Add GitHub Actions CI/CD pipeline and test framework

- Add .github/workflows/build-test.yml for automated testing
- Add tests/ directory with TypeScript test runner
- Add docs/CICD.md documentation
- Remove .gitlab-ci.yml (migrated to GitHub Actions)
- Update .gitignore for test artifacts

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

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

tests/testcases/inference/TC-INFERENCE-001.yml

@@ -0,0 +1,30 @@
+id: TC-INFERENCE-001
+name: Model Pull
+suite: inference
+priority: 1
+timeout: 600000
+
+dependencies:
+  - TC-RUNTIME-003
+
+steps:
+  - name: Check if model exists
+    command: docker exec ollama37 ollama list | grep -q "gemma3:4b" && echo "Model exists" || echo "Model not found"
+
+  - name: Pull model if needed
+    command: docker exec ollama37 ollama list | grep -q "gemma3:4b" || docker exec ollama37 ollama pull gemma3:4b
+    timeout: 600000
+
+  - name: Verify model available
+    command: docker exec ollama37 ollama list
+
+criteria: |
+  The gemma3:4b model should be available for inference.
+
+  Expected:
+  - Model is either already present or successfully downloaded
+  - "ollama list" shows gemma3:4b in the output
+  - No download errors
+
+  Accept if model already exists (skip download).
+  Model size is ~3GB, download may take time.

tests/testcases/inference/TC-INFERENCE-002.yml

@@ -0,0 +1,28 @@
+id: TC-INFERENCE-002
+name: Basic Inference
+suite: inference
+priority: 2
+timeout: 180000
+
+dependencies:
+  - TC-INFERENCE-001
+
+steps:
+  - name: Run simple math question
+    command: docker exec ollama37 ollama run gemma3:4b "What is 2+2? Answer with just the number." 2>&1
+    timeout: 120000
+
+  - name: Check GPU memory usage
+    command: docker exec ollama37 nvidia-smi --query-compute-apps=pid,used_memory --format=csv 2>/dev/null || echo "No GPU processes"
+
+criteria: |
+  Basic inference should work on Tesla K80.
+
+  Expected:
+  - Model responds to the math question
+  - Response should indicate "4" (accept variations: "4", "four", "The answer is 4", etc.)
+  - GPU memory should be allocated during inference
+  - No CUDA errors in output
+
+  This is AI-generated output - accept reasonable variations.
+  Focus on the model producing a coherent response.

tests/testcases/inference/TC-INFERENCE-003.yml

@@ -0,0 +1,34 @@
+id: TC-INFERENCE-003
+name: API Endpoint Test
+suite: inference
+priority: 3
+timeout: 120000
+
+dependencies:
+  - TC-INFERENCE-001
+
+steps:
+  - name: Test generate endpoint (non-streaming)
+    command: |
+      curl -s http://localhost:11434/api/generate \
+        -d '{"model":"gemma3:4b","prompt":"Say hello in one word","stream":false}' \
+        | head -c 500
+
+  - name: Test generate endpoint (streaming)
+    command: |
+      curl -s http://localhost:11434/api/generate \
+        -d '{"model":"gemma3:4b","prompt":"Count from 1 to 3","stream":true}' \
+        | head -5
+
+criteria: |
+  Ollama REST API should handle inference requests.
+
+  Expected for non-streaming:
+  - Returns JSON with "response" field
+  - Response contains some greeting (hello, hi, etc.)
+
+  Expected for streaming:
+  - Returns multiple JSON lines
+  - Each line contains partial response
+
+  Accept any valid JSON response. Content may vary.

tests/testcases/inference/TC-INFERENCE-004.yml

@@ -0,0 +1,32 @@
+id: TC-INFERENCE-004
+name: CUBLAS Fallback Verification
+suite: inference
+priority: 4
+timeout: 120000
+
+dependencies:
+  - TC-INFERENCE-002
+
+steps:
+  - name: Check for CUBLAS errors in logs
+    command: cd docker && docker compose logs 2>&1 | grep -i "CUBLAS_STATUS" | grep -v "SUCCESS" | head -10 || echo "No CUBLAS errors"
+
+  - name: Check compute capability detection
+    command: cd docker && docker compose logs 2>&1 | grep -iE "compute|capability|cc.*3" | head -10 || echo "No compute capability logs"
+
+  - name: Verify no GPU errors
+    command: cd docker && docker compose logs 2>&1 | grep -iE "error|fail" | grep -i gpu | head -10 || echo "No GPU errors"
+
+criteria: |
+  CUBLAS should work correctly on Tesla K80 using legacy fallback.
+
+  Expected:
+  - No CUBLAS_STATUS_ARCH_MISMATCH errors
+  - No CUBLAS_STATUS_NOT_SUPPORTED errors
+  - Compute capability 3.7 may be mentioned in debug logs
+  - No fatal GPU-related errors
+
+  The K80 uses legacy CUBLAS functions (cublasSgemmBatched)
+  instead of modern Ex variants. This should work transparently.
+
+  Accept warnings. Only fail on actual CUBLAS errors.