feat: add RasterKind.SEMANTIC and make intake layout-aware

[leiverkus]

Closes the intake prerequisite for v0.8 (ROADMAP.md) and the intake-layout open decision (docs/architecture.md).

The bug

Effigies ships a per-pixel class field at odm_semantic/orthophoto_semantic.tif. Point Structura at such a delivery tree today and that raster is tagged ORTHO — the filename contains "ortho", and discover_inputs tested that substring first. It then goes to the 2D segmenter, which reads its class codes (0–3) as pixel intensities and emits spurious STONE polygons.

Nothing downstream re-checks the kind, so there is no error, no warning, and the output looks plausible. The added pipeline test reproduces it against the old code as assert 4 == 2, with the extra features carrying orthophoto_semantic.tif as their source raster.

The fix

Classification prefers the delivery layout. A raster under odm_orthophoto/, odm_dem/ or odm_semantic/ takes its kind from the directory, because the directory states what the product is — independent of what the file is called. This is what architecture.md said the heuristic should eventually do.

The filename fallback survives for flat or ad-hoc input directories, but "semantic" is now tested before "ortho". That ordering is load-bearing and the comment says so, because the failure it prevents is invisible.

RasterKind.SEMANTIC is added and the field is tagged but not consumed — consuming it as a prior is the rest of v0.8. The explicit branch in pipeline.run documents that rather than letting semantic products fall through the dispatch unmentioned. Tagging it is precisely what keeps it out of the segmenter in the meantime.

classify is split out of discover_inputs, so the rule is testable without a directory walk.

One behaviour change worth naming

Rasters matching neither the layout nor the filename rules are now skipped rather than guessed at. In practice this only affects files that were previously mis-tagged — the old code had no "unknown" outcome for anything containing ortho/dem/dsm/dtm, and skipped everything else already.

Tests

16 new tests (41 → 57). tests/test_intake.py covers the regression, layout-beats-filename, the filename fallback table, skipping unrelated rasters, and deterministic ordering. tests/test_pipeline.py gets the end-to-end guard described above.

The pipeline regression test was run against the pre-fix src/ to confirm it actually fails there — it does, assert 4 == 2.

Verification

Clean venvs on both matrix versions:

	3.11	3.12
ruff check .	pass	pass
mypy src	26 files clean	26 files clean
pytest	57 passed, 2 skipped	57 passed, 2 skipped

Plus a CLI run against a mock delivery tree — dem/dem/ortho/semantic tagged correctly, the sibling orthophoto_semantic.legend.json untouched.

🤖 Generated with Claude Code