[froststeam]

Summary

• prefer the pairwise mtmlDeviceCountMtLinkLayouts() query for NVLink-compatible P2P detection
• avoid materializing handles for host GPUs hidden by MTHREADS_VISIBLE_DEVICES
• retain the per-link scan for older runtimes that do not support the layout API
• add regression tests for the pairwise path and the legacy fallback

Root cause

In an 8-GPU host with only four GPUs exposed, mtmlDeviceGetMtLinkSpec() still reports all 14 physical links as UP. mtmlDeviceGetMtLinkRemoteDevice() correctly returns MTML_ERROR_NOT_FOUND for links whose remote endpoint is outside the visible device set. Pairwise P2P detection should not enumerate those hidden endpoints.

Validation

• 15 passed (test_mtlink_visibility.py, test_abi_layout.py, test_library_loading.py)
• S5000, driver 5.2.0-server, MTHREADS_VISIBLE_DEVICES=0,1,2,3: all six visible pairs returned NVML_P2P_STATUS_OK, each reported two MtLink layouts, and the patched pairwise path performed zero remote-device lookups
• legacy fallback test skips MTMLError_NotFound links and still finds a visible peer

The repository baseline currently has pre-existing full-file Black/flake8 failures; the new test file passes Black and isort checks.