[yujiaxu-ops]

Step 1 of the consolidation agreed on 2026-11-13: remove what was measured and left off, before the harness moves into the repo and the data layout is restructured. Nothing a scene shows changes.

Removed

• The temporal pass's two round-3 experiment switches (THREERT_TEMPORAL_FLAGS 1 and 2) and the flag word that carried them (LiveParams.mode.w, now unused).

• The emitter-reservoir experiment (THREERT_RESTIR, ReSTIR DI at camera hits). It was measured and left off: first-frame SSIM 0.608 → 0.613, motion +0.1 ΔE, +0.44 ms. Removed with it:

  • its kernel code and the helpers only it used;
  • the reservoir buffer (group 1, binding 4);
  • five vec4 of Params and the F_RESTIR bit.

  The kernel now binds 19 storage buffers as written, and the merged layout's MergedLive has six arrays.

• Five A/B and override switches:

  • THREERT_NO_MV: no per-object motion vectors.
  • THREERT_LT_DYNAMIC=0: emitters added in place by power share. bench-voxel-light's third row goes with it.
  • THREERT_EMIT_SYNC: unreferenced.
  • The overrides of three constants: THREERT_EMITTER_CULL, THREERT_EMITTER_CULL_VIEW, THREERT_LIGHT_LOD.

Kept

• What the benches and tests use: THREERT_PROFILE, _PROFILE_BLAS, _FREEZE_TIME, _EXPOSURE_SNAP, _NO_GRAIN, _MOTION_BLUR, _EDIT_REBUILD, _INSTANCING.
• Diagnostics: _EDIT_LOG, _VIEW_MODE, _TL_HIDE, _CACHE, _CACHE_FLAGS.
• Configuration: _GOVERN_CLOCK, and _RASTER (the primary-visibility path kept for devices without hardware ray tracing).

The other renderer files have about twenty more switches; they are not reviewed here.

Checked

• test-render, native layout: main run twice and this build run twice, every output compared byte for byte.
  • Every output a run repeats to the byte is byte-identical to main's.
  • The fifteen or so that differ between two runs of one build (async emitter builds, physics, skinning, media) differ here within that.
• test-render with THREERT_MERGED=1 (the browser layout, natively):
  • The same 166 pass and the same 24 fail as on main.
  • One volume test differs at a relative 2e-6: the kernel is compiled from a shorter source.
• cargo test --release --locked 124 + 2 (the merged tests count one buffer fewer), npm test 159, typecheck clean.

Found on the way, not touched here

On main, 24 tests fail in the merged (browser) layout: particles, GPU particles, translucency order and depth, volume lighting, wind sway, surface windows. They fail identically before and after this change.

🤖 Generated with Claude Code

---

Note

Low Risk
Mechanical deletion of disabled code paths with documented byte-identical test-render parity; no change to default scene output.

Overview
Removes live-renderer experiment code and env overrides that benchmarks showed were not worth keeping, so default rendering and tests stay unchanged while the shader/kernel surface area shrinks.

ReSTIR DI at camera hits (THREERT_RESTIR) is deleted end-to-end: large blocks in pathtrace.wgsl, the resv group-1 buffer and bind layout, related Params fields, and LiveCore wiring. Emitter direct lighting always uses the existing RIS/CDF path again. The merged layout drops one MergedLive array (six instead of seven; kernel 19 storage buffers).

Temporal round-3 toggles (THREERT_TEMPORAL_FLAGS) and live.wgsl's xflag path are removed; à-trous noise gating and specular filter reach stay on the shipped logic only.

Other retired switches bake in the measured defaults: per-object motion vectors always on, light LOD / emitter cull constants fixed in Rust, async emitter worker no longer gated by sync/lt-dynamic flags, and bench harnesses stop treating removed THREERT_* keys as part of the ref cache key.

docs/ROADMAP.md documents what was removed vs kept for benches/diagnostics.

^{Reviewed by Cursor Bugbot for commit 3b84760. Bugbot is set up for automated code reviews on this repo. Configure here.}