perf: share preprocessing for async palette graphics

[2dubu]

Summary

Reuse decoding, sampling, and forward OKLCH conversion when async palette graphics request both a palette and semantic swatches with the automatic or CPU quantizer.

Background

A cache miss in AsyncPaletteGraphicLoader currently extracts a palette and then calls swatches, which repeats the image pipeline with at least 16 colors. With the default 10-color palette, the two quantizations are necessary to preserve their different results, but preprocessing the same source twice is redundant.

Solution

Add an internal paired extraction path with request-local prepared pixels. It preserves both quantizations at K and max(K, 16), including when K is already 16 or greater. The loader uses the pair and retains its cancellation check before caching or publishing success.

Custom and Metal quantizers keep the existing two-call path. Public API signatures, filtering, fallback, source color-space selection, and the first palette's timing boundary are preserved. A second extraction failure still produces a successful palette with nil swatches. The internal cancellation method returns the canceled task so tests can await completion deterministically.

Related Issue

N/A.

Verification

• Added ordered palette RGB/population/proportion and role-by-role swatch equivalence checks over K=2/10/16/32, all output color spaces, sRGB/P3 sources, and automatic/CPU quantizers. Also cover encoded input, resizing/sampling/filter options, fallback, invalid options, pre-cancellation, custom call order, and second-call failure.
• Added UIKit loader tests for failure/cache behavior and cancellation during a suspended second custom quantization. The optimized CPU path's cancellation boundaries are additionally source-reviewed; the suspension test intentionally uses a controllable custom quantizer.
• Commands run:
  • swift package resolve
  • swift build -c debug
  • swift test --filter PaletteKitTests --filter PaletteKitInsightsTests — 67 tests passed.
  • swift test -c release --filter SharedExtractionTests — 8 tests passed.
  • xcodebuild test -scheme PaletteKit-Package -only-testing:PaletteKitTests -only-testing:PaletteKitInsightsTests -destination 'id=5F0334CA-5AD2-47FE-B837-F802AFFB20FD' — 164 tests passed on iOS 26.5 Simulator (155 core and 9 Insights).
  • Release SharedExtractionBenchmarks/testSeparatePaletteAndSwatches and testSharedPaletteAndSwatches, three process runs each with order alternated.
  • git diff --check
• An additional temporary integration of all three independent PRs passed 71 macOS tests and 168 iOS Simulator tests.
• Environment: Apple M5 Max, macOS 26.6.2, Swift 6.3.2. iOS compilation reports existing existential/CIContext warnings in unchanged renderer files and generated App Intents metadata warnings; they are outside this change.

Compatibility and Impact

• Public API: None.
• Behavioral impact: The paired CPU path reuses one source's preprocessing while preserving the palette and richer swatch results. Independent public calls and custom/Metal extraction retain their existing selection and error semantics.
• Performance impact: Removes duplicate preprocessing in the paired CPU path. Prepared buffers remain alive through the pair; peak memory and iPhone performance have not been measured. No memory-reduction claim is made.
• Affected platforms: Shared extraction logic on iOS/macOS; the async graphic loader is UIKit-backed.

Additional Context

Release benchmark: a reused 1024×1024 deterministic noise CGImage, K=10, OKLCH, CPU, default downsampling. The same full output checksum consumes both results; input generation is outside measurement. Fifteen clock samples per path from three process runs:

Palette plus swatches	Median
Separate extraction	15.955 ms
Shared preprocessing	8.555 ms

This is 46.4% lower latency for this Mac fixture, not an iPhone or all-image improvement claim. Both paths are measured from this branch; a separate base-versus-head standalone API control uses three JPEG fixtures, gradient, noise, and transparent fallback, with collectTimings: true and 45 samples per input. Ordered output signatures match, and median changes range from -0.5% to +1.1%; these measurements do not establish zero overhead or peak-memory equivalence.

The benchmark target also provides testStandalonePalette. This branch is based directly on main and is independent of the LUT and MMCQ cache PRs.

Checklist

• This pull request is focused and contains no unrelated changes.
• Tests were added or updated where behavior changed, or I explained why they are not needed.
• Regression tests cover the intended behavior or root cause rather than only a reported sample.
• Public API changes are documented and reflected in the changelog, or this pull request has none.
• I considered source compatibility, behavioral changes, and performance impact.