Harden Capability OS release candidate gates

[rblake2320]

What changed

• adds baseline-subtracted cold VRAM measurement and preserved RTX 5090 evidence
• moves TPM attestation to a real unelevated user-scoped Platform Crypto Provider key
• makes shadow approvals explicitly human-reviewed and permanently non-publishing
• maps all 12 architectural invariants to collected pytest node IDs
• introduces explicit no-platform-substitution test doctrine
• separates cross-platform hermetic CI from the Windows platform tier
• preserves fresh 10-run known and holdout Qwen 3.6 vs GPT-OSS matrices

Why

This closes the RC-to-GA review gaps without turning unavailable hardware or integrations into mocked passes. The benchmark was rerun only after identifying and stopping an unrelated NVIDIA NIM container that distorted VRAM and latency.

Validation

• 947 passed, 15 justified skips, 0 failed
• real unelevated TPM quote self-test passed
• 9 cold Qwen VRAM trials plus 3 cold Qwen3-VL trials
• 12/12 invariants resolved to 962 collected test nodes
• Qwen: 10/10 known and 10/10 holdout eligible with full outcomes
• GPT-OSS: faster and lower VRAM, but failed known hard gates in 7/10 runs
• Ruff passed
• wheel built with all 33 required modules
• 29/29 release claims and 25/25 tagged README claims validated