Graphite shader-error handler + fix iOS-simulator dst read (mono/SkiaSharp#4555)

[ramezgerges]

Description

Three related changes that together fix mono/SkiaSharp#4555 (Graphite/Metal GradientBlend cell returning null Recorder::Snap() on the iOS simulator) and give SkiaSharp callers a way to see the shader/PSO compile diagnostics behind similar failures in the future.

Commits:

1. Add sk_graphite_context_options.fShaderErrorHandler callback — expose Graphite's skgpu::ShaderErrorHandler on the C API so callers can capture SkSL→driver shader compile failures. Skia's default handler is SkDebugf + assert, which is invisible on iOS/tvOS. Bridged through a heap-allocated FfiShaderErrorHandler referenced by ContextOptions.fShaderErrorHandler (Skia stores it as a raw non-owning pointer, so the wrapper is deliberately leaked — one per Context creation with a handler set).

2. Report Graphite Metal PSO-creation failures through the ShaderErrorHandler — a Graphite Metal shader can pass the MSL front-end compile yet still be rejected at newRenderPipelineStateWithDescriptor time. Previously that failure only surfaced via SKIA_LOG_E; now the NSError text goes through ShaderErrorHandler::compileError alongside the MSL compile path. This is exactly the #4555 failure mode: MSL compiled, PSO creation failed with CompilerError Code=2 "reading from a rendertarget is not supported", Recorder::snap() returned bare null.

3. Disable Graphite framebuffer fetch on the Metal simulator — MtlCaps enabled framebuffer fetch (sk_LastFragColor [[color(0)]]) for every Apple-family GPU, and Caps::getDstReadStrategy() then always picked kFramebufferFetch for dst reads. The iOS/tvOS simulator advertises an Apple GPU family but its Metal implementation cannot read from a render target in a fragment shader. Root cause of #4555: GradientBlend's Multiply-blended circle needs a dst read → picked framebuffer fetch → PSO rejected → null Snap. Ganesh Metal never uses framebuffer fetch (dst-copies instead) and real Apple devices/macOS support it, which is why every other cell of the matrix rendered fine. Fix: keep FB fetch off under TARGET_OS_SIMULATOR so dst reads fall back to DstReadStrategy::kTextureCopy, which the simulator handles correctly.

Diagnosed on the iOS simulator using commit 1's callback + commit 2's PSO forwarding — the chain of null-paths led directly to the FB-fetch PSO rejection.

SkiaSharp issue

Related to mono/SkiaSharp#4555

Required SkiaSharp PR

Requires mono/SkiaSharp#4586

Areas affected

• C API (include/c, src/c)
• Native dependency / DEPS
• Build (gn / build files)
• Upstream Skia merge or rebase
• Rendering output / behavior
• Other

Changes

C API

// new — shader/PSO compile error diagnostic callback
typedef void (*sk_graphite_shader_error_handler_proc)(
    void*       userData,
    const char* shader,
    const char* errors,
    bool        shaderWasCached);

// sk_graphite_context_options_t gains
void* fShaderErrorHandler;         // nullable; cast from proc typedef, stored as void*
void* fShaderErrorHandlerUserData;

Behavior

• Graphite Metal PSO-creation failures are now visible to installed shader-error handlers (previously invisible outside SkDebugf).
• iOS/tvOS simulator: Graphite dst reads fall back to kTextureCopy instead of kFramebufferFetch. Slightly more work per dst read, but correct — was previously silently broken for any non-coefficient blend needing a dst read (Multiply, Screen, Overlay, etc.).
• Real Apple devices and macOS unchanged.

Testing

Verified in the companion SkiaSharp PR — full SkiaSharp.Tests.Devices run on iPhone 16 / iOS 26.1 simulator (arm64): 1219 total, 0 failed, 1138 succeeded, 81 skipped, GradientBlend cell rendering and matching its golden.

Locally on Linux/Lavapipe (Vulkan; validates commit 1 wiring only — the Metal-specific commits 2/3 need Apple hardware): shader-error-handler smoke test passes, no double-free / no EntryPointNotFoundException on teardown.

Checklist

• Targets the skiasharp branch
• Changes above lists every added/changed C API export (or "None.")
• Companion mono/SkiaSharp PR linked above (submodule bump at minimum; regenerates bindings + adds tests for C API changes)