[prakashUXtech]

Anatomy round 1, batch r1-b1 for affine-composite: the review findings on transformed, padded and the packet. Part of #434.

Fixes

1. Overshoot error says what to do. The non-finite output error now reads "transform by with : lanczos3 overshoot made a component non-finite; use .bilinear or keep components finite". Proven by AffineOpsTests.overshootThrows, which checks that the message contains "lanczos3".
2. padded doc states its placement rule. It grows with transparent pixels and shrinks by cropping, and the top-left moves by the size change times the anchor's place (0, 0.5 or 1), rounded down. CanvasSizeTests.centreRoundsDown already pins this; the change is docs only.
3. Packet pseudocode matches the API as built. resampled is now resized(width:height:filter:). The scale-only separable fast path is dropped, with the reason: resized reads clamp-to-edge and transformed reads transparent outside. The packet also says the kernel widens for any scale below 1. Docs only.
4. Downscale widening is tested. New AffineOpsTests.downscaleWidensKernel: a 0.25 shrink of one-pixel stripes (columns, then rows), placed so each output centre lands on a source centre, is flat 0.5 grey within 1e-3 for bilinear and lanczos3. Stripes rather than a checkerboard, because a checkerboard averages to grey through either axis alone, so forcing stretchX to 1 escaped. New mutation entry: let stretchX = 1.0.
5. Filtered values are normalised means and fade at the edge. New AffineOpsTests.halfPixelTranslationBlends: a bilinear 0.5 px shift of a seeded opaque layer gives alpha 0.5 and half the colour in column 0, and two-neighbour means inside (within 1e-5). Lanczos3 on a uniform opaque layer stays alpha 1 inside. New mutation entry: dropping / total.
6. Huge whole translation is exercised. New AffineOpsTests.hugeWholeTranslationIsClear: translations of (1e300, -1e300) and (Int.max x 4, 0) return a clear 10 x 8 buffer. New mutation entry deletes the abs(m.tx) <= limit, abs(m.ty) <= limit clause; with it removed, Int(m.tx) traps and the runner counts the crash as caught.
7. Overshoot throw is exercised. New AffineOpsTests.overshootThrows. Columns follow a period-4 +max, +max, -max, -max pattern of Float.greatestFiniteMagnitude, which matches lanczos3's lobes; plain alternation cancels at Nyquist and cannot overflow. Through a 0.3 px shift they overshoot by about 1.37x and throw RasterError.invalid. New mutation entry removes the finite guard.
8. Kernel stretch is capped per axis. stretchX is capped at width and stretchY at height; before, both were capped at max(width, height). New AffineOpsTests.extremeShrinkOnNarrowAxisIsBounded: a seeded 2 x 20000 layer through a = 1e-7 and through a = 1e-3 (lanczos3, 4 x 20000 output) gives equal results in about 0.15 s. Optional mutation entry added (swap back to max(width, height)); the runner catches it in about 40 s.

Checks

• scripts/mutate --run --changed-since origin/dev: 11/11 caught. That is 9/9 in CoreRaster.AffineTransform (4 existing, 5 new) and 2/2 in CoreRaster.CanvasSize.
• scripts/ci prints CI: PASS on this head, after rebasing on origin/dev. Before the rebase, two runs failed on an unrelated wall-clock budget, CoreAssetTests.ThumbnailStoreRecoveryTests/fuzzedIndexThrowsOrRecovers (1.5 to 2.4 s against a 1 s budget). The machine's load average was 8 to 13 on 4 cores. The test takes 0.3 s when run alone, and this diff does not touch CoreAsset.
• APIs: Swift stdlib only (Float.greatestFiniteMagnitude, Double.squareRoot(), isMultiple(of:)). No platform or library API is new; I did not check the docs MCP for these.
• I did not touch the AC-5 files (CompositeAffine.swift, AffineStubTests.swift).
• Branch is lead/affine-composite-r1-b1: the existing lead/affine-composite ref blocks a lead/affine-composite/r1-b1 ref.

Part of #434

Belt-Lead: act-1a150bfbcf7-sfwy
Feature: affine-composite
Slice: AC-geo
Mutate: 9/9 CoreRaster.AffineTransform
Fast-Check: pass
Head: 3a1bc68

[github-actions]

Fast lane: merged into dev (176 lines, 5 files, Fast-Check marker).