test(formatters): query_clarification times out on CPU during full test run

[planetf1]

Summary

test/formatters/granite/test_intrinsics_formatters.py::test_run_transformers[query_clarification] hangs for 180+ seconds during torch.nn.Linear.forward (stuck in F.linear()) when running on CPU/MPS without GPU.

Context

Discovered during full test run for #1294 (aLoRA input_ids fix). This is a separate pre-existing issue.

Details

• Model: ibm-granite/granite-4.1-3b
• Context: 4 conversation turns + 12 retrieved documents (~2000+ tokens)
• Generation: max_completion_tokens: 512
• Hardware: Apple Silicon (MPS), no CUDA

This is a performance/resource limitation, not a code bug. The test is already gated on CI via gh_run == 1 xfail, but it blocks local CI-aware runs.

Proposed Fix

Add a @pytest.mark.slow marker or a CPU-specific skip to prevent it from blocking local test runs, or increase the timeout for this specific test case.

[planetf1]

Analysis

The test works — the xfail is not a correctness issue. The existing gh_run == 1 xfail is an intentional, well-documented CI time-budget workaround (the test file comment explains it: CI has a 15-minute hard budget for 1900+ tests, any test over 5s must xfail or the whole run fails). The test completes correctly on GPU.

The gap is local CPU/MPS runs, where the test hangs for 180+ seconds because granite-4.1-3b is too slow for CPU inference on a large context. The @pytest.mark.slow marker excludes tests from the default uv run pytest invocation (via addopts: -m "not slow" in pyproject.toml) and would fix this.

Fix

Add @pytest.mark.slow to test_run_transformers for the query_clarification case, or at the fixture/parametrize level if that's cleaner. One-line change.

Broader context

• The gh_run xfail pattern is systemic across test_intrinsics_formatters.py — this isn't unique to query_clarification
• No overnight/nightly CI run currently exists. test: on-demand nightly test runs for PRs #734 tracks on-demand nightly runs for PRs — slow/GPU tests would naturally move there once that lands
• Once test: on-demand nightly test runs for PRs #734 exists, query_clarification could run properly on GPU hardware in CI rather than being permanently excluded