SU.749 PartCraft: large assemblies, measured first at 1,000, 5,000 and 20,000 parts, then fixed where it is slow

Asked (#749, Oct 16): managing large, complex assemblies, with Catia named as the reference. PartCraft's stress tests so far are 750 features in one part (D7.51) and closed-loop mates (D7.40); no row measures an assembly of thousands of parts.

Do: (1) generate the test assemblies ONCE under testfiles/generated/ (registry testfiles/LARGE-FILES.md, memory 'large-test-files'), from repeated instances of a few parts, not from files typed by hand; (2) measure open, rotate, select, mate-solve and save at each size in the real page, one browser test at a time with nice and timeout; (3) fix what the numbers show, in this order of likelihood: instancing (one mesh, many transforms), level of detail and a lightweight display mode, hiding by visibility sets, mate solving only for what moved, saving without rewriting unchanged parts. Record every before and after number in the log and in the release notes (memory 'test-results-in-release-notes').

Not claimed: Catia's feature set, any size without a measurement. Untested: everything, including the sizes above.

assembly-09 2026-11-06: Measured at 1,000, 5,000 and 20,000 copies of three small parts in the real page (headless Chromium on a CPU renderer, never a GPU), before and after; fixed: one draw group per part (40,000 draw calls to 3 at 20,000), a joint turned places only the parts behind it (5.0 s to 65 ms at 20,000), the part list shows its first rows (a click 11 s to 26 ms), boxes while the view moves (Views > Lightweight display keeps them on; a frame 15.7 s to 0.9 s on the CPU renderer), and assemblies open up to 50,000 copies (500 before). A frame in full detail is not faster on the CPU renderer; assemblies of heavy parts and a real GPU are not measured. No claim about any other program. Table and tests: docs/0.9-sessions/assembly-09.md.