Generalization hardening from 27号 (rounds 5-8): 6 gap fixes, all TDD
- detect_sheet_frames: auxiliary-titled frames (非机动坡道/无障碍改造…) never become the alignment reference; wall-profile correlation fallback when a frame has no axis geometry (coverage+peak-ratio gated); bogus x-pitch from stray same-row titles distrusted (pitch_x < 0.7*pitch_y); plan titles inside cached-xref blocks scanned too - cad_dxf_report: cached_xref_content warning (msp INSERT stubs + in-file block cache -> expand_blocks ≈ bind, not a re-export ask) - derive_showcase_spec: floor_order gaps warn loudly in both directions - assemble_campus: per-building rotation_deg overrides the global one - SKILL.md updated; regression suite +14 checks (auxref/nogrid/mixedrow/ blocktitle/cachedxref/floor_order/rotation scenarios) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
10
SKILL.md
10
SKILL.md
@@ -52,7 +52,7 @@ The Agent owns interpretation. Scripts are tools for evidence and repeatable exe
|
||||
|
||||
The pipeline is designed so a first pass either succeeds or STOPS at a loud gate. Do not hand over a model unless every line below has evidence (a report line, a gate status, or an image):
|
||||
|
||||
1. Read report clean: every `Warnings` entry resolved. `multi_frame_sheet` -> ran `detect_sheet_frames.py` and used its clip/offset config; `likely_xref_shell` -> user re-exported with XREFs bound.
|
||||
1. Read report clean: every `Warnings` entry resolved. `multi_frame_sheet` -> ran `detect_sheet_frames.py` and used its clip/offset config; `cached_xref_content` -> the XREF content is CACHED as in-file block definitions, expand blocks (≈bind) instead of asking for a re-export; `likely_xref_shell` (no cached content) -> user re-exported with XREFs bound.
|
||||
2. Frame alignment proven: detect report shows `grid-aligned, votes x>=3 y>=3` per floor (any `WEAK grid match` floor was verified on its overlay SVG).
|
||||
3. Audit exit code 0, and gates `floors_aligned` / `glazing_captured` / `no_annotation_layers_classified` all PASS. Never pass `--no-strict-gates` to proceed.
|
||||
4. No `WARNING topless_group` from derive (roof handled via caps or `roof_specs`), and no `WARNING mesh_from_polygons` from the GLB build (else slabs/roofs silently vanished — install mapbox_earcut).
|
||||
@@ -76,13 +76,13 @@ The pipeline is designed so a first pass either succeeds or STOPS at a loud gate
|
||||
- Install missing parser dependencies into the current project work area, not globally:
|
||||
`python3 -m pip install --target work/vendor ezdxf`
|
||||
- Inspect the Markdown/JSON report before modeling. Confirm drawings, units, bounding boxes, and the top layers for walls/glazing/roof/columns/beams/railings/doors/detail lines/floor lines/grid.
|
||||
- Resolve the report's `Warnings` section before modeling. `likely_xref_shell` / `near_empty` means a sheet's real geometry lives in unbound XREFs and will silently become a missing floor: ask the user to bind/insert the XREFs in CAD and re-export that DXF (do not model the shell and report the floor as done).
|
||||
- Resolve the report's `Warnings` section before modeling. `cached_xref_content` means the "empty" sheet's INSERT stubs reference block definitions cached INSIDE the file (AutoCAD keeps unbound-XREF content as blocks; both libredwg and ODA export the cache) — expand blocks in detect/derive (`expand_blocks` ≈ bind), no re-export needed. Only `likely_xref_shell` / `near_empty` WITHOUT cached content means geometry truly lives outside: ask the user to bind/insert the XREFs in CAD and re-export that DXF (do not model the shell and report the floor as done).
|
||||
|
||||
3. Build the extraction config and `building_spec.json`.
|
||||
- Use `references/pipeline.md` and `references/building_spec_schema.md`.
|
||||
- Map CAD layers to normalized categories: `walls`, `glazing`, `roof`, `columns`, `beams`, `railings`, `doors`, `detail_lines`, `floor_lines`, and `grid`.
