fix(raster): colorBalance keeps luma at the clamp; pin mix formulas

[prakashUXtech]

Anatomy round 1, batch r1-b2 for tone-pack: the ColorMix fixes (blackAndWhite, channelMixer, colorBalance). Part of #432.

What changed and which test proves it:

• colorBalance keeps luma at the clamp (TP-4 contract). With preserveLuminosity the result was scaled by L/lo and then clamped, so a channel pushed above 1 lost luma: encoded [1, 1, 0.2] with highlights [0, 0, -100] lost about 6.6e-3. Now, when the scaled result has a channel above 1, the chroma is pulled toward the grey at L by the largest k in 0...1 that fits (out = L + (s - L) * k). Both s and L are non-negative, so only the upper bound limits k. The result's luma is L exactly. Proof: ColorMixTests.preserveKeepsLuma now covers inputs over the full 0...1 range, channels at 1, and that exact case. It fails before the fix (mutation G3: Scene commands #4 puts the old behaviour back and is caught).
• Non-finite errors name the field. checkFinite is now called once per field: "blackAndWhite yellows", "channelMixer green", "colorBalance highlights" and so on. Proof: ColorMixTests.nanNamesField.
• Docs. Each of the three doc comments now says "Out-of-range values are clamped; non-finite values throw RasterError.invalid". colorBalance's comment also describes the chroma pull. The clamp is pinned by GrayMix(reds: 500) == GrayMix(reds: 300) in betweenRangesAndTint.
• Zone weights and the negative branch. New test ColorMixTests.highlightsAndMidtonesByZone (with preserveLuminosity false) checks greys 0, 0.1, 0.5, 0.95 and 1 under highlights [100,0,0], midtones [100,0,0] and shadows [-100,0,0]. Each value is compared within 1e-5 to the documented formula.
• channelMixer formula. New test ColorMixTests.matrixRowsAndConstant: on a seeded opaque buffer, each output channel equals clamp(row·(r,g,b,1), 0, 1) within 1e-6, and a pixel whose sum is above 1 gives exactly 1.
• blackAndWhite interpolation, tint and weight clamp. New test ColorMixTests.betweenRangesAndTint covers three things. Hue-30 orange gives the interpolated grey 0.5. A tint on mid grey matches the W3C soft-light reference with r > g > b, and gives half the offset at tint alpha 0.5. The weight clamp is pinned as above.

Mutation entries added to Tests/mutations/CoreRaster.ColorMix.json: the luma k limit, the highlights smoothstep edge, the negative-shift direction, the channelMixer constant column and row order, the hue interpolation and the tint alpha. scripts/mutate --run --changed-since origin/dev reports 10/10 caught. I also checked by hand that the tests catch these mutants: the weight clamp removed, the output clamp removed, midtones dropped and the tint dropped.

No platform API is new here. The change uses only Swift stdlib SIMD operators that this file already uses, so I did not look anything up in the docs tools.

Belt-Lead: act-1a150bfbcf7-sfwy
Feature: tone-pack
Slice: TP-mix
Mutate: 10/10 CoreRaster.ColorMix
Fast-Check: pass
Head: 46a38b9

---

[github-actions]

Fast lane: merged into dev (198 lines, 3 files, Fast-Check marker).