Typed values were quantized to the f32 grid before reaching the
entity (e.g. a length of 2000.8 committed as 2000.8020), worsening
with magnitude. The interaction point layer was all glam::Vec3 (f32).
Two-layer fix:
1. Coordinate backbone -> f64. last_point, last_cursor_world,
dyn_anchor, dyn_ref become DVec3; UcsXform transforms, parse_coord,
the ucs helper fns, dyn_resolve_point, the OTRACK typed-distance
paths, command-line coordinate entry and the coordinate readout all
compute in f64. Screen/GPU/alignment stays f32 (snap engine, render
geometry, projection, UCS-icon/viewcube drawing) with casts only at
those boundaries.
2. Per-command committed storage -> DVec3. on_point already received a
DVec3, but many command impls narrowed it back to f32 (let pt =
pt.as_vec3(); Vec<Vec3> fields) and re-widened at build, quantizing
the committed coordinate. Migrated PLINE, SPLINE, MLINE, LENGTHEN,
INSERT, 3D primitives, every dimension/leader/mleader, ray, revcloud,
wipeout, attdef, table, tolerance, array center, mview, plot_window,
plus the QDIM_PLACE handler and qdim_collect_points. Preview,
hit-test and rubber-band paths keep f32 at the boundary.
Hatch manual boundaries are left for a follow-up: HatchModel.boundary
is structurally Arc<Vec<[f32;2]>> and needs a model-level widen.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
57710eec2b