Bruno Postle
aaaaad2e53
operators.unfold_shared_leaves(): materialise each live shared leaf (share=k) into k distinct same-code sibling leaves, splitting its footprint into k equal-target children with squarest per-cut rotation (_size_subtree_equal), clearing the share stamp. Surgical — siblings' evolved geometry is untouched. Investigation result (harbor-house, evolved-3M sharing seed): - unfold alone (zero search) closes all 15 critical missing-room fails and lifts the canonical score 6.73e-29 -> 1.46e-19 (90->59 fails). - warm-starting a --no-leaf-sharing evolve from the unfolded seed runs ~7-8 orders of magnitude ahead of the naive (un-materialised) warm start at equal budget. The count deficit, not the sizing, was what stranded Schedule A deep in the fail hole. Adds test_unfold_shared_leaves_materialises_deficit. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01M8566xAxTnwtJTkpXjYNZm |
||
|---|---|---|
| .. | ||
| test_dom_corpus.py | ||
| test_driver.py | ||
| test_evolve_cli.py | ||
| test_fitness.py | ||
| test_genome.py | ||
| test_geometry.py | ||
| test_graph.py | ||
| test_innerloop.py | ||
| test_operators.py | ||
| test_oracle.py | ||
| test_staging.py | ||
| test_superposition.py | ||