Windows and Metal image portability diagnostics (2026-09-10)
Pre-release
Diagnostic evidence for draft PR #157 at 0dd8f67480f154f462f9db74b155431d948a6609.
Both canonical images pass on hosted Metal. Windows still fails with the same modern tier and software GI. Raw depth exposes 9,745 differing geometry/clear coverage decisions in skinned/alpha motion before TAA. The precise cutout cause remains under investigation; no baseline, threshold, or performance budget changes.
The ZIP contains full raw Metal and Windows captures, the controls, exact commands/source snapshots, checksums, and a report. All 68 raw attachments on each backend match their declared sizes and checksums. On Windows, enabling raw export preserves final PNGs byte-for-byte. All 24 hosted checks pass at this source; two scheduled physical-hardware jobs are skipped.
Archive: windows-metal-image-portability-20260910-v1.zip
SHA-256: b28a646e6063c3a0d6996e86fe3f594dcd7fb3dd610925df332b1672c9c36ad1
Size: 109,739,974 bytes; 637 entries including a hash manifest.
The hosted Apple Paravirtual device lacks GPU timestamps. Retained test output explicitly confirms that its profiler regression skips. This corrects the earlier Metal timestamp-coverage claim; Radeon Vulkan/DX12 validation is unaffected. Shared-runner timing and the recorded busy-host Windows sample are not hardware performance qualification.
The two Windows visual failures, HD temporal stability, and the broader engine completion plan remain open.