Enforce warning-free tests and add hardware reporting

.github/ISSUE_TEMPLATE/hardware-validation.yml

@@ -0,0 +1,69 @@
+name: Hardware validation report
+description: Report native Intel GPU results and the checks still needed.
+title: "[Hardware] GPU / OS / backend"
+labels: ["help wanted", "untested"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Help record which Intel GPU environments have been tested. Include failures as well as successes, and mark checks you have not performed as "not run".
+        See the [validation checklist](https://github.com/offbyonebit/arc-llama/issues/71). Remove tokens and private details before attaching logs.
+  - type: input
+    id: hardware
+    attributes:
+      label: GPU and memory
+      description: Exact GPU model, dedicated/shared memory, and number of GPUs.
+      placeholder: Arc A770 16 GiB, one GPU
+    validations:
+      required: true
+  - type: textarea
+    id: environment
+    attributes:
+      label: Software environment
+      description: OS/build, kernel (Linux), Intel graphics driver, Arc Llama version, Python version, llama.cpp runtime build/tag, and oneAPI version if using SYCL.
+    validations:
+      required: true
+  - type: dropdown
+    id: backend
+    attributes:
+      label: Backend tested
+      options:
+        - Vulkan
+        - SYCL
+        - Both (report results separately)
+        - Not run
+    validations:
+      required: true
+  - type: textarea
+    id: diagnostics
+    attributes:
+      label: Diagnostic output
+      description: Paste output from arc-llama --version, arc-llama gpus, and arc-llama doctor. These commands do not run inference. State any command you did not run.
+      render: text
+    validations:
+      required: true
+  - type: textarea
+    id: configuration
+    attributes:
+      label: Model and configuration
+      description: Model repository/file, quantization, recipe, context size, GPU selection/offload, and any changed settings. If inference was not run, say so.
+    validations:
+      required: true
+  - type: textarea
+    id: results
+    attributes:
+      label: Observed results
+      description: Record cold start, model switching, streaming, non-streaming, and shutdown as pass/fail/not run. Include symptoms, timing, or throughput if measured. For multiple backends, distinguish each result.
+      value: |
+        Cold start:
+        Model switching:
+        Streaming:
+        Non-streaming:
+        Shutdown:
+    validations:
+      required: true
+  - type: textarea
+    id: logs
+    attributes:
+      label: Logs and remaining problems
+      description: Relevant errors, reproducible steps, or logs. Include remaining checks and unexpected GPU/memory behavior.

.github/workflows/ci.yml

@@ -46,7 +46,7 @@ jobs:
       - name: Test with pytest
         # Print every skip reason so an environment change cannot hide missing
         # coverage in an otherwise green job.
-        run: pytest tests/ -q -ra --tb=short -p no:cacheprovider
+        run: pytest tests/ -W error -q -ra --tb=short -p no:cacheprovider
 
   browser-tests:
     # Real-browser regression suite for the bundled web UI. Kept separate

README.md

@@ -164,6 +164,23 @@ curl http://127.0.0.1:11437/v1/chat/completions \
   }'
 ```
 
+## Tested hardware
+
+The recorded native validation below used the **0.9.0rc2 candidate** and the
+models and recipes in the linked report. It does not establish validation for
+every driver, model, or later release.
+
+| GPU | Platform | Backend | Recorded validation |
+| --- | --- | --- | --- |
+| Arc Pro B60 (24 GiB) | Linux | Vulkan and SYCL | [Native release validation](docs/release-validation-linux-2026-10-03.md) |
+| Alchemist (A-series) | Linux / Windows | Vulkan / SYCL | Hardware reports [requested](https://github.com/offbyonebit/arc-llama/issues/71) |
+| Lunar Lake | Linux / Windows | Vulkan / SYCL | Hardware reports [requested](https://github.com/offbyonebit/arc-llama/issues/71) |
+| Arc Pro B60 | Windows | Vulkan / SYCL | Native Windows validation [requested](https://github.com/offbyonebit/arc-llama/issues/69) |
+
+Have one of these devices? [Submit a hardware validation report](https://github.com/offbyonebit/arc-llama/issues/new?template=hardware-validation.yml)
+with your environment, configuration, and observed results. Reports can include
+checks you have not run; mark them as “not run” so the remaining coverage is clear.
+
 ## Requirements
 
 ### Linux

pyproject.toml

@@ -44,6 +44,7 @@ mcp = [
     "mcp>=1.0,<2",
 ]
 dev = [
+    "httpx2>=2.13.1",
     "pytest>=8.0",
     "pytest-asyncio>=0.23",
     "ruff>=0.5",

tests/test_launcher.py

@@ -319,9 +319,12 @@ def poll(self):
         try:
             srv.start(log_dir=log_dir)
             assert log_dir.exists()
+            assert srv.is_running is True
         finally:
             subprocess.Popen = original_popen
-        assert srv.is_running is True
+            # Discard the mock child before closing real log/lock handles.
+            srv.process = None
+            srv.stop()
 
     @pytest.mark.asyncio
     async def test_wait_ready_true_when_healthy(self, monkeypatch: pytest.MonkeyPatch):
@@ -573,12 +576,15 @@ def poll(self):
         subprocess.Popen = _fake_popen
         try:
             srv.start(log_dir=log_dir)
+            assert called["kwargs"]["creationflags"] == getattr(
+                subprocess, "CREATE_NEW_PROCESS_GROUP", 0
+            )
+            assert "preexec_fn" not in called["kwargs"]
         finally:
             subprocess.Popen = original_popen
-        assert called["kwargs"]["creationflags"] == getattr(
-            subprocess, "CREATE_NEW_PROCESS_GROUP", 0
-        )
-        assert "preexec_fn" not in called["kwargs"]
+            # Discard the mock child before closing real log/lock handles.
+            srv.process = None
+            srv.stop()
 
     def test_stop_sends_ctrl_break_then_force_kills_tree_on_timeout(self, monkeypatch):
         from arc_llama import launcher as launcher_mod