[performance] perf(harfbuzz): avoid defensive Feature[] copy in Font.Shape

[github-actions]

Description

HarfBuzzSharp.Font.Shape(Buffer, IReadOnlyList<Feature>, IReadOnlyList<string>) called features?.ToArray() unconditionally, allocating a fresh Feature[] copy on every shape call — even on the common Shape(Buffer, params Feature[]) path where the caller already owns a real array. hb_shape_full only reads the features (const hb_feature_t*), so the copy is pure managed waste.

This change pins the caller's array directly when it already is a Feature[], falling back to ToArray() only for non-array lists:

var featuresArray = features as Feature[] ?? features?.ToArray ();

The pinned pointer and count handed to native are identical either way, so shaping is unchanged.

Changes

• binding/HarfBuzzSharp/Font.cs — pin caller-owned Feature[] instead of copying.
• benchmarks/SkiaSharp.Benchmarks/Benchmarks/HarfBuzzShapeFeaturesBenchmark.cs — New-vs-Old proof.
• tests/Tests/HarfBuzzSharp/HBFontShapeFeaturesTest.cs — equivalence + sensitivity coverage.

Required skia PR

None.

Areas Affected

• HarfBuzzSharp (text shaping)

Testing

Proof 1 — BenchmarkDotNet (New vs Old, InProcessEmit, net10.0, x64, Linux):

FeatureCount	New (fix)	Old (ToArray)	Allocated New	Allocated Old
1	575.8 ns	610.3 ns	~0 B	72 B

~5.7% faster; per-call Feature[] allocation eliminated (72 B → 0 B). No allocation regression. Empirically measured on the run's hardware.

Proof 2 — equivalence test HBFontShapeFeaturesTest (5 tests, all pass): array vs non-array feature lists produce bit-identical GlyphInfos/GlyphPositions across empty / 1 / 3 features and the null-list case; a guard confirms the comparison catches a genuinely different result.

Built binding/HarfBuzzSharp (net10.0) and ran the equivalence tests via tests/SkiaSharp.Tests.Console — all green.

ABI

Body-only change to an existing method — ABI-safe. Renders identically before and after.

---

Produced by the performance-fixer agentic workflow + performance-fixer skill (managed-C# hot-path scan, text-and-fonts focus area). Measured numbers are empirical (net10.0/x64/Linux); ns figures vary by runner, the 72 B → 0 B allocation delta is the stable primary signal.

Fixes #4689

Generated by Fixer - Performance · opus48 · 693.3 AIC · ⌖ 37 AIC · ⊞ 13K · ◷