BCL patterns: interop & marshalling

The techniques specific to a P/Invoke binding — reducing the marshalling the managed wrapper adds around a native call. The generated SkiaApi P/Invoke declarations in *.generated.cs are off-limits to hand-edit; the lever you control is the managed wrapper. Prove byte-for-byte parity for anything that reaches native (../measuring.md).

Keep managed signatures blittable and exact

• Do: keep wrapper signatures blittable (structs/spans passed directly), match the native C declaration exactly.
• Instead of: marshalled arrays/strings where a pinned span/pointer works; "simplifying" bool/char marshalling.
• Why: blittable types pin/pass directly instead of copying through a marshal buffer.
• Complexity: low–medium · TFM: any · ABI: internal.
• Watch out: never change bool/char marshalling without checking the native layout — C++ bool is 1 byte; default .NET bool marshals as 4.

Avoid temp strings/byte[] in text-encoding interop

• Do: on modern TFMs, encoding.GetByteCount(ReadOnlySpan<char>) then encode into a stackalloc/pooled Span<byte>.
• Instead of: appending a NUL by string concatenation and allocating an encoded byte[] per call (Util.GetEncodedText).
• Why: removes both the temp string and the array on per-text-call paths.
• Complexity: low–medium · TFM: better on net-modern; keep the byte[] path as old-TFM fallback · ABI: internal (don't change the return type unless all callers are updated).
• Watch out: GetEncodedText is not UTF-8-only — it switches on SKTextEncoding (Utf8→UTF8, Utf16→Unicode, Utf32→UTF32) with an addNull variant. Prove parity for every encoding AND both addNull cases — the NUL terminator is 1/2/4 bytes.

Cache delegates / prefer static callbacks

• Do: cache a static delegate for a native callback, or use delegate* unmanaged / [UnmanagedCallersOnly] on modern TFMs (retain the delegate fallback for old ones).
• Instead of: allocating a lambda per call on callback-heavy paths (managed streams, drawables).
• Why: avoids per-call delegate allocation and marshalling setup.
• Complexity: medium · TFM: function pointers net5+ · ABI: internal if the callback signature is unchanged.

LibraryImport source-gen marshalling (net7)

• Already conditional in this repo (USE_LIBRARY_IMPORT). It reduces stub overhead and helps AOT/trimming, but the declarations live in generated SkiaApi — never hand-edit them. Your lever is keeping the managed wrapper signatures blittable.

[SuppressGCTransition] (net5, advanced — usually not actionable here)

• Skips the GC-mode transition for tiny, non-blocking, non-callback, non-throwing leaf getters (width/height/sample-count). But it must annotate the P/Invoke declaration, which lives in the generated bindings — applying it needs a binding-generator/config change plus maintainer approval, not a hand-edit. Never on draw/encode/decode/stream/lock/callback calls. Treat as a proposal to file, not a quick fix.