[matmen]

Adds RotateOptions.fast, on by default like ResizeOptions.fast. Rotations by angles that aren't multiples of 90° now sample with 8-bit weights on 8-bit premultiplied pixels, so red+blue and green+alpha each share one multiplication. fast: false keeps the floating-point sampling.

Speed (1920×1080 photo rotated by 30°)

	fast: false	fast (default)
opaque	119 ms	51 ms
with alpha	114 ms	61 ms

How it works (src/util/rotate.ts)

• Transparent border: The source is copied with a transparent border one pixel wide.
  • Bilinear sampling reads the 4 neighbours of each source position. Neighbours outside the image now land on transparent pixels, so no bounds checks are needed.
  • Edges fade out exactly as before, when out-of-image taps were skipped.
  • The border exists only in the copy; the output size is computed from the original dimensions, as before.
• Weights: The 4 weights come from 8-bit fractions and are adjusted to sum to exactly 256, so a lane never overflows.
• Row spans: Each output row only visits the span that maps into the source (span()). At 30°, about 40% of the output is empty corners.
• Shared helpers: It uses the helpers from src/util/alpha.ts. unpremultiply8 now writes fully transparent pixels directly instead of doing three table lookups; the result is identical.

Precision

Measured over the corpus (1298 images, 4 angles), fast against fast: false:

• Opaque pixels are within 1. Only 0.5–2% of all values differ.
• Translucent pixels, including the faded edges of opaque images, differ by at most about 2.5 steps where they show (channel × alpha).
• Alpha differs by at most 2.
• tests/rotate.js checks these limits on a photo crop, with and without an alpha gradient.
• The rotate-45 targets are regenerated. They differ from the old ones by at most 1.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YJgRK5qpwJEbHPuFv6kgAQ