qig-geocoding

Fisher-Rao attention, Fourier position features and recursive integration for geometric neural networks. Install with pip install qig-geocoding; the Python import is geocoding.

Compatibility and metric units

The attention metric retains radius 2, 2 * acos(BC), matching qigkernels 0.4.4. Current qig-core uses radius 1, acos(BC). Convert explicitly when comparing distances. Halving attention distance without also halving its temperature changes the model, so this release preserves existing attention outputs. The finite-gradient guards also differ near coincident and sparse points; they are not exactly the same function throughout the domain. Both full and compute-skipping banded attention remain faithfulness-gated.

Scientific provenance: qig_core.torch.geometry_simplex.fisher_rao_distance_simplex in qig-core 2.15.3 records the PI ruling reconciling its import paths to radius 1. The qig-verification EXP-009 causal-sweep results document the historical radius-2 convention. These are compatible unit conventions, not grounds to rewrite frozen experiment results. Dense-point conversion, sparse-point error bounds and the live Duchi simplex projection are tested separately. Current qig-core already handles coincidence gradients safely.

Release changes

• Preserve development's D1 local-critical curvature and opt-in basis telemetry, D2 query-only inter-layer geodesic sync, D4 basin_layer_drift naming, and the D3 countersigned preregistration amendment.
• Correct the obsolete equal-radius compatibility assertion; preserve attention arithmetic.
• Reject non-callable curvature providers at construction rather than silently swallowing invalid wiring.
• Align the runtime version with package metadata; require the current validated qig-core dependency.
• Install qigkernels in development/release checks so compatibility cannot silently skip. Lint, types and all CPU tests gate release. Resolve published dependencies in standalone checkouts without a parent workspace.

D3 is a preregistration, not a completed empirical result. This release does not claim that its order-preservation kill experiment has passed. Basis reductions remain opt-in; sync enables them explicitly.

Verify and publish

Install CPU PyTorch, then pip install -e '.[dev]'. Run ruff check src tests, mypy src, and pytest -q. The trusted-publishing workflow uses the existing pypi environment and OIDC. Publication currently requires a version tag after promotion; merging alone does not trigger release.yml. Manual dispatch runs checks/build only. Release operators must verify the wheel version and PyPI availability before declaring publication complete.