WebGL2 draws the particle pools

[pasquelin]

Why

Split from #755: the particle draw on both backends passes one pull request. #755 ships the WebGPU draw and refuses the pools by name on WebGL2; a WebGL2 world then shows no particles. This issue gives WebGL2 the same draw, about 200 hand-written lines.

To do

• WebGL2 draws the particle pools: additive fire blended order-free; premultiplied smoke sorted coarsely per emitter; soft edges fading with the scene depth. Same image rules as Particles draw without a global sort: additive fire, sorted-per-emitter smoke, soft edges #755 on WebGPU; the refusal Particles draw without a global sort: additive fire, sorted-per-emitter smoke, soft edges #755 puts on WebGL2 is removed.
• Depth readable on the default framebuffer: the soft-edge depth copy works when the scene renders to the default framebuffer; the copy target's format is probed in order DEPTH24_STENCIL8 then DEPTH_COMPONENT24; a context offering neither refuses the pools by name, never draws hard edges as soft.
• A refusal never fails the session (only if Particles draw without a global sort: additive fire, sorted-per-emitter smoke, soft edges #755 did not already deliver it): a non-throwing refusal channel for particle pools, like effects.refused; a refused pool is named with a notice and the world session carries on.

Code context

• The WebGPU draw of Particles draw without a global sort: additive fire, sorted-per-emitter smoke, soft edges #755 (branch 755-particle-draw): the blend modes, the per-emitter sort and the soft-edge fade are reused, never rebuilt for WebGL2.
• The WebGL2 particle step of Particles update on WebGL2 and keep their precision anywhere in the world #759: the pool is read in place, it stays the only particle store.
• effects.refused: the existing non-throwing refusal channel the particle refusal follows.

Proof

• Fake GL test that bites: a strict fake WebGL2 context where the default framebuffer offers only DEPTH_COMPONENT24 (the copy works), then neither format (the pools are refused by name, the session keeps running); the test fails if the probe order or the refusal is removed.
• The measurer's run of the Fluids S2 (spike): one GPU particle system on WebGPU and WebGL2 #420 bench scene (fires and smoke) on WebGL2: additive fire, per-emitter premultiplied smoke and soft edges checked in the image, draw cost per tier published on Fluids S2 (spike): one GPU particle system on WebGPU and WebGL2 #420; ten-thousand-bodies (no particles) at 0 px on both backends.
• Gates: check:changed, test:changed, validate.

Links

Parent #417. Blocked by #755. Split from #755.

[pasquelin]

Split from #755 (CTO): the WebGL2 draw, with the soft-edge depth copy on the default framebuffer. Starts once #755 merges; the non-throwing refusal To-do applies only if #755 did not deliver it.

[pasquelin]

Lead physics, for whoever takes #844: the WebGL2 draw and its depth-format fix (probe DEPTH24_STENCIL8 then DEPTH_COMPONENT24, strict fake GL) are parked on branch 755-webgl2-draw (3c08581), based before #755's final shape. When wiring it back: the composer no longer runs the #778 WebGL2 step (particles/webglParticles.ts), so wire the step and the draw together, and move the step's throwing PARTICLES_UNSUPPORTED to the non-throwing particlesRefused channel #755 adds (world/render/compose.ts). Known leftover for #423: in a WebGL2-beside-WebGPU comparison the shared pools' refused flag is set by the WebGL2 side every frame (one step state per engine).

[pasquelin]

taken by lead physics (picked before the older #756 because #756 touches the TAA pass, which the running #432 coder may reach; #844 shares no file with it). Design note: resume the parked WebGL2 draw (branch 755-webgl2-draw, 3c08581) onto today's develop: reuse #755's blend modes, per-emitter sort and soft-edge fade; wire the #778/#759 WebGL2 step with the draw in the composer; PARTICLES_UNSUPPORTED moves to the non-throwing particlesRefused channel; depth copy probes DEPTH24_STENCIL8 then DEPTH_COMPONENT24, neither → refused by name. Paths: WebGL2 only, WebGPU untouched. Proof: strict fake-GL test (reviewer runs it), recette image after merge.