|
||||
- Identify one drawing per floor or split sheets that contain multiple floor plans side by side.
|
||||
- MULTI-FRAME SHEETS (the common Chinese pattern: one DWG per building holding ALL its plans + elevations + sections stacked in modelspace): run `scripts/detect_sheet_frames.py` on the sanitized DXF. It finds `XX平面图` title texts (filtering catalog rows and title-block copies), proposes a wall-bbox-refined `clip` per floor frame, and computes per-floor `offset`s by voting-based grid-line correlation (label-free; TArch grids live on layer `DOTE`, not `AXIS`). Never model such a file as one floor (all sheets superimpose) and never stack frames without offsets (floors staircase away diagonally — the classic broken output). Read its report: `WEAK grid match` floors need overlay verification; elevation/section frames are listed for floor heights (`z`/`height` from DIM_ELEV 标高 texts like `8.100`), not as floors.
|
||||
- MULTI-FRAME SHEETS (the common Chinese pattern: one DWG per building holding ALL its plans + elevations + sections stacked in modelspace): run `scripts/detect_sheet_frames.py` on the sanitized DXF. It finds `XX平面图` title texts (filtering catalog rows and title-block copies; titles INSIDE cached-xref blocks are scanned too), proposes a wall-bbox-refined `clip` per floor frame, and computes per-floor `offset`s by voting-based grid-line correlation (label-free; TArch grids live on layer `DOTE`, not `AXIS`). When a frame has NO axis geometry at all, it falls back to wall-profile correlation (length-weighted wall-position histograms, x and y independently, coverage+peak-ratio gated) before surrendering to the clip-corner anchor — clip corners track wall bboxes and drift by whole bays when balconies differ. Auxiliary-titled frames (非机动坡道/无障碍改造/人防…) are flagged and never become the alignment reference (a bike-ramp frame once anchored a whole tower to nothing). Never model such a file as one floor (all sheets superimpose) and never stack frames without offsets (floors staircase away diagonally — the classic broken output). Read its report: `WEAK grid match` floors need overlay verification; `wall-correlated` floors deserve one overlay check too; elevation/section frames are listed for floor heights (`z`/`height` from DIM_ELEV 标高 texts like `8.100`), not as floors. Frame clips can swallow the sheet above/below (refine bbox includes between-sheet strips): if a floor's footprint depth is ~2-4x its siblings, cluster its wall y-midpoints and cut the clip band to the plan's own row.
|
||||
- Align floors by structural grid or repeated control axes. Use millimeters-to-meters scale (`0.001`) unless the report proves the CAD is already in meters.
|
||||
- Floor keys: `floor_sort_key` treats a leading `B` as basement (sorted below ground). Do not invent keys like `B1F` for "Building 1 Floor" — use per-building specs instead, and set explicit `floor_order` whenever keys are nonstandard.
|
||||
- When CAD linework has enough exterior/interior boundary signal, enable generated footprints for floor slabs and roof caps. Prefer footprint polygons over rectangular floor bounds for visual/digital-twin presentation.
|
||||
@@ -129,11 +129,13 @@ The pipeline is designed so a first pass either succeeds or STOPS at a loud gate
|
||||
A folder like `建筑-1#厂房 / 2#厂房 / 3#辅房 / 总图` is a CAMPUS: several buildings plus a site plan — NOT floors of one building. Never map "one DWG = one floor" here.
|
||||
|
||||
- Model each building separately: one detect-frames pass, one extraction config, one `building_spec.json`/GLB per building DWG. Small structures (机械停车坑/水池) are usually one or two frames.
|
||||
- Assemble with `scripts/assemble_campus.py`: record each building's transform (model anchor -> surveying anchor + campus rotation) in `site_manifest.json` and let the script do the merge — it also runs the mandatory neighbour-gap check automatically.
|
||||
- Assemble with `scripts/assemble_campus.py`: record each building's transform (model anchor -> surveying anchor + campus rotation) in `site_manifest.json` and let the script do the merge — it also runs the mandatory neighbour-gap check automatically. Buildings on one campus can face different ways (27号: 4#楼 sits ~27° off its siblings): a per-building `rotation_deg` in the manifest overrides the global one.
