Correct SSGI surface reconstruction and HD startup stability

[proggeramlug]

SSGI probe placement reconstructed point-sampled depth at a different screen coordinate, and its normal guard rejected valid small pixel footprints. Together these produced horizontal indirect-light bands and failed HD stationary TAA after its required 16-frame warm-up.

Reconstruct receivers at the selected depth texel center and scale surface edges before normalizing their cross product. Resolve diagnostics use the same geometry. HD now runs in the regular GPU suite; an analytic-plane regression exercises the actual placement pipeline at four extents and four placement phases.

• Original HD limits pass on Radeon/Vulkan, Radeon/DX12, and hosted Metal. Both partial corrections fail; restoring either old placement calculation also fails the analytic regression.
• All nine Radeon images pass in three runs. Runs 2 and 3 pass every configured check, and their 257 compared artifacts are byte-identical with matching metadata and timing differences inside the existing noise bounds. Run 1 retains its host-load failures.
• The complete local shared suite passes (488 library tests, 93 reported golden passes). Metal executes the new plane and HD tests and passes both canonical images.
• All 24 hosted CheckRuns report success, but the Windows job logs reveal a separate shell-invocation gap: no Cargo/test execution is recorded. That status is not hosted Windows execution evidence. The local Windows logs are actual execution; fixing CI is follow-up [CI] Make local, PR, hardware, and release quality gates truthful #140 work.

Published report and evidence include exact source, captures, commands, raw checksums, and the failed controls. Archive SHA-256: 0ea0d9ff4769c7ee9eaed2ed977bf2205f6eda52d328120819743037a5f5645a. No baseline, warm-up, threshold, or performance budget was relaxed. Named RTX/physical Metal qualification and the wider temporal contract remain open.

Stacked on #158. Related to #135, #149, and #128. The engine plan also retains example linking/startup, clean installation, streaming, generated APIs, components, and UI work. No PR has been merged.

[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.}