Conftest GPU auto-skip incorrectly skips CPU xpress tests

[FBumann]

Problem

test/conftest.py::pytest_collection_modifyitems auto-marks any parametrized test where the solver param is a backend with SolverFeature.GPU_ACCELERATION as a GPU test and skips it without --run-gpu.

Xpress carries GPU_ACCELERATION=True because it has an optional GPU pathway, but the vast majority of xpress test runs are pure CPU and don't need a GPU. The current rule treats CPU xpress and GPU xpress as the same thing, so any parametrize("solver", [..., "xpress", ...]) silently drops the xpress cell from the default test run.

Reproduction

pytest test/test_sos_masked.py -k xpress -rs
# every xpress case → SKIPPED [reason: need --run-gpu option to run GPU tests]

pytest test/test_sos_masked.py -k xpress --run-gpu
# every xpress case → PASSED

Same pattern hits test_piecewise_constraints.py::test_sos2_per_entity_nan_padding[xpress-*] and any other suite that parametrizes on solver name.

Why this matters

Xpress's SOS path was a regression target in #688 — xpress / direct and xpress / lp are exactly the cells we want covered by default. Today they're hidden behind --run-gpu, which CI does not pass, so a real regression in xpress would slip through.

Suggested fix

Distinguish "GPU-only solver" from "solver with an optional GPU mode":

• cuPDLPx (GPU-only) → keep the skip
• xpress → don't skip; xpress GPU acceleration is opt-in at solve time via solver options, not at import/parametrize time

A small option here: replace solver_supports(solver, SolverFeature.GPU_ACCELERATION) in the conftest hook with a stricter check (e.g. a GPU_ONLY feature, or just an explicit set of GPU-only backends). Then xpress runs by default and cuPDLPx-style backends still require --run-gpu.

Source

test/conftest.py, lines 37–57.