Standalone CRS benchmark validation

The standalone crate was run directly through the reporter's unmodified judges and frozen inputs at rust-proj-capability revision ddd9fac70ec370acd9651ebb8b46035207347ad5. The Terrio source revision was f454562 on perf/crs-optional-fallback. Machine-readable results include input hashes, oracle environments, per-fixture results and projection error maxima.

CRS study

The 93-fixture study retains its frozen PROJ 9.9.0 / pyproj 3.8.0 oracle, with no datum grids and network disabled. The judge's "exact" classification requires less than 1 mm; it does not assert bitwise equality.

The fallback configuration retains the reported Terrio 0.5.0 score of 74 exact fixtures. The native configuration refuses method-eck4, method-natearth and method-hammer, with an error naming wbprojection-fallback. Both configurations retain the existing approximately 0.815 m forward difference for form-2264-proj-bound and refusal at the Mercator pole (edge-merc-pole). Geocentric, PROJJSON, IAU and additional projection capabilities are tracked in #269– #272.

Projection graticules

The 54-projection suite uses its separate PROJ 9.10.0 development oracle, revision 5fa31c29f0a806568ed7bc76c1a557604db4a804, with pyproj 3.8.0. Its "exact" classification requires forward error below 1 mm and inverse error below 1e-8 degrees, with longitude differences wrapped and scaled by cos(reference latitude).

Both configurations classify the four corrected projections as exact across each 91-point graticule, including poles and the reported orthographic case:

These are native build observations for the pinned fixtures. WASM pole rounding has separate allowances in the repository's regression tests; see testing.

Winkel Tripel remains "off" in the original projection benchmark because its bare operation oracle supplies PROJ's operation default lat_1=acos(2/pi), while Terrio's CRS parser supplies the CRS default lat_1=0. With an explicit lat_1=50.467…, the pinned regression agrees within its documented tolerances. The original benchmark inputs were retained, so the table exposes this semantic difference. No oracle was regenerated with the local PROJ installation.

Reproduction

In the pinned benchmark checkout, replace its two terrio Cargo dependencies with aliases of the standalone package at the Terrio revision above:

terrio = { package = "terrio-crs", git = "https://github.com/opengeos/terrio", rev = "f454562", default-features = false }

Retain optional = true in engines/rust-crs/Cargo.toml. Replace the adapter imports terrio::crs:: with terrio:: (also in the CRS compile check), and label the candidate/engine terrio-crs, version 0.1.0. The measured run used a local path alias to this same source. The adapters, fixtures and judges otherwise stay unchanged.

Use an empty directory for TERRIO_GRID_DIR, TERRIO_GRID_CACHE, PROJ_DATA and PROJ_LIB, and an isolated XDG_DATA_HOME; set TERRIO_NETWORK=OFF and PROJ_NETWORK=OFF. Retain the workspace's bounded local-build settings.

From engines/rust-crs, run:

cargo run --release --no-default-features --features c-terrio -- \
  --candidate terrio-crs ../../suites/crs/crs.json
cargo run --release --no-default-features \
  --features c-terrio,terrio/wbprojection-fallback -- \
  --candidate terrio-crs ../../suites/crs/crs.json

From engines/rust-projections, run:

cargo run --release -- --engine terrio-crs ../../suites/projections/reference.json
cargo run --release --features terrio/wbprojection-fallback -- \
  --engine terrio-crs ../../suites/projections/reference.json

Save each JSON-lines result separately and pass projection records to the pinned report/projections.py::judge against the corresponding reference entry. The CRS executable includes its judge and emits verdicts directly. Keep generated lockfiles to reproduce the resolved third-party dependencies.