Keeps the numbers next to the code rather than in a commit message, since the
adoption decisions turn on them. Two results are not what the plan assumed:
- face_surface_props gains *nothing* from leaving Python — the serial C++ loop
costs exactly what the Python loop costs. That agrees with the app's own
profile, which found SurfaceProperties to be 94 % of face_candidate_anchors:
the time is inside the kernel, so the interpreter round trip was never the
problem. The whole 10.9x is OSD_Parallel, reachable only because the GIL is
released.
- BinTools is already fast enough. Dropping the BytesIO detour saves nothing
measurable, so the pools' overhead is not in serialisation — worth knowing
before rewriting seventeen call sites for it.
Mesh extraction is the clear win at 35x, where the Python loop overhead really
does dominate; edge polylines 2.6x.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DfriM8XUkn7uYf5Dwe2xo6
