Worktree split geological model 2 - #298

…ualisation workspace package

Ports the loopstructural-visualisation repo in as a uv-workspace member,
following the src-layout pattern proven by packages/loop_common and
packages/loop_interpolation in Stage 2. Import name stays
loopstructuralvisualisation; only the on-disk container path changes.

- Routed the one print() call in __init__.py's trame-ui import guard
  through LoopStructural.utils.getLogger, consistent with the rest of
  the ecosystem's logging conventions.
- Left LoopStructural.datatypes imports in _3d_viewer.py as-is rather
  than swapping to LoopStructural.geometry: that module only exists in
  this repo's unreleased dev tree, not in any published LoopStructural
  release, and the package's declared floor is LoopStructural>=1.6.17.

Ports Loop3D/map2loop (upstream commit 078f8a6, master @ 3.3.1) into this
repo as packages/map2loop, following the packages/loop_interpolation
src-layout convention (Stage 2 of ROADMAP.md), for Stage 4's map2loop half.

- Source moved from flat map2loop/ to src/map2loop/; test suite ported to
  packages/map2loop/tests/ unchanged apart from one CWD-relative fixture
  path fix (tests/sampler/geo_test.csv -> path relative to the test file)
  needed because tests now run from the monorepo root, not map2loop's own
  repo root.
- pyproject.toml written from the loop_interpolation template (build
  system, classifiers, src-layout packages.find, tests extra) but keeps
  map2loop's own real name/version/requires-python/dependencies.
- Dependency-declaration gaps fixed (imported at module level but
  undeclared upstream): pandas and packaging. GDAL (osgeo) is also a hard,
  module-level import but is deliberately left undeclared, matching
  upstream's own convention of treating it as an externally-provisioned
  (conda/system) dependency rather than a pip one -- their own
  dependencies.txt explicitly skips it in check_all_dependencies(), and
  pip-building it from source fails without system libgdal headers.
- dependencies.txt (consulted at runtime by map2loop's own import-time
  version-compatibility self-check) copied to src/dependencies.txt to
  match the relative path the existing code resolves it from, and pytest
  removed from it since map2loop's own self-check otherwise hard-requires
  pytest to be installed for a plain (non-tests-extra) install to import
  successfully.

Verified in an isolated uv venv: `import map2loop` succeeds, and
`pytest packages/map2loop/tests` gives 78 passed, 4 failed, 1 skipped, 1
collection error -- all 5 non-passing outcomes trace back to the absence
of native GDAL bindings in this pip/Windows-only verification environment
(4 of them) or a monorepo-embedding artifact where one test's internal
`pytest.main()` call with no explicit path picks up this repo's own
unrelated tests/conftest.py pytest_plugins (1 of them), not to the port
itself.

Audited for loop_common type reuse per ROADMAP.md Stage 4: no swap made.
map2loop's bounding_box is a plain dict of map-CRS extents threaded
through dozens of call sites, architecturally unrelated to
loop_common.geometry.BoundingBox's local/global mesh-frame transform
model; its own map2loop.logging module is already map2loop's public,
already-used API surface, so routing its handful of ad-hoc print()
statements through loop_common.logging would need adding loop-common as a
new hard dependency for a cosmetic swap; no custom point/orientation data
structures exist to compare against loop_common.observations (dip/dip
direction data stays as GeoDataFrame columns throughout). No clean,
low-risk win found.

…visualisation

…visualisation

Completes Stage 4 (ROADMAP.md, outcomes 5/7) for map2loop and
loopstructural-visualisation, following on from the two ported packages
already merged in from agent worktrees (packages/map2loop,
packages/loopstructural_visualisation).

- Add both to [tool.uv.sources]; loopstructural_visualisation also needs
  LoopStructural = { workspace = true } since it depends on LoopStructural
  itself, which uv treats as an implicit workspace member once referenced
  by name.
- CI: loopstructural_visualisation joins the existing packages-test matrix;
  map2loop gets its own ubuntu-only job (map2loop-test) with an
  apt-installed libgdal, since GDAL has no pip wheel on Windows/macOS and
  upstream map2loop itself only tests via conda for the same reason.
- Add a minimal import-smoke test for loopstructural_visualisation, which
  shipped with zero test coverage upstream.
- Wire both into pypi.yml sdist/publish, release-please-config.json +
  .release-please-manifest.json (loopstructural_visualisation uses
  extra-files to bump version.py's __version__, since it has no static
  pyproject version field), and grandfather D/ANN ruff rules for both
  (matching the loop_common/loop_interpolation precedent).
- Correct ROADMAP.md: loopresources (Stage 4a) was never actually merged
  to master, despite a prior status-log entry claiming otherwise.

Verified: uv sync resolves cleanly for both packages standalone and as a
full workspace sync; tests/unit 695 passed/4 skipped (zero regressions).

…raphic_column from wiping state

evaluate_gradient's tetrahedron-refinement loop set resolved=True
unconditionally after one pass, so it never actually retried after
shrinking the tetrahedron near a fault. set_stratigraphic_column
cleared the model's stratigraphic column before unconditionally
raising DeprecationWarning, destroying existing state on every call.

GeologicalModel mixed recipe/pickle I/O in with feature-construction
orchestration, evaluation, and export logic (2580 lines total). Move
the recipe/JSON and dill/pickle serialization logic into a standalone
ModelSerializer, keeping every GeologicalModel public method as a
thin unchanged-signature wrapper so the CI-enforced stable API
surface (API.md, tests/unit/test_public_api_contract.py) is
unaffected.

Move surface/block-model export logic (get_fault_surfaces,
get_stratigraphic_surfaces, get_block_model, save) into a standalone
ModelExporter, continuing the GeologicalModel decomposition. Public
methods remain unchanged thin wrappers.

Move fault/unconformity relationship bookkeeping (_add_faults,
_add_domain_fault_above/below, _add_unconformity_above,
add_unconformity, add_onlap_unconformity) into a standalone
FeatureRelationshipManager, continuing the GeologicalModel
decomposition. Public methods remain unchanged thin wrappers;
_add_feature stays on GeologicalModel since it owns
features/feature_name_index mutation directly.

Move the seven _build_* feature-construction methods (foliation,
fold frame, folded foliation, folded fold frame, intrusion, domain
fault, fault) -- collectively ~1000 lines, the largest remaining
chunk of GeologicalModel -- into a standalone ModelFeatureFactory.
The FeatureBuilderRegistry wiring at the bottom of geological_model.py
now dispatches to ModelFeatureFactory.build_* instead of the removed
private methods. Public create_and_add_* wrappers are unchanged;
their docstring cross-references are updated to point at the new
factory location. geological_model.py drops from 2579 to 1505 lines.