Found building hardware/rose-pro. Unioning a link's sheet plates into one solid — where lapped and stacked plates touch face-to-face and fold corners overlap by exactly one thickness — sent R_shank past 400 s while the mirror-image L_shank (same plates, y negated) took 10 s. Same topology, same dimensions, only the sign of y differs, so this is a float-coincidence degeneracy in the coincident-face union path, not a shape cost.
Workaround in the generator (rose/pro/rose_pro_loon.py): emit one root per plate and never union plates. That is also what a STEP export / BOM wants, so it stuck, but the asymmetry is a kernel bug worth a torture case: two plates sharing a full face (lap), two plates sharing an edge strip (fold corner overlap of one thickness), at both y signs.
The plates are built with hardware/rose-pro/plates.loon; the old unioned-link form can be reconstructed from the per-link groups in rose-pro.loon.
Found building
hardware/rose-pro. Unioning a link's sheet plates into one solid — where lapped and stacked plates touch face-to-face and fold corners overlap by exactly one thickness — sentR_shankpast 400 s while the mirror-imageL_shank(same plates, y negated) took 10 s. Same topology, same dimensions, only the sign of y differs, so this is a float-coincidence degeneracy in the coincident-face union path, not a shape cost.Workaround in the generator (
rose/pro/rose_pro_loon.py): emit one root per plate and never union plates. That is also what a STEP export / BOM wants, so it stuck, but the asymmetry is a kernel bug worth a torture case: two plates sharing a full face (lap), two plates sharing an edge strip (fold corner overlap of one thickness), at both y signs.The plates are built with
hardware/rose-pro/plates.loon; the old unioned-link form can be reconstructed from the per-link groups inrose-pro.loon.