vchelaru RaylibGum's Rendering Pipeline
This skill covers the raylib backend only. MonoGame/KNI/FNA go through a different renderer entirely — see gum-monogame-rendering.
Architecture Seam: BatchDrawCallCounter
Renderer.Draw routes every raylib scissor/blend/mode-2D/shader state change through BatchDrawCallCounter (Runtimes/RaylibGum/RenderingLibrary/BatchDrawCallCounter.cs) rather than calling raylib directly, so draw-call stats stay accurate. Per-element renderables (Sprite.cs, NineSlice.cs, Text.cs) call BatchDrawCallCounter.BeginBlendMode(Blend.Value)/EndBlendMode() around their own draw whenever Blend.HasValue — and every Sprite/NineSlice gets Blend = Blend.Normal by default (StandardElementsManager.AddBlendVariable), so this fires for virtually all content, not just elements with an explicit non-default Blend.
Known Gap: Blend.Normal/Additive Bypass the Render-Target Premultiply Pass
BeginRenderTargetBlend()/_renderTargetBlendActive (BatchDrawCallCounter.cs) exist so content baked into an offscreen render target composites back without a "double-blend dark fringe" — see Renderer.cs's BakeRenderTarget/CompositeRenderTarget. But BeginBlendMode(Blend blend) calls TryGetSimpleRaylibBlendMode (BlendModeExtensions.cs) first, and that method maps Blend.Normal/Additive straight to a canned raylib BlendMode without ever consulting _renderTargetBlendActive — only Replace/ReplaceAlpha/SubtractAlpha/MinAlpha reach the ambient-aware custom-separate factors. So the premultiply-pass protection never actually engages for ordinary translucent content, which is the common case. Tracked in #4204; fix that engages this path for Normal/Additive too, not a caller-side workaround.
The Underlying Alpha-Channel Gotcha
Raylib's canned BlendMode.Alpha uses the same GL factors (SrcAlpha/OneMinusSrcAlpha) for the color and alpha channels. A single translucent draw over an opaque backdrop still computes the correct color (color blending doesn't depend on the backdrop's alpha), but the resulting alpha comes out short of 255 (srcA² + dstA*(1-srcA) instead of the correct srcA + dstA*(1-srcA)). That's invisible as long as nothing downstream reads the render target's own alpha again — but any consumer that composites that render target a second time (blits it onto another target, reads it back for export) re-multiplies by that leftover alpha and visibly darkens exactly the translucent regions.
A host application that bakes Gum's draw output into its own render target and later composites that render target elsewhere (e.g. blitting it to the window) must not weight that second pass by the render target's own alpha — make it a straight replace-copy, or flatten the alpha first. Airpig PR #200 works through a concrete instance of this. gumcli screenshot --backend raylib and the Gum tool never hit it, since both do a single compositing pass and never re-consume that leftover alpha.
No user-facing docs exist for this yet — MonoGame has an equivalent gotcha documented (docs/code/rendering/gumbatch.md's "RenderTargets" section, with a corrected BlendState); raylib has no counterpart. Tracked in #4205.