[yujiaxu-ops]

What was wrong

Dragging the view over Tokyo at night, the tower broke up: its legs and most of its lamps gone while the camera turned, back half a second after it stopped. Measured against offline renders, these are two faults of the upscaler, both older than #59.

• Held still, a field of lamps a third of a pixel wide kept a third of its light. judge takes a surface for gone from in front when what was nearest among the nine samples a frame ago is nearer than what is nearest now, and the history's colour is not in this frame. A lamp the jitter hit last frame and missed this one is exactly that, so its history was dropped every frame it was missed.
• In motion the history is clamped into the frame's nine samples, and a lamp none of them hit is outside them. The 4-sample cap in motion is not the cause: without it nothing changed.

The change

The remedy is the one FSR 2 (pixel locks) and UE's TSR (thin geometry detection) have: where the picture is a thin feature, the rules that hold the history to this frame's samples stand back.

Lock in live.wgsl is a half word per traced pixel, in space the fast mean's buffer had free:

• compose makes it where the sample's emission is thin: no 2 × 2 block of the nine like it (FSR 2's rule), and above or below both of some pair of opposite neighbours.
• It lasts a jitter sequence, is carried with the surface, and is never set on animated shading.
• A locked pixel's history is not clamped in motion (the lock of any of its nine samples counts).
• A locked sample's distance from the history's range is no finding.
• Nothing is gone where one of the nine samples is locked.

The test runs on emission (what the camera ray saw directly) because that part of a traced sample has no noise. FSR 2's test on the composed sample called most of a night street "thin".

Measured

	main	this	offline
lamps field held still, mean radiance	0.035	0.106	0.103
the same, the camera sliding 1.4 px a frame	0.027	0.112	0.103
Tokyo Tower in a turn, ΔE / SSIM / luminance against offline	2.81 / 0.867 / 0.77	2.23 / 0.928 / 1.05
the same, the whole frame	2.69 / 0.889 / 1.01	2.56 / 0.915 / 1.06
the tower half a second after the stop	1.89 / 0.948 / 1.07	1.47 / 0.964 / 1.00
the tower settled	1.48 / 0.962 / 0.94	1.39 / 0.965 / 0.99

• test-render 190 pass, cargo test --release --locked 114 + 2, npm test 159, typecheck clean.
• New tests upscale/lamps-still-live and upscale/lamps-in-motion-live: main fails both (0.021 and 0.027 of 0.1).
• compose + judge + upscale: 1.73 → 1.80 ms at 1440p output.

What it costs

Where a view translates past emitters, the locked ones trail a little. bench-ghosting still passes on both scenes.

	main	this
Kabukichō signs, mid-strafe	6.42	6.67
Kabukichō signs, 0.1 s after the strafe (trail)	5.27 (26 px)	5.55 (31 px)
Kabukichō signs, at the dolly's stop	6.75	7.07
Bistro whole frame, 0.1 s after the strafe (trail)	2.49 (13 px)	2.63 (24 px)
Bistro slice, half a second after its explosion	5.28	5.41
Thames orbit, moving / held	2.24 / 1.34	2.29 / 1.24
Tokyo at dusk, skyline p99 at 6 s	11–14	16–18

• Kabukichō's ground is the same to 0.01 in every shot; its whole frame is within 0.08.
• Bistro's other shots are within 0.05.
• Held still, the fire shot and the lighting-edit shot are as they were.

Left

• Thin things that do not emit (a wire, a lattice by day) are still taken for gone under the jitter.
• The locked lettering of a still sign comes back from a motion's blur by the blend alone.
• A lamp that comes on arrives over a quarter of a second.

What was tried and left out is in docs/ROADMAP.md § 1a.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes core temporal upscaler behavior (compose, judge, upscale) with intentional ghosting tradeoffs on fast translation; validated by new tests but affects all live upscaled output.

Overview
Adds a thin-feature lock (Lock in live.wgsl) so sub-pixel emitters (lamps, lit windows) are not discarded by temporal upscaling when jitter misses them or motion clamps history into the nine-sample neighborhood.

compose detects thin emission (FSR2-like local similarity + opposite-neighbor contrast), persists lock state in spare ae_in f16 bits for a jitter sequence, and skips animated shading. judge no longer treats locked samples as disoccluded “gone” or penalizes history-range distance; upscale relaxes in-motion history clamping by LOCK_RELAX (0.75) when any contributing sample is locked; reject propagates max lock across the nine neighbors.

docs/ROADMAP.md documents the problem, tuning tradeoffs, and benchmark deltas (e.g. lamp field radiance 0.035→0.106 still, Tokyo Tower metrics improved). New render tests upscale/lamps-still-live and upscale/lamps-in-motion-live guard mean radiance (~0.1) for procedural sub-pixel lamp fields.

^{Reviewed by Cursor Bugbot for commit fa89a24. Bugbot is set up for automated code reviews on this repo. Configure here.}