Add NASA IMERG Early and Late V07 analysis datasets

[aldenks]

Summary

Adds two global half-hourly precipitation analysis datasets from NASA GPM IMERG V07:

• nasa-imerg-analysis-early-v7
• nasa-imerg-analysis-late-v7

(IMERG Final is intentionally deferred to V08 — V07 Final stopped 2025-09-30.)

The two runs differ only in their source product/URL and update latency, so they share NasaImergRegionJob and a NasaImergAnalysisTemplateConfig base under src/reformatters/nasa/imerg/, mirroring the NOAA GFS forecast/analysis shared-code layout. A single NasaImergSourceFileCoord carries a run field; thin per-run RegionJob subclasses set it via a ClassVar.

Details

• Source: GES DISC HTTPS with NASA Earthdata Login (reuses the SMAP earthdata_auth pattern). download_file tries the date-appropriate V07C/V07B filename label and falls back to the other.
• Reading: rasterio over the HDF5 /Grid subdatasets, reorienting the source (time, lon, lat) layout to (latitude, longitude), converting fill values (-9999.9 float / -9999 int16) to NaN. precipitation is converted mm/hr → kg m-2 s-1 to match the existing precipitation_surface / precipitation_flux convention.
• Variables: precipitation_surface, probability_of_liquid_precipitation_surface, precipitation_quality_index_surface.
• Grid / time: global 0.1° (1800 × 3600), record from 1998-01-01 (verified against the GES DISC archive listing).
• Chunking: chunk (1440, 45, 45) / shard (1440, 450, 450) = 30 days/shard (~37 GB per-worker buffer). The memory ↔ read-locality ↔ update-cost tradeoff and lean (720) / premium (2880) alternatives are documented in template_config.py.
• Storage: Icechunk on AWS Open Data (s3://dynamical-nasa-imerg).
• Added "0.1 degrees (~10km)" to the shared SpatialResolution literal; added the two non-CF/non-ECMWF variables to the compliance-test exemption lists.

Testing

• Unit tests for all custom logic (URL building for both runs, V07B/V07C selection, the HDF5 reorientation, fill/scale, coord generation, version-fallback download, operational update jobs, K8s/validators).
• A fully-mocked backfill_local → update integration test that runs offline and exercises the whole download → read → write → append pipeline.
• Slow tests that download and read a real granule from each run (Early & Late) — require Earthdata credentials, so they don't run in plain CI.
• ruff, ty, and the full non-slow suite pass.

Follow-ups before backfill

• Verify lat/lon orientation against a real file: the (time,lon,lat) → (lat,lon) transpose + latitude flip is implemented from the source spec but only confirmed by the slow tests. Run uv run pytest tests/nasa -m slow with Earthdata credentials before backfilling.
• Add the datasets to the dynamical.org catalog (separate repo).

Relates to #462

🤖 Generated with Claude Code