pub const FACETED_BREP_DEDUP_FACE_LIMIT: usize = 20_000;Expand description
Face-count threshold above which [faceted_brep_face_count] tells the
caller to skip the structural signature walk entirely (#1909). The walk in
[try_faceted_brep_signature] mirrors the mesher’s own traversal
face-by-face, bound-by-bound, point-by-point — for dedup to pay off there
must be at least one OTHER item sharing this exact signature. Tekla-style
duplication (thousands of connection plates/bolts, the case this cache
exists for) tops out at low hundreds of faces per part; a SINGLE huge
faceted import — a terrain, a scanned mesh, one big BREP with no sibling —
has none, so hashing every point is pure loss that DOUBLES the traversal
cost of an already-expensive mesh with no possible payback. Set well above
any plausible detail-part face count and well below the scale where a
single giant mesh becomes measurably worse — a 2.5 M-triangle model
reported ~30 s where web-ifc took ~2.85 s, traced to this hash.
pub (re-exported through router and the crate root — see lib.rs,
same pattern as GEOMETRY_DIAGNOSTICS_SCHEMA_VERSION) rather than
pub(super)/pub(crate), specifically so the #1909 regression test
(tests/issue_1909_large_single_brep_dedup.rs, an external integration
test crate that only sees this crate’s public API) reads the real
threshold instead of hard-coding 20_000 a second time (PR #1977
review: that duplication is exactly the shape of the 341/88 checklist
drift on #1963 — assertion and constant silently disagreeing).