|
||||
- Units differ by sheet type: building plans are almost always millimeters (`scale 0.001`), while 总图 site plans are often METERS at real surveying coordinates (huge x/y like 349xxx/663xxx, texts ~2-4 units tall). The extraction config has a single global `scale`, so never mix a meters site plan and mm building plans in one spec.
|
||||
- Place buildings with a site manifest: from 总图/定位图 (prefer a "带轴号(可点坐标)" version if present) measure each building's anchor point and rotation, then assemble per-building GLBs in the viewer/scene with those transforms. Record the manifest (building -> position/rotation/source) in the notes file.
|
||||
- Proven placement recipe (贝昂 case): find the building's envelope rectangle in the site drawing (layer 面积核算/建筑轮廓线) — its edge lengths must equal the model's wall-envelope bbox 1:1 (here 90.2 x 42.2 m exactly), edge angle gives rotation (7.57°), the corner matching the model's (min x, min y) gives translation. Resolve the 180° flip by densest-feature voting (e.g. machine-room walls inside a known sub-rect). Then EITHER merge at spec level (transform all segments/footprints/caps into site-local meters and append a SITE context floor) and export one GLB, OR keep per-building GLBs plus transforms in the manifest. Pick a site-local origin (用地红线 min corner), never raw surveying coordinates — float32 breaks at 349 km.
|
||||
- Anchor placement on WALL evidence, never roof/eave envelopes: 屋面 outlines include canopies and asymmetric overhangs — concentric placement against one put 2# 0.45 m off (26 cm INSIDE its neighbour). Best anchors, in order: corner coordinate callouts (X=/Y= texts), wall lines, envelope rects whose edges equal the wall bbox exactly.
|
||||
- Site plans without per-building envelope rects (27号 pattern): building outlines may hide on odd layers (EVTR!) — match each closed polyline to its nearest 楼栋 label, then register the model footprint onto the outline by point-to-edge ICP (brute-force angle + translation refine, trim worst 10%). Constrain the angle window when the outline is visibly axis-aligned — balcony bumps love a tilted false minimum. A building with NO drawn outline can still be anchored from its DIM_COOR corner callouts alone (pair X=/Y= texts, drop points that fall on the garage outline / fire-pad rects), and registering the UNION of several floors' footprints beats one floor when upper wings overhang. Surveying convention on both 13号/27号 site plans: drawing x = 测绘Y (east), drawing y = 测绘X (north), meters.
|
||||
- Buildings named on the site plan but with NO delivered drawings (27号 3#/4#): build honest massing specs from the site outline x storey count x storey height (labels like `22F` sit next to each outline; heights from the sibling tower's ladder). Mark them as 总图轮廓体量 in the spec description and campus notes — never silently fake a detailed model.
|
||||
- Adjacent-building gap sanity check is mandatory: compute the wall-to-wall gap in the shared-edge frame. ~0.15-0.30 m means 贴建 with an expansion joint (变形缝) — corroborate with the neighbour's roof labeled on this building's roof plan and matching floor elevations. Several metres means fire separation. A NEGATIVE gap (overlap) is always a placement bug. Cross-check the chosen position against the boundary: if a hypothesis (e.g. a 10 m fire gap) pushes the building through 用地红线, the hypothesis is wrong.
|
||||
- Site context is a deliverable, not decoration — a campus scene with only a few boundary lines reads as broken. Run `scripts/site_context_from_dump.py` on the 总图/定位图 dumps: it emits ready-to-merge floors for ground (用地红线 footprint slab), roads + parking stalls (lines, arcs, block-expanded 车位 INSERTs), fence (围墙 as 1.8 m railings), street trees (行道树 INSERT points as green posts, custom `trees` category), and low-confidence massings for context structures drawn only in the site plan (连廊/门卫/消控室, named by nearest label). It takes one plan-copy window (site sheets usually hold several scheme copies), a per-source unit scale (总图 in meters, 定位图 copies in mm), and an exclude-quad so the real building model replaces its own site outline. Buildings that have their own drawing sets get real models placed by the manifest — never fake them as massings.
|
||||
|
||||
|
||||
Reference in New Issue
Block a user