Retain raw attachment and capability evidence for image portability

[proggeramlug]

Sponza and skinned/alpha motion fail the portable image thresholds on Windows while the same canonical scenes pass on hosted Metal. Final PNGs alone cannot identify the first divergent stage, and independently normalized depth previews cannot prove numerical depth equality.

Add a hosted Metal diagnostic lane using SHA-256-pinned Perry 0.5.1182 and the canonical scene commands. Add opt-in BLOOM_QUALITY_RAW=1 export of exact attachment rows, format/dimension/checksum metadata, and the existing raw HDR/material/motion/albedo readback. The lane retains success and failure evidence and separately rejects incomplete or wrong-backend captures; shared-runner timing remains informational.

The cutout-input tool makes temporary opaque-card builds, exports exact f32 U/V bits and the original coverage decision, and restores the source and native library afterward. It exposes accepted and rejected nearest-card fragments for comparison across backends.

Hosted shared tests now retain adapter and capability-skip output. This exposed that the Metal profiler test skips because timestamps are unavailable on the Apple Paravirtual device. The prior Metal timestamp-coverage claim is corrected; Radeon Vulkan/DX12 validation is unchanged.

Validation:

• Both initial hosted Metal images pass existing thresholds: Sponza SSIM 0.997442544 and skinned/alpha 0.999417603, on source 98cce62 (generated executable is the only dirty path).
• Windows still fails both images with the same modern tier and software GI. The portability cause remains under investigation.
• The raw export passes its row-padding/truncation test, contracts, and lint. All 68 attachments from two real Windows captures match their declared byte counts and checksums. Final PNGs are byte-identical with raw capture disabled/enabled.
• All 24 hosted checks pass at 30e7625; the two scheduled physical-hardware jobs are skipped. Both raw Metal captures pass their visual thresholds. The cutout probes produce 103,644 corresponding masked-card pixels on each backend, and their U/V runs retain identical depth and decision data.

Stacked on draft #156. Related: #128, #140, #153. Approved images, thresholds, and performance budgets are unchanged.

Initial raw evidence and investigation report. The exact inputs identify different Bayer phase extents from floating-point exp2/floor; follow-up draft #158 makes the approved phase grid explicit with integer arithmetic. Its evidence release includes these original inputs, the corrected captures, and the final validation results.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}