[matmen]

This makes lossless WebP faster without changing the output. On a 1920×1080 photo:

	master	this PR
encode	1220 ms	735 ms
decode	95 ms	82 ms

Both builds ran alternately in the same run, because the local machine is noisy. On the runner, encoding took 621 ms in the 3.2.1 benchmark.

Changes (src/codecs/vp8l.ts)

• predictRun: Predictions now run with one loop per predictor mode, not through predict() with a switch for every pixel. When encoding it writes a block row's predictions into a buffer; when decoding it restores the pixels in place.
  • Encoding uses it to choose a mode per 16×16 block (14 modes per block), decoding uses it in inversePredictor.
  • predict() has no callers left and is removed.
  • A copy specialised for decoding (no restore check in the loop) measured no faster, so there is only the one function.
• Residual cost: looked up in a 256-entry table instead of four branches.
• tokenize: each pixel pair is hashed once into an array, instead of once for the search and again for the insert.
  • A larger hash table or a shorter chain made no measurable difference in speed or size.
  • Most of the remaining time is random memory access along the hash chains (about 17 candidates per search).

Verified

• The VP8L data is byte-identical to master for all 1298 corpus images that were encoded. The rest of the container can differ only by the EXIF timestamp.
• Decoded pixels are identical for all 153 WebP files in the corpus. Both builds reject the same 3 broken files.
• npm test passes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YJgRK5qpwJEbHPuFv6kgAQ