fix: PT-9 - SVGF disocclusion seeding + ReSTIR default-on (indoor motion noise)

[proggeramlug]

A true disocclusion started from the raw 1-spp sample with variance
EXACTLY zero (m2 - m1 squared of a single value), so freshly
streamed-in pixels - the whole viewport periphery whenever the camera
translates - showed raw path noise until history rebuilt. Indoors
(bounce-dominated light, the shooter house) that read as a
salt-and-pepper ring around a clean centre, worst at low frame rates
where history rebuilds slowest.

A newborn pixel usually sits next to texels holding converged history
for the very same surface: the wall streaming in at the screen edge
continues inward. accum[] and moments[] are the previous frame
buffers, race-free to read at any texel, so the temporal kernel now
borrows from depth-consistent, converged (n >= 4) neighbours in a 5x5
window, weighted by history length, and inherits half their history
(capped at 8) - the canonical 1/N blend then folds real fresh samples
in over the next frames. A pixel with no depth-compatible neighbour (a
genuinely new surface rounding a doorway) keeps the honest raw start.
Realtime mode only; progressive accumulation is untouched.

ReSTIR DI (PT-4) flips to ON by default: interiors carry 7+ point
lights and uniform 1-of-N NEE picks the light that matters for a pixel
~6% of the time; the indoor A/B showed a visible win at no measurable
frame cost (24 vs 25 fps). BLOOM_PT_RESTIR=0 restores plain NEE.

Verified: golden_pt_progressive + golden_pt_realtime_motion both pass;
panning indoor capture (shooter INDOORCAM harness) shows the
disocclusion band gone at unchanged fps; outdoor RT full-run smoke
clean.

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

Summary by CodeRabbit

• New Features

  • ReSTIR direct illumination is now enabled by default for path-traced rendering, while remaining configurable through the existing environment setting.

• Bug Fixes

  • Improved temporal accumulation in disoccluded areas to reduce visual artifacts and preserve more reliable lighting history.
  • Added fallback history seeding from nearby valid samples when standard reprojection cannot find a match.

[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: 34ef5dd6-bf49-4814-982a-ea0dc42b43b8

📥 Commits

Reviewing files that changed from the base of the PR and between 5720b6a and 2fe9641.

📒 Files selected for processing (2)

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

---

📝 Walkthrough

Walkthrough

The renderer now enables ReSTIR DI by default and seeds missing SVGF temporal history from valid neighboring trace texels during path tracing.

Changes

Renderer path tracing

Layer / File(s)	Summary
Enable ReSTIR DI by default native/shared/src/renderer/mod.rs	BLOOM_PT_RESTIR now remains enabled when unset or set to any value other than 0.
Seed SVGF disocclusion history native/shared/src/renderer/shaders/pt.rs	The no-tap-matched path scans a 5x5 neighborhood for depth-consistent converged texels, seeds color and moment history, and caps inherited history at 8 after halving it.

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

Possibly related PRs

• Bloom-Engine/engine#92: Modifies the related SVGF temporal accumulation path.
• Bloom-Engine/engine#93: Changes the same ReSTIR toggle and SVGF history fallback behavior.

✨ 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/pt9-disocclusion-seeding

---

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.}