refactor(shaders): make all WebGL fragment shaders branchless for Mali 400

[chiefcll]

What

Removes every if/else/ternary from the WebGL fragment shaders, per the Mali 400 / OpenGL ES 2.0 GPU floor documented in CLAUDE.md (in-order scalar fragment pipelines serialize both sides of any branch — including uniform-driven ones).

• Ternary quadrant-radius selects (r.xy = (p.x > 0.0) ? r.yz : r.xw; ...) in RoundedWithShadow, RoundedWithBorder, RoundedWithBorderAndShadow, HolePunch → the branchless step()+mix() select already used in Rounded (shared quadRadius() helper).
• Varying-driven branch in RoundedWithBorderAndShadow (worst offender — diverges per fragment) selecting between two shadowBox calls → step()/mix() parameter select into a single shadowBox call.
• Uniform-driven early-return paths (borderZero, u_borderGap, u_color.a, u_shadow_color.a) in RoundedWithBorder, RoundedWithBorderAndShadow, Border, Shadow, RoundedWithShadow → compute all composites, mix()-select the result.
• Gradient stop loops with data-driven if + return-inside-loop in LinearGradient, RadialGradient, RadialProgress → the previously-unused genGradientColors() mix-accumulation in ShaderUtils, rewritten to smoothstep-per-segment so it is pixel-identical to the old segment select, and unrolled per colors:N #define variant.
• RadialProgress uniform ternaries (u_duration, u_countdown, progress normalize) → step()/mix(), with mix()-guarded denominators so the divisions stay finite at 0 without overflowing fp16 mediump to Inf/NaN (a max()-epsilon guard would).

Why step/mix instead of #define variants for the uniform branches

Shader props are runtime-mutable — animations write shaderProps directly and CoreShaderNode prop setters fire RecalcUniforms — while the program is fixed at createShader() time (getCacheMarkers is evaluated once). Keying a compile-time variant on border-w == 0 / border-gap == 0 would silently break a border animating from 0. The structural colors:N markers (array length can't animate) keep their #define unrolling. On Mali 400 the old serialized branch already paid for all paths per batch, so computing-everything+mix is equal or cheaper.

Reviewer notes

• The border vertex shaders now initialize v_outerBorderUv/v_innerBorderUv/v_*BorderRadius in the zero-border case — the branchless fragment always evaluates the border SDFs, and an unwritten varying is undefined and can poison the final mix() with NaN. (The remaining vertex-shader if(borderZero == 0.0) is fine: the Mali 400 constraint is the fragment pipeline.)
• The gap select uses step(0.0001, abs(u_borderGap)) — abs() matters, negative gaps are valid (see shader-border-gap) and the original condition was == 0.0.
• No varyings added (RoundedWithBorder* stays at 7/8 rows), #ifdef GL_FRAGMENT_PRECISION_HIGH guards kept, every non-void GLSL function ends in an unconditional return, uniform values remain a pure function of resolvedProps + node w/h.
• Canvas backend untouched (no branching constraint there).

Verification

• Visual regression vs the certified chromium-ci snapshots (generated from the old branchy shaders), run in Docker: 180/180 pass, pixel-identical.
• New shader-border-zero example test + certified snapshot covering the merged degenerate paths (zero-width border on both RoundedWithBorder* shaders, zero-alpha shadow on Shadow/RoundedWithShadow) — manually inspected.
• New unit tests execute the generated genGradientColors GLSL (via mix/smoothstep shims) against the original segment-select reference across a 100-point sweep.
• pnpm build, pnpm test (274 tests), pnpm lint clean.

🤖 Generated with Claude Code