Pattern tiles were sampled nearest-cell, exact when a pattern cell
(0.75 pt) is a whole number of output pixels (a 960-wide export of a
720 pt slide) but harsh otherwise: at 960 px for a 960 pt slide (0.75 px
per cell) PowerPoint shows greys where a pixel straddles two cells, and
DeckCraft showed only the two colours, in different places. The tile is
now sampled bilinearly. Fitted against PowerPoint's export of all 54
presets at 0.75 px per cell: mean error 0.067 (bilinear) against 0.082
(nearest) and 0.095 (area average); still not exact (narVert reads
255/128/0 in PowerPoint, 213/128/42 here). Unchanged at whole-pixel sizes.
Wild synthetic set: patterned backgrounds 21.15 -> 11.31 and
18.00 -> 9.75, pattern galleries 22.50 -> 19.44 and 13.96 -> 11.64.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Pattern fills were approximated (dots, bars or a grid per family, on an
8-point cell), so pct30 looked like sparse dots and wave like a dot grid.
Every one of the 54 DrawingML presets now draws its 8x8 tile, each cell one
96-dpi pixel (0.75 pt), laid out from the slide's origin. The tiles were
read from PowerPoint's own export of each preset at two output pixels per
cell; the pattern probe deck now matches it to a mean pixel difference of
0.2 (from 62).
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
