Portable cutout phase correction: Windows and Metal evidence

Evidence for draft PR #158, source 662a44f85b23fd91cff53b6479e6373f208dc321, stacked on diagnostic PR #157.

Sponza and skinned/alpha motion now pass the unchanged portable images on Windows. The correction computes the approved Bayer phase extent with integer arithmetic in scene and shadow shaders. Exact cutout-input captures explain the previous Metal/Vulkan discrepancy; alpha sampling and mip generation are unchanged.

Both complete Radeon 760M/Vulkan runs pass all nine image gates. The reproducibility check passes, with 257/257 artifacts byte-identical. Both hosted Metal images pass, and all 24 hosted checks pass on the exact source. The GPU regression executes on Vulkan, DX12, and Metal; the old Vulkan shader fails its first reference point.

The two Radeon runs retain overall FAIL status for host-load rejections. Run 1 invalidates Bistro timing; run 2 invalidates Bistro, draw/light stress, and weighted-transparency timing. Windows System-process CPU was above the unchanged 75% limit. The reproducibility result does not override those invalid timing windows. Named RTX 4080 and Apple M1 Max qualification remains unperformed; the hosted Metal adapter has no timestamp-query support.

The ZIP contains the report, full result bundles and host snapshots, focused captures, original and corrected Metal captures, exact cutout inputs, negative control, local and hosted validation logs, source snapshots, and a SHA-256 manifest. All 1,820 entries pass CRC verification; every manifest payload passes SHA-256 verification. A separate receipt verifies the declared byte counts and FNV-1a hashes of all 350 raw attachments (799,028,480 bytes) inside the ZIP.

• ZIP: alpha-phase-662a44f-windows-metal-20260910-v1.zip
• Size: 135,304,167 bytes
• SHA-256: 0cfe31d256cece617e15addba3ffd6d35039f04aec3831021e65283b924d7dfb

This publishes image-correction evidence. No PR is merged, and the broader engine plan remains open.