[kylebarron]

Note

This PR was written by Claude (Claude Code), not by @kylebarron.

Stage 4 of the GeoKey-to-CRS work: the user-defined CRS path of _crs.py is now driven by the GeoTIFF CT method table shared through geotiff-test-data, built with pyproj's typed classes, and tested against GDAL's reading of every fixture in the conformance corpus. Design: dev-docs/specs/2026-10-26-geokey-crs-conformance-design.md.

What changed

• _crs_table.py loads the vendored data/geotiff-ct-methods.json into typed records (method, variants, parameters with their GeoKey fallback lists).
• _crs_geographic.py builds the geographic side the way libgeotiff's GTIFGetDefn and GDAL's GTIFGetOGISDefn do: angular and linear units from EPSG codes (via proj.db) or explicit sizes, the ellipsoid from its code or axes (sphere when the axes are equal), the datum and prime meridian from their codes or from GDAL's citation names (Datum = ..., Primem = ..., unknown fallback, never undefined), and coordinate systems in the file's units and axis directions.
• _crs_conversion.py interprets the table: first-present GeoKey wins (a stored 0.0 counts), angular and linear values scaled to degrees and metres, the Mercator and Polar Stereographic variant rules, the polar axis rule, pyproj's typed conversion classes where they exist and a PROJ string otherwise, then a check that every table parameter reached the conversion with its value. A table key async-tiff cannot read (ProjRectifiedGridAngle today) raises instead of defaulting.
• _crs.py is now dispatch only; crs_from_geo_keys and projjson_from_geo_keys keep their signatures.
• All builders take the GeoKeys as a plain mapping, so the 70-odd unit tests use dicts.

Conformance

KNOWN_FAILURES in tests/test_crs_conformance.py drops from 34 rows to 7:

• tmerc_towgs84, omerc_variant_a, omerc_variant_b, somerc: need GeoKeys 2062/3096, merged in feat: Expose GeogTOWGS84GeoKey and ProjRectifiedGridAngleGeoKey async-tiff#368 but not yet released.
• lcc_1sp_paris_grad: GDAL's own write and read of a prime meridian in grads disagree (spec Q3); we follow libgeotiff's single conversion.
• krovak, moll: ESRI PE String rows, where GDAL layers the Geog* keys and ESRI name morphing on top of the WKT and pyproj parses the WKT alone. The PE String path is unchanged in this PR.

The fresh-context review found three inputs the GDAL-written corpus never produces and that misbehaved: a coded datum without ellipsoid keys raised, an ellipsoid code with only a semi-major axis silently became a sphere, and the degree code 9102 scaled angles by a factor a hair off 1.0 (and some sexagesimal codes pyproj reports with a zero factor would have zeroed every angle). All three are fixed with tests.

uv run pytest: 511 passed, 4 skipped, 9 xfailed (the 7 above plus 2 non-strict proj.db-drift rows). Ruff and mypy clean.

Follow-ups

• geotiff-test-data: the table's Laborde (CT 4) parameters inherit the Hotine keys lonc/alpha, but PROJ's labrd takes lon_0/azi; the builder overrides them (PROJ_KEY_OVERRIDES) and the validation step is what caught it. The table should be corrected so the override can go.
• When async-tiff ships keys 2062 and 3096, add their accessors to KEY_ACCESSORS (and a BoundCRS for TOWGS84) and remove the four rows.
• Also includes the .gitignore entry for dev-docs/plans/ that did not land with test: Add the CRS conformance suite from geotiff-test-data #181.

🤖 Written by Claude Code

🤖 Generated with Claude Code