fix: PT-9b - honest variance for unseedable disocclusions; 7x7 borrow

[proggeramlug]

The PT-9 seeding handled ROTATION, where a few pixels per frame stream
in at the screen edge next to converged history. Under TRANSLATION -
the actual walking-indoors case - a whole COLUMN of texels streams in
per frame, the 5x5 borrow window finds nothing but fellow newborns,
and the unseeded pixels fell back to the old degenerate state: a raw
1-spp sample with variance EXACTLY zero (m2 - m1 squared of a single
value). Zero variance tells the wavelet filter the pixel is CONVERGED,
so the raw outliers survived all four a-trous iterations as a visible
noise band along the leading screen edge.

Two changes, both in the temporal kernel, realtime mode only:

• the borrow window grows to 7x7 (newborn pixels only, so the cost is
  motion-proportional and small);
• a newborn with NOTHING to borrow now writes a frank variance
  (max(l squared, 0.25)) instead of the degenerate zero - a 1-sample
  estimate has unknown variance, not none - so the a-trous blurs those
  pixels hard until real statistics exist one frame later.

Verified: both golden PT tests pass; the translating-dolly capture
(sideways translation, worst-case parallax) shows the noise band gone
at unchanged fps; progressive mode untouched.

Claude-Session: https://claude.ai/code/session_018574BdCfSjdpNK3WLgx1hP

Summary by CodeRabbit

• Bug Fixes
  • Improved real-time path-traced rendering during fast camera movement.
  • Reduced disocclusion artifacts by making better use of nearby compatible visual history.
  • Improved handling of newly revealed areas, producing more stable noise and variance estimates.
  • Enhanced image quality and temporal stability in scenes with significant camera translation.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 8340792e-ee3a-4361-bbbd-f13d53a8003b

📥 Commits

Reviewing files that changed from the base of the PR and between 6a5d63d and dd5644d.

📒 Files selected for processing (1)

• native/shared/src/renderer/shaders/pt.rs

---

📝 Walkthrough

Walkthrough

The PT shader’s SVGF accumulation path expands disocclusion neighbor searches from 5×5 to 7×7, tracks successful history seeding, and applies a fallback variance estimate for young unseeded history.

Changes

SVGF seeding updates

Layer / File(s)	Summary
Neighbor seeding and variance handling native/shared/src/renderer/shaders/pt.rs	The shader tracks successful neighbor-history seeding, expands the disocclusion search window, and derives variance from l_new*l_new with a floor for very young unseeded history.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related PRs

• Bloom-Engine/engine#92: Introduced the realtime SVGF/PT accumulation logic extended by this change.
• Bloom-Engine/engine#96: Also modifies SVGF disocclusion neighbor-history seeding in the same shader.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/pt9b-translation-band

---

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.}