[yujiaxu-ops]

Step 3 of the consolidation agreed after #60: the live passes' data layout, and renderer/src/live.rs split into files. One PR, two commits; neither changes a picture (tier R1).

1. The buffers by what they hold (e4718cd). The state had grown by reuse: a pair named alb held the diffuse fast mean, the upscaler's flicker and lock, and judge's findings over one another; the output colour history sat behind the traced grid's diffuse history in one buffer; velocity had a ping-pong partner nothing read; the g-buffer was allocated at twice what it holds; every pair was swapped by binding the other of two bind groups. Now each lasting state has a buffer of its own, named for what it holds, in two halves (this frame's and the last, chosen by LiveParams.state): diffuse_hist, diffuse_mom, diffuse_fast, spec_hist, gbuffer, up_state, color_hist. The transients that share a buffer are listed at the top of live.wgsl with who uses each when. up_state has half of z and all of w free: the lock's depth (the next fix, for lamps missed in motion) goes there.

• Storage bindings 28 → 24, one bind group, 16 bytes a pixel less (59 MB at 1440p). Merged layout: 21 live buffers.
• Live passes 6.27 → 6.13 ms (Bistro, 1440p at 0.375).

2. live.rs in files (8c3ae2a). 4,931 lines → live/mod.rs (LiveCore, settings, camera, governor, readbacks), targets.rs (pipelines, buffers, bind group), load.rs (a scene loaded, the sky's and emitters' workers), edits.rs (edits in place), frame.rs (one frame), session.rs (the headless runner, native only). The items moved as they were, made visible to the parent where it uses them. Every reference to live.rs in comments, docs and tests points at the file the item is in now (one-line edits in world.rs, scene.rs, gpu.rs, docs/VOXEL.md, docs/RUNTIME.md and others).

Checks

	result
render-identity, buffers vs main, native	nothing changed (after confirm runs: grow-live a flake, three tests within their spread)
render-identity, buffers vs main, THREERT_MERGED=1	nothing changed (physics-instance and shared-pose within their four-run spread)
render-identity, split vs buffers, native	190 outputs: 172 byte-identical, 18 within run-to-run spread, 0 changed
render-identity, split vs buffers, merged	190 outputs: 175 byte-identical, 15 within spread, 0 changed
bench-window still-bistro, tokyo-turn	radiance cache on: as close to main as main is to itself (0.03–0.05 ΔE between frames). Cache off: 0.03 against main's 0.01 to itself. A shader compiled from other source rounds differently
cargo test (renderer)	124 passed, 6 ignored
wasm32 library, web crate, npm typecheck / test	build; 159 passed

render-identity changes too. It confirms what it flags with four more runs a side, alone (two runs a side under-read a few tests' spread). A confirming run that leaves no output is now left out and counted instead of aborting the gate. geometry/grow-live's frame script adds 1,100 objects and sometimes runs past the script's time limit, so no picture is written. This happens on either build.

ROADMAP § 1a records the step.

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches the live denoise/upscale shader bindings and buffer sizing used every frame; mitigated by render-identity checks, but merged-layout and FP rounding can still differ slightly from main when the cache is off.

Overview
Reorganizes live post-process GPU state so each temporal quantity has its own storage buffer (diffuse_hist, diffuse_mom, diffuse_fast, spec_hist, gbuffer, up_state, color_hist, sample_vel, etc.) instead of overloaded ping-pong pairs (hist_in/out, ae_in/out, xhist). Frame-to-frame halves are selected via new LiveParams.state and this_half / last_half in live.wgsl; bind layout drops from 28 → 24 storage bindings and saves ~16 B/pixel at 1440p.

Splits the former renderer/src/live.rs (~4.9k lines) into live/ modules (mod.rs, targets.rs, load.rs, edits.rs, frame.rs, session.rs) with no intended behavior change; docs, comments, and tests now reference the new paths (e.g. live/edits.rs for edit_geometry, live/session.rs for headless live).

ROADMAP notes render-identity parity with main, a small live-pass win (~6.27 → 6.13 ms on Bistro), and tighter render-identity confirmation rules.

^{Reviewed by Cursor Bugbot for commit 8c3ae2a. Bugbot is set up for automated code reviews on this repo. Configure here.}