GODEEEP CF registry: config-declared retrieval with no fallback paths

[ktehranchi]

Closes #803.

What

Replaces the hardcoded Zenodo record table and the download hidden inside build_renewable_profiles with a single declarative registry for GODEEEP compressed capacity-factor files:

• godeeep_cf_registry.py — sources (local mirror / Zenodo) are declared in config["godeeep_cf_registry"] with exactly the (dataset, year) pairs they hold. Sources are walked in config order; the first one declaring the request wins. There are no fallback paths: no default hub height, no nearest-year substitution, no unscreened-variant swap. All configured scenario × year × technology combinations are resolved eagerly at snakemake parse time; failures aggregate into one error naming the dataset key, the requested year, and the available years per source.
• rule retrieve_godeeep_cf — retrieval is now an explicit DAG step. Local-mirror hits are symlinked (or copied via copy_local) after sha256 verification against the mirror's SHA256SUMS; Zenodo hits download with checksum validation and retries. build_renewable_profiles consumes