perf(webgl): batch distinct shader nodes with equal uniform collections

[chiefcll]

What

The render-op batch test in WebGlRenderer.reuseRenderOp rejected any two quads whose shader nodes were different instances (curShader !== shader). Apps create one shader node per card (renderer.createShader('Rounded', {...}) at construction), so a rail of visually identical rounded cards never batched — one render op per card, each paying useShader + bindRenderOp + scissor + drawElements dispatch on the driver CPU. On TV SoCs (~0.5–2µs per GL call) that's a meaningful per-frame tax for the standard left-nav/rows layout.

How

Shader nodes with equal value keys already share their uniform collection by reference (the shader-value cache in WebGlShaderNode.update), and collections are immutable after fill — so reference equality implies value equality. The gate now batches distinct shader nodes when they resolve to:

1. the same program, and
2. the same uniform collection (reference compare).

Program identity is required for correctness: the shader value key (CoreShaderNode.createValueKey) does not include the shader type, so an equal collection reached from a different program (key collision across shader types) must never merge. A regression test covers this case.

WebGlShaderProgram.reuseRenderOp gets the same one-reference compare in place of the per-node for..in over resolvedProps (the prop walk is kept as a fallback for distinct collections), and no longer invokes the node.time getter for programs without a time uniform.

Measured

?test=stress-tv&debug=true, 200 rounded-rect cards (tier "rounded rect only"), clean dev server per side:

	draws	quads	GL calls/interval
main	205	204	845
this PR	6	204	48

Reviewer notes

• The per-op system checks (worldAlpha / dimensions / time) still run in WebGlShaderProgram.reuseRenderOp after the gate, so alpha-animated or differently-sized cards split correctly.
• Verified visually that different-props nodes still split: the shader-rounded example (every node has different corner/border/shadow props) renders pixel-identical — borders, shadows, and per-corner radii intact. The stress-tv interleaved tier (cards + unique image textures) still splits on texture changes as before.
• Batching semantics for same-instance shader nodes, the default shader, RTT, and clipping are unchanged.

Testing

• Full unit suite: 295 tests pass
• New: WebGlRenderer.reuseRenderOp.test.ts (8 tests — shared-collection batching, cross-program rejection, clipping/default paths, program-level veto still consulted) and WebGlShaderProgram.reuseRenderOp.test.ts (7 tests — reference short-circuit, prop-compare fallback, alpha/dimension rejects, time getter untouched, canBatch delegation)
• pnpm build clean; Prettier/ESLint clean (0 errors)

🤖 Generated with Claude Code