          # JSON response
+          json_resp=$(curl -sf \
+            -X POST http://localhost:18080/v1/audio/transcriptions \
+            -F "file=@tests/fixtures/sample2.wav;type=audio/wav" \
+            -F "model=cohere-transcribe" \
+            -F "language=en" \
+            -F "response_format=json")
+          echo "JSON: $json_resp"
+          echo "$json_resp" | python3 -c "
+          import sys,json
+          d=json.load(sys.stdin)
+          assert 'text' in d, 'Missing text'
+          print('text:', d['text'])
+          "
+
+          # Text response
           curl -sf \
             -X POST http://localhost:18080/v1/audio/transcriptions \
             -F "file=@tests/fixtures/sample2.wav;type=audio/wav" \
-            -F "model=cohere-transcribe" | python3 -c "import sys,json; assert 'text' in json.load(sys.stdin)"
+            -F "model=cohere-transcribe" \
+            -F "response_format=text"
+          echo
+
+          # verbose_json response
+          curl -sf \
+            -X POST http://localhost:18080/v1/audio/transcriptions \
+            -F "file=@tests/fixtures/sample2.wav;type=audio/wav" \
+            -F "model=cohere-transcribe" \
+            -F "response_format=verbose_json" | python3 -c "
+          import sys,json
+          d=json.load(sys.stdin)
+          assert d['task']=='transcribe'
+          assert 'text' in d and 'duration' in d and 'segments' in d
+          print('verbose_json OK — duration:', d['duration'])
+          "
+
           kill $SERVER_PID
+          echo "All server integration tests passed"
 
   # ─────────────────────────────────────────────────────────────────────────────
   # Integration test — macOS Apple Silicon, mlx backend
@@ -340,6 +386,12 @@ jobs:
         env:
           MACOSX_DEPLOYMENT_TARGET: "14.0"
 
+      - name: Copy mlx.metallib next to binaries
+        run: |
+          # MLX runtime looks for mlx.metallib in the same directory as the binary
+          find target/release/build -name "mlx.metallib" -exec cp {} target/release/ \;
+          ls -lh target/release/mlx.metallib
+
       - name: CLI — transcribe sample2.wav
         run: |
           result=$(./target/release/transcribe \
@@ -349,17 +401,63 @@ jobs:
           echo "Transcript: $result"
           echo "$result" | grep -qi "fox\|lazy\|dog"
 
-      - name: Server — health + transcription
+      - name: Server — start, health check, transcription, stop
         run: |
+          # Start server in background
           ./target/release/transcribe-server \
-            --model-dir "$MODEL_DIR" --port 18080 &
+            --model-dir "$MODEL_DIR" \
+            --port 18080 \
+            --verbose &
           SERVER_PID=$!
+          echo "Server PID: $SERVER_PID"
+
+          # Wait for server ready (up to 120 s — model loading takes time)
           for i in $(seq 1 120); do
-            curl -sf http://localhost:18080/health > /dev/null 2>&1 && break; sleep 1
+            if curl -sf http://localhost:18080/health > /dev/null 2>&1; then
+              echo "Server ready after ${i}s"; break
+            fi
+            sleep 1
           done
+
+          # Health endpoint
           curl -sf http://localhost:18080/health | grep -q '"ok"'
+          echo "Health OK"
+
+          # JSON response
+          json_resp=$(curl -sf \
+            -X POST http://localhost:18080/v1/audio/transcriptions \
+            -F "file=@tests/fixtures/sample2.wav;type=audio/wav" \
+            -F "model=cohere-transcribe" \
+            -F "language=en" \
+            -F "response_format=json")
+          echo "JSON: $json_resp"
+          echo "$json_resp" | python3 -c "
+          import sys,json
+          d=json.load(sys.stdin)
+          assert 'text' in d, 'Missing text'
+          print('text:', d['text'])
+          "
+
+          # Text response
           curl -sf \
             -X POST http://localhost:18080/v1/audio/transcriptions \
             -F "file=@tests/fixtures/sample2.wav;type=audio/wav" \
-            -F "model=cohere-transcribe" | python3 -c "import sys,json; assert 'text' in json.load(sys.stdin)"
+            -F "model=cohere-transcribe" \
+            -F "response_format=text"
+          echo
+
+          # verbose_json response
+          curl -sf \
+            -X POST http://localhost:18080/v1/audio/transcriptions \
+            -F "file=@tests/fixtures/sample2.wav;type=audio/wav" \
+            -F "model=cohere-transcribe" \
+            -F "response_format=verbose_json" | python3 -c "
+          import sys,json
+          d=json.load(sys.stdin)
+          assert d['task']=='transcribe'
+          assert 'text' in d and 'duration' in d and 'segments' in d
+          print('verbose_json OK — duration:', d['duration'])
+          "
+
           kill $SERVER_PID
+          echo "All server integration tests passed"