Fix B200 compiler isolation and validate Colfax GPU labs

[cfregly]

The complete B200 suite previously timed out while Llama numerical checks shared compiler state with earlier tests. Run the unchanged CUDA comparison in an isolated worker with owned cleanup. The integrated rerun passed 5,795 tests with 79 explicit skips and no failures, then exited and drained normally.

The new Colfax FA4 labs now verify the complete upstream Python source tree and exact installed Git revision. CUDA graph outputs are published after actual replay, and real GPU tests cover the setup detector, poisoned output buffers, changed inputs, and an independent oracle. README generation preserves the lab guidance and the reports retain repeated measurements, original failures, and unsupported boundaries.

Validation:

• Complete B200 suite on 46ba89929: 5,795 passed, 79 skipped, zero failures/errors, 31m29s.
• Final Colfax source 356490bd1: 32 GPU tests passed in each pinned environment; both normal full-workload deep-dive runs, both four-seed ABBA repeats, and all eight Nsight report inspections passed.
• Repeated full-output ratios: memory fusion 31.74309x; Colfax decode 1.173108x; Colfax backward 1.080924x. Cache-aware two-GPU inference measured 0.988428x, so the earlier single-run speedup did not reproduce.
• Explicit two-GPU ZeRO2, Blackwell distributed, and local-model vLLM checks passed.
• 936 benchmark entrypoints: zero errors/warnings; repository-wide Ruff correctness passed; README generation checks passed.
• Hosted CI on 568c9102c passed: 5,341 CPU tests passed, 536 skipped, zero failures/errors; static analysis, dashboard validation, and all configured CUDA architecture builds passed.
• The exact TE 2.18 FP8 CUDA template passed in a separate environment with PyTorch 2.9.1+cu130 retained. Its supported build option explicitly excluded optional NCCL expert parallelism; that separate feature is not qualified.
• The final distributed NCU repeat reported driver UnknownError; the failed capture and owned-process interruption/drain receipts remain preserved alongside earlier successful five-metric reports.

These are portable current-host observations. Unavailable hardware/model assets, independent approximate-kernel accuracy budgets, explicitly unimplemented hardening detectors, and intermittent distributed NCU replay remain documented. No performance or correctness gate was weakened.