[matmen]

This PR cleans up the fast resampling from #71 and the 8-bit alpha code it duplicated from blur().

Changes

• One fast path: resampleOpaqueFast and resampleTransparentFast are merged into resampleFast.
  • The red+blue and green+alpha lanes work for opaque pixels as well. Alpha stays exactly 255, because the weights sum to 256.
  • The function is split into verticalFast and horizontalFast.
  • Opacity is still detected without an extra scan: first the first and last rows are checked, then the vertical pass ANDs the pixels together. If it finds transparency in the middle of the image, the vertical pass runs again on premultiplied pixels.
• Shared helpers in src/util/alpha.ts:
  • isOpaque, premultiply8, unpremultiply8, and a single littleEndian flag.
  • Used by both resample and blur.
  • unpremultiply8 uses a 64K lookup table that rounds like the clamped arrays, so blur() gives the same results as before.
• blur()'s 8-bit path now explicitly requires little endian, because the helpers work on 32-bit words. On big-endian machines it falls back to the float path.

Verified

• blur() is bit-identical to master on 1313 corpus images, 5 sigmas each, with and without alpha.
• Resizing matches master except where unpremultiplying lands exactly on .5. That affects 0.003–0.007% of the values, by 1, and only in the unreleased fast path for transparent images.
• tests/resize.js adds an image whose only transparent pixel is in the middle, which exercises the rerun path.

Speed (resize() of a 1920×1080 photo, master → this PR)

	opaque	alpha 200
halving	12.8 → 12.6 ms	24.0 → 24.2 ms
doubling	63.2 → 63.9 ms	117 → 103 ms

🤖 Generated with Claude Code

https://claude.ai/code/session_01YJgRK5qpwJEbHPuFv6kgAQ