v0.7.0 — semantic field v0 + change-detection accuracy

The semantic-field bridge to Structura plus a hardening pass on change-detection accuracy. Mechanism only — Effigies ships the field machinery; the fine-class archaeological material model is a downstream Structura deliverable and is never baked into the MIT image.

Added

• Semantic orthophoto v0 (helpers/semantic_ortho.py, opt-in --semantic). The
  geometry-derived first increment of the semantic field (ROADMAP v0.7.0): it
  rasterises the OpenPointClass point classes already written into the LAZ onto the
  orthophoto grid — pixel-aligned with odm_dem/dsm.tif — taking the per-cell
  majority class, folded into ground / vegetation / structure, as
  odm_semantic/orthophoto_semantic.tif (Byte GeoTIFF + colour table) plus a legend
  JSON. No trained model: it ships what the cloud classification already knows; the
  fine archaeological material classes (stone/earth/paving/ceramic/mortar) are a
  downstream 2D-model deliverable. Needs a classified cloud (enable --classify);
  self-skips, non-fatally, otherwise. Runs after the orthophoto/DSM so the grid exists.
  Unit-tested (majority + ASPRS→v0 mapping + GeoTIFF round-trip) and image-validated
  end-to-end. This is the bridge's v0 to Structura's vectorisation.
• Multi-epoch semantic propagation (helpers/semantic_propagate.py). Carries the
  semantic field across daily epochs (runs under --semantic when --align-to gives a
  reference epoch). Because change detection re-lands this epoch into the reference frame,
  its semantic ortho is already co-registered with the reference epoch's; the step writes
  (1) a carry-forward field — odm_semantic/orthophoto_semantic_propagated.tif, where
  this epoch's unobserved cells inherit the reference epoch's class (temporally consistent;
  honest "unobserved = unchanged" assumption) — and (2) a semantic-change raster —
  odm_semantic/semantic_change.tif with per-pixel class transitions + per-transition area
  in odm_report/semantic_change.json (the class complement of the DoD/M3C2 geometric
  change: e.g. structure→ground = a feature removed, vegetation→ground = clearing). The
  reference is resampled nearest-neighbour (categorical). Self-skips without both semantic
  orthos; non-fatal. Unit-tested (carry-forward + transition + stats) and end-to-end
  validated.

Changed

• Change detection — M3C2 level-of-detection now includes the co-registration
  residual (change_detect.py). The per-point LoD previously reflected local
  roughness only; it now passes the post-ICP cloud-to-cloud residual to py4dgeo as
  registration_error (Lague 2013), so a cm-level alignment error is folded into the
  significance test instead of being treated as zero — small (few-cm) changes are no
  longer over-reported as significant. Recorded as registration_error_m in
  odm_report/change_detection.json. Unit-tested: a 5 cm residual lifts the LoD
  median (≈ 0.3 cm → 10 cm). Stable-area-masked ICP remains v2 (see ROADMAP).
• Change detection — DoD now thresholded at a minimum level-of-detection
  (change_detect.py). The DEM-of-Difference changed area and fill/cut volumes now
  count only cells whose |Δz| exceeds a minimum LoD (Wheaton 2010 — a robust noise
  floor of the difference distribution, floored by the co-registration residual;
  min_lod_from_dod), so sub-LoD noise is no longer booked as excavation / back-fill.
  A raw un-thresholded net volume is kept as a cross-check, and the minLoD is recorded
  as min_lod_m in odm_report/change_detection.json. Unit-tested.
• Change detection — co-registration is now stable-area-masked (change_detect.py).
  ICP runs in two passes: a whole-cloud fit, then a re-fit on only the unchanged
  ground (changed cells dropped via stable_mask), so a localised excavation change no
  longer biases the rigid transform, and the residual over the stable area is a clean
  registration-only error that now feeds the M3C2 LoD and the DoD minLoD
  (coreg_reg_error) instead of the conservative full-cloud C2C. Reports
  stable_fraction + registration_error in change_detection.json; degrades to the
  whole-cloud fit (and records why) when too little stable ground remains. Unit-tested
  (stable_mask separation; the two-pass ICP smoke is pdal-gated).
• Change detection — --align-to now re-lands the deliverables into the reference
  frame by default (change_detect.py), ODM --align parity. The recovered ICP
  transform is applied in place to the delivered mesh OBJ (offset-aware, transform_obj)
  and the LAZ (PDAL; EPT rebuilt); because re-landing runs before the raster stages, the
  DSM/DTM, orthophoto, contours, glTF and 3D Tiles inherit the reference frame natively
  (no re-warp). Previously the alignment was additive-only (deliverables untouched) —
  pass --no-reland to keep that. report["relanded"] lists what moved; non-fatal per
  asset. Camera assets (shots.geojson) are a known re-land gap (v2). Offset-exact
  transform unit-tested; the full raster re-derivation is Docker-validated.
• Change detection — --align-to now accepts a DEM GeoTIFF as the reference, not
  only a point cloud (change_detect.py). A reference .tif (a prior DSM/DEM) is
  read as cell-centre points for the ICP co-registration and M3C2, and used directly
  (resampled onto the shared grid) as the reference DSM for the DEM-of-Difference —
  so a prior epoch's DSM, or any external reference DEM, can drive the comparison.
  is_dem / dem_to_xyz / resample_dem; unit-tested.
• Change detection — camera assets are re-landed with the rest (camera_exports.py).
  When an --align-to run re-landed this epoch into the reference frame, shots.geojson
  camera centres and orientations are now transformed by the recorded re-land transform
  (read from odm_report/change_detection.json, gated on the relanded marker), so the
  camera positions stay consistent with the re-landed mesh / cloud / orthophoto instead
  of sitting in the old frame. cameras.json is unchanged (intrinsics are
  frame-independent). The re-land gate is unit-tested; the pyproj WGS84 path is
  Docker-validated. Closes the last v2 gap of the change-detection item.
• Change detection — fix: M3C2 no longer misses deep changes (change_detect.py).
  py4dgeo's M3C2 max_distance (the cylinder search depth along the normal) defaulted
  to 0, which on the auto-scaled small cylinder of a dense cloud is too shallow — a deep
  excavation (a change larger than the cylinder scale) had no matching surface in range
  and came back NaN / not-significant, even where the DoD clearly measured it. It is now
  set generously (max(30·cyl_radius, 3 m)). Found by the new end-to-end smoke
  (scripts/smoke_change_detect.py), which the fix takes from M3C2 0 %→significant on a
  0.4 m block.

Carried forward (post-v0.7.0)

• The full mesh-classify → z-buffer --semantic path (v0 approximates it via per-cell cloud majority).
• The fine-class material model (stone / earth / ceramic / mortar) — Structura's deliverable, itself data-blocked.
• Cross-project contract finalisation.

---

Test gate: pure-Python unit suite green; the only scripts/test.sh failure is the pre-existing macOS bash-3.2 issue in test_autoscale.sh (passes on the bash-5 Docker image).