The CSV export and the BricsCAD drawing sit in different coordinate systems. A
silent offset here puts every annotation in the wrong place while still looking
plausible, so this module derives the transform, proves it, and refuses when it
cannot.
Approach: vote for the translation over point pairs, refine the best candidates
by Procrustes fit (rotation included), pick the one with the most exact matches,
then report the residuals.
Two things the naive version got wrong, both fixed here:
Matching is constrained by role. The plant is built from regular grids - the
store lines, their infeeds and their outfeeds each form a row of 20 points at
800 mm pitch. Voting over all points is therefore ambiguous: a wrong offset hits
the infeed row exactly as well as the right one hits the separator row, and it
won by a single vote (21 to 20), producing an offset 8.6 m out with a spurious
0.52 degree rotation. Separators are now matched only against separator blocks
(S-SP, S-LP) and scanners only against SCAN, which makes the result unique.
The fit uses exact matches only (50 mm). For extended objects the CSV reports the
bounding-box centre while the block reports its insertion point; such pairs are
100-500 mm apart and would tilt the fit if they were included. They still count
towards the reported residuals.
For the Mubea plant this yields dx +50905.7 mm, dy +8414.5 mm, rotation exactly
0, from 20 exact matches - the same values a hand analysis produced.
A verified transform is stored in SKEL_CFG/dxf_registration.json per CSV/DXF
pair, so later runs are reproducible; a stored transform is still re-checked
against the drawing on every run.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
s.ayadi
7a8c25f3b9