Particles update on WebGL2 and keep their precision anywhere in the world

[pasquelin]

Why

Split from #420 (a came out at ~645 lines): the particle pool and its WebGPU step land with #420; WebGL2 needs its own update backend. Half-float targets lose steps under ~1.5 cm past ~32 m from the origin.

To do

• WebGL2 particle update: the half-float ping-pong backend behind Fluids S2 (spike): one GPU particle system on WebGPU and WebGL2 #420's shared backend interface, and the composer hook that runs it.
• Precision held anywhere in the world: particle positions are stored relative to their emitter, so a step of 1 mm holds at any distance from the origin (one mechanism on both backends; no 32-bit special case).

Code context

#420's backend interface and pool, the WebGL2 composer, the parked branch of #420's coder.

Proof

Tests by behaviour: the WebGL2 backend matches the WebGPU step on a reference emission within half-float tolerance; a 1 mm step holds at 10 km from the origin. The measurer's branch proof on WebGL2.

Links

Parent #417. Blocked by #420. Split from #420.

[pasquelin]

taken by lead physics — built on #420's branch (head 054a73f) while #420 waits for its GPU proof; its PR opens after #420 merges.

[pasquelin]

Design note (lead physics). Approach: rebase the parked branch 420-particle-webgl2 onto 420-particle-pool (054a73f) as a new branch 759-particle-webgl2, adapted to what #420's review changed: no shared interface today (it comes back here: the WebGL2 and WebGPU steps behind one backend type, only now that there are two), emitter-relative records (the pool's origin; the step is translation-invariant, so no origin in the uniforms), the refused-pool state, the idle-pool rule and the frame hold. The composer runs the WebGL2 step where #420 refuses today, and the refusal goes. Budget: two half-float RGBA ping-pong targets per pool allocated once, zero allocation per frame, no readback; EXT_color_buffer_half_float/_float required, refused by name without it. Paths: WebGL2 fragment ping-pong only. Scenes: fast tests (the WebGL2 step's words and targets against a fake context; the same reference emission stepped by the CPU model of both steps within half-float tolerance; a 1 mm step at 10 km); the measurer's WebGL2 proof with a probe beside #420's particles-step-gpu.ts (WebGL2 readback in the probe only).

[pasquelin]

CTO decision (recorded by lead physics), replacing the design note's half-float targets: one format, RGBA32F ping-pong targets (EXT_color_buffer_float); without the extension WebGL2 particles are refused by name (a named degrade) — no half-float path, no lifetime cap. Positions stay emitter-relative. Reason: a half-float age stops growing once dt is below its precision (~16 s at 144 Hz), so long-lived particles never died on WebGL2. Proof added: the probes show a 60 s lifetime dying on both backends.

[pasquelin]

Deviation from the design note (lead physics), before the merge: the two steps do not share one backend type — no caller needs it (WebGPU steps inside a command encoder, WebGL2 on the context); they share poolStates.ts (per-pool state, emitted-slot count, the 'any pool moving' rule), which is what removed the duplication. The half-float wording of this issue's body is overridden by the CTO decision above (RGBA32F only).