Add bounded CUDA hardware probe - #70

Summary

• implement explicit miniverl plan --probe calibration while keeping normal planning weight-free
• load actor and teacher sequentially, measure tiny rollout candidates, one selected-position backward and one teacher top-k score, then verify CUDA allocation returns near baseline
• cache only under an exact hardware/software/model/token/plan identity and keep cache transport metadata out of the immutable plan digest

Measured RTX 4080 probe

• device: NVIDIA GeForce RTX 4080, compute capability 8.9, driver 596.49
• Torch/CUDA: 2.13.0+cu130 / 13.0
• pinned Qwen3-0.6B student and Qwen3-1.7B teacher, NF4
• duration: 13.63 s
• successful rollout candidates: batch 1, 2, 4
• parameter updates: 0
• checkpoint published: false
• allocation after role release: 17,040,384 bytes
• exact-cache reuse reproduced byte-identical plan bytes

Validation

• non-GPU/non-network: 2,203 passed, 9 skipped, 21 deselected; 84.54% branch coverage
• RTX 4080 GPU: 8 passed
• network: 13 passed, 2 environment-dependent skips
• Ruff, format, mypy, actionlint, strict MkDocs, link/text/release-state checks and 4-viewport Playwright passed
• dev wheel SHA-256: 1929e34ab873c35717fd2b8e5d6c0c1210d98535dbd2506bf5b06c42fa3f6615
• dev sdist SHA-256: 050c30a230d46f5fa28806e12c94a8d7cf452789a5bd240e12ac7fd056a073a1
• frozen calculator SHA-256 remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc

The probe is calibration, not training, a throughput benchmark or an alignment result.