All three found on rose-pro and worked around in its generator; each deserves a minimal repro from the rose repo's commit history:
- Corner-radius cost is non-monotonic: same plate, whole-link eval: r=8 6.6s, r=12 5.4s, r=16 23.8s, r=20 5.2s, r=25 68.3s. Looks like coincident-tessellation-vertex degeneracy at particular radii. (rose_pro_loon.corner_radius carries the table.)
- Union of face-coincident plates: R_shank >400 s while its mirror L_shank took 10 s — float-coincidence in the kernel, not shape cost. rose-pro now emits one root per plate and never unions, purely to dodge this.
- 'Difference removed no volume from a subtrahend that overlaps the minuend — rejecting the B-rep result and re-cutting with the mesh boolean' fires on every skinned render; the fallback works but the B-rep path is losing these cuts.
🤖 Generated with Claude Code
All three found on rose-pro and worked around in its generator; each deserves a minimal repro from the rose repo's commit history:
🤖 Generated with Claude Code