The background voted on the pixel grid: 309 sprites un-chunked at no cost

Oliver put two pictures side by side — z_rufall3 as the fam_ru sheet drew it,
and z_rufall3 as it shipped — and they were not even close. The sheets were
never the problem: every one of the eleven family rolls came back FINE, and the
pipeline pixelated seven of them on the way down.

_fine.block_size() scores b x b blocks by uniformity, and it scored ALL of
them. key() zeroes the background, a zeroed block is uniform at every b, and on
a 9-row family sheet the background is ~85% of the bounding box — so the
detector's 0.82 bar was cleared by the background alone, at any block size it
tried. Measured on the ru sheet: 0.0% of ink blocks uniform at b=11, 99.2% of
empty ones, overall 0.821. Detected block: 11. _zsheet then honestly snapped
428 px figures down to 39 art px and scaled them back up — the pixelation was
MANUFACTURED there, not drawn. br survived at 0.775 (its ink coverage is 4
points higher) and the whole per-family chunkiness table — br fine, ch/ru/sh/st
at 6 px blocks — was ink-coverage noise around one threshold.

The fix is one line where every caller routes through: background-only blocks
get no vote. Then every family re-cut from the sheets already on disk — no new
rolls, no tokens — and all eleven came down at art block 1. br's re-cut is the
regression check: byte-identical result. Dimensions drift at most 3 px, so
nothing on court changes size. SPRV 112 -> 113 so phones drop the year-long
cache on the chunky art.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

assets/_fine.py

@@ -68,6 +68,12 @@ def block_size(a):
         for y in range(0, h - b, b):
             for x in range(0, w - b, b):
                 blk = sub[y:y+b, x:x+b].reshape(-1, 4)
+                # ONLY BLOCKS THAT TOUCH THE ART GET A VOTE. key() zeroes the background, so an
+                # empty block is uniform at EVERY b — and on a 9-row family sheet the background is
+                # ~85% of the bounding box, so it alone cleared the 0.82 bar and reported b=11 for
+                # sheets whose figures were drawn fine. _zsheet then snapped those figures down 11x
+                # and back up: that is where the pixelated families came from, not the model.
+                if not blk[:, 3].any(): continue
                 tot += 1
                 if (blk == blk[0]).all(): ok += 1
         if tot and ok / tot > 0.82: best = b