Membership by midpoint/endpoint kept whole segments, so long table borders
and annotation rules carried the full sheet width into a floor — 13号's 56m
frame emitted 143m of geometry and footprint hulls grew stray plates.
Segments are now clipped to the frame rect; fully-inside geometry passes
through unchanged. Regression: a 140m line crossing a 40m frame comes out
frame-wide. Dense sheet layouts hit this; 贝昂's wide frame pitch never did.
Round-3 log: tower-1 twin towers aligned and gated (union-clip standard
floors; PH centered on cores, RF parapet bbox-anchored, wing offsets
self-consistent to 50mm); podium F1-F3 zoning still open.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Conversion/encoding chain (libredwg on Chinese TArch DWGs):
- run_dwg2dxf: capture bytes, decode leniently (GBK console output crashed
the batch under subprocess text=True)
- sanitize_dxf_comments: byte-level, code/value pair walk — the UTF-8
errors="ignore" round-trip mangled every GBK CJK string ("一层平面图"
-> "一平图") and eating XRECORD "999" VALUE lines corrupted files
- read_doc: recover fallback also on ValueError; scrub_surrogates keeps one
corrupt drawing from killing the whole report
- decode_mif in report + frame detector: \M+5XXXX escapes (whole 黄桥 set)
restored so plan-title detection works at all
- ODA auto-fallback: reconvert DWGs whose libredwg DXF fails inspection or
reads an empty modelspace (黄桥 plans: 0 entities via dwg2dxf, 21k via
ODA); env ODA_CONVERTER override, fake-converter regression test
Frame detection:
- detect_sheet_frames expands INSERTs (memoized, depth<=3) when collecting
axis/wall lines — 黄桥 wraps each frame's grid in anonymous A$C... blocks
(msp had 8 DOTE lines, blocks held 700); all 7 frames went from WEAK
clip-corner fallback to grid-aligned votes x=14
Extraction judgment rules (both loud, both default-on, both switchable):
- drop_isolated_clusters: tiny segment clusters far from the main mass are
sheet-legend content drawn on real layers (title-block window symbols made
every floor's footprint grow a sliver fin)
- drop_unsupported_roof_lines: roof outlines with no footprint/cap plate
within 2m beneath are hatch-only-support canopies that extrude into
floating ribbons (F3 SW 坡屋面, B1 ground canopy at -5.85m)
Tests: 64 -> 75 checks, incl. GBK/MIF synthetic DXFs, block-wrapped grid
alignment (non-uniform bays — uniform grids make subset correlation
ambiguous), legend-cluster and unsupported-roof fixtures.
references/generalization_test_log.md tracks the test protocol, per-project
gate table, and every gap with its regression coverage.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Agent-led CAD-drawing → 3D web showcase pipeline. Inspects CAD reports,
maps layers/floors, extracts a normalized building spec (with block-probed
glazing and audit gates), builds Blender/trimesh GLB variants, and assembles
a Three.js web viewer. Includes XREF-shell detection and a regression test.
Large inputs/outputs (work/, outputs/, source_cad/, renders) are gitignored.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>