[fMindstormer]

Whole-program search for 0.7.0: the re-search now compares finished programs instead of committing to one shape at a time.

What changes

• Redraw a color whose region falls into parts (search.parts, search.add_parts, a new improve neighborhood). The greedy search takes one shape over a whole region (two circles stacked on an oval become one tall box plus patches) because that shape fixes the most pixels at its step. The new step splits the region where its distance to the edge comes apart, climbs one shape per part against that part alone, refills, and keeps the result only if the finished program scores better.
• Re-search the 3 best results, not only the best (compiler.RESEARCH = 3, --research N). They run in parallel when there are several processes. --research 1 keeps the old behavior. The browser demo passes its checkbox as on/off, so it still re-searches one result.
• Code generation: inset drops the lengths CSS repeats (inset:100px), without breaking a padding that elements share.
• Benchmarks:
  • bench/run.py --research N and a matching input on the Benchmark workflow;
  • ablation rows no-parts and research-1;
  • evaluate.py methods v0.6 / v0.6-N, for the held-out baseline at equal compute;
  • a Held-out workflow.
• Version 0.7.0: committed Linux results, ablation, heldout4 programs and known optima. The README, method notes, blog and results page are updated.

Results

The 66 development targets in Chromium (Benchmark workflow):

	Total J	Characters	≥ 99% match	Shorter than human
0.6.0	52,153	23,329	66 / 66	56 / 66
0.7.0	47,263 (−9.4%)	21,175	66 / 66	60 / 66

The lowest match stays at 99.21%, and the mean goes from 99.82% to 99.84%.

Ablation (cssynth's own scores):

• switching the whole re-search off now costs 14.4%;
• the parts step alone, 6.1%;
• re-searching only the best result, 2.4%.

heldout4, spent once on this version, with all methods on one runner:

	Solved	Characters, on the 43 all solve	CPU
0.6.0's search	43 / 48	16,722	4,667 s
0.6.0's search, 48 starts (equal compute)	43 / 48	16,728	9,143 s
0.7.0	43 / 48	14,696 (−12.1%; 95% interval −22.7% to −1.4%)	8,141 s

Known optima: unchanged; optimal on 33 of the 35 reproduced exactly.

Cost: a compile takes 1.5 to 3 times the CPU of 0.6.0, depending on the picture. With several processes the wall time grows much less (the lotus: 16.9 s against 17.8 s on 8 cores).

After merging

The Pages workflow publishes the results page, blog and demo from main. Then tag v0.7.0 on main and create the GitHub release with the wheel, sdist, examples, benchmark zip and checksums. No PyPI.

🤖 Generated with Claude Code