[prakashUXtech]

What

AutoTone.suggest(for:) replaces the TP-0 stub in Sources/CoreRaster/Tone/AutoTone.swift. It returns a white balance (temperature, tint) and linear levels (inBlack, inWhite, gamma). The suggestion is applied as whiteBalance and then levels with output 0...1.

• White balance: computed from the mean straight linear rgb of the brightest 5% of pixels by luma, counting only pixels with no channel at 0 or 1. The mean goes back through whiteBalance's own gain model (gains 2^(T/200), 2^(−M/200), 2^(−T/200)), which gives T = 100·log2(b/r) and M = 100·log2(g²/(r·b)), each clamped to −100...100. Applying it makes that mean grey.
• Levels: read from the Rec. 709 luma Histogram (4096 bins, linear) of the white-balanced buffer. Black and white points sit at 0.1% and 99.9%. Gamma = ln t / ln 0.5 with t = (median − black) / span, which puts the median at 0.5, and is clamped to 0.1...10.
• AutoToneTests is turned on (the three .disabled("waits on TP-6") traits are removed and no assertion is changed). The autoToneStubThrowsNotBuilt row is removed from ToneStubTests.swift.
• Tests/mutations/CoreRaster.AutoTone.json holds the packet's two mutations.

Why

TP-6 of the tone-pack packet (slice TP-map): auto tone suggests levels and a white balance that neutralise a cast.

Packet ambiguities resolved

• Order: the white balance is computed first, from the input. The levels are then read from the white-balanced buffer (by reusing whiteBalance itself) because the tests apply them in that order.
• "Non-clipped": a pixel with any channel at or below 0, or at or above 1, is left out of the bright set. This keeps blown highlights and pure black out of the cast estimate.
• Neutral suggestions (documented in the doc comment; suggest never throws or returns NaN):
  • Temperature and tint are 0 when the bright set is empty (a black, blown-out or fully transparent buffer).
  • Levels are 0, 1, 1 when nothing is counted or when the black and white points are less than 1/256 apart (a flat image).
  • Gamma is 1 when the median falls outside the open span.
• Quantile precision: each quantile is the mean of the rank read from the bottom and from the top of the histogram, interpolated within its bin. With one-sided bin reads, the ramp's median lands 2/4096 low and gamma misses the 1e-3 bound. The symmetric read puts it at exactly 0.5.
• Helpers: none of the ColorMix/ColorAdjust helpers (encoded/decoded/checkFinite/mappingStraight) are needed. Auto tone reads linear values and maps no pixels itself. It reuses Histogram, Histogram.luminanceWeights and PixelBuffer.whiteBalance.
• ToneStubTests.swift: autoToneHoldsItsValues is kept. It tests the struct's memberwise init, which TP-6 does not outgrow. Only the stub's "throws not built" row is removed.

Proof

• AutoToneTests (castNeutralised, rampIsIdentity, blackSuggestsIdentity) and autoToneHoldsItsValues pass.
• scripts/mutate --validate: OK, 664 anchors resolve exactly once.
• scripts/mutate --run --plan CoreRaster.AutoTone: 2/2 caught.
  • Clip 1% instead of 0.1% is caught by rampIsIdentity.
  • Dividing by an empty bright set (guard count > 0 changed to >= 0) is caught by blackSuggestsIdentity.
• scripts/ci: CI: PASS on this head after rebasing on origin/dev. Earlier attempts failed only on wall-clock budget tests under shared-container load (ThumbnailStoreRecoveryTests.fuzzedIndexThrowsOrRecovers), and the load-guarded retry passed.

Docs MCP: checked Sequence.reversed() (returns an Array) and the SIMD3 min()/max() reductions with apple-docs. The math functions (log, log2) come from Glibc/Darwin, as in ColorAdjust.swift.

Part of #432

Belt-Lead: act-1a150bfbcf7-sfwy
Feature: tone-pack
Slice: TP-map
Mutate: 2/2 CoreRaster.AutoTone
Fast-Check: pass
Head: 1e2457c

---

[github-actions]

Fast lane: merged into dev (139 lines, 4 files, Fast-Check marker).