[yujiaxu-ops]

Follows #87. A camera hit read its light list at its own pixel: with the camera turning, the list of another part of the scene. Now it reads the list of the tile it was seen in when the lists were built: its point, moved back by its motion, through that frame's camera (Params::lt_cam); the nearest tile's when it was off screen.

New shot: bench-window rooms-turn (lamp-rooms-wide, the camera turning 2 degrees a frame for 24 frames, then still).

Measured (M5 Pro, dE / p99 against the offline reference)

frame	before #87	#87	this
last moving	2.00 / 8.57	0.86 / 4.82	0.80 / 3.54
6 frames after	2.24 / 8.75	0.80 / 4.09	0.75 / 3.09
30 frames after	1.96 / 7.43	0.69 / 2.80	0.67 / 2.42
settled	1.00 / 3.56	0.67 / 1.93	0.67 / 1.92

(The reference is the cached light-tree render #87's text mentions.) Still shots as #87.

Cost: 0.19 ms of the trace in a scene with lists (lamp-rooms 3.57 -> 3.76 ms). Bistro 8.43 ms on both.

Checks

• cargo test --release: 143 passed. test-render: 192 / 192, with the loop guard.
• Loop guard at 5 limits on Bistro and both lamp-rooms scenes: no stop.

Not done

• No depth test: a surface newly in view reads the list of what was there before (a guide only: noise, not bias).
• The 0.19 ms is more than the arithmetic explains: the lookup now sits in the path loop's first iteration. Not reduced.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes live light sampling and shader uniforms on camera motion; wrong projection would add noise (not bias per design), with a small trace cost increase on list scenes.

Overview
Light tile resampling now follows the camera instead of the current pixel. Paths look up the 8×8 tile list where each motion-compensated hit point appeared when lists were built, by reprojecting through a new lt_cam uniform (last frame’s position, forward, right, up). Live frame setup stores that camera with the list metadata; lt_list_start takes a world position rather than gid.xy and runs in the path loop after motion is known.

Docs (RUNTIME.md, ROADMAP.md) describe the behavior and bench results for a turning camera on lamp-rooms-wide. bench-window gains a rooms-turn shot (2°/frame for 24 frames, then settle) to regression-test moving views.

^{Reviewed by Cursor Bugbot for commit f587c00. Bugbot is set up for automated code reviews on this repo. Configure here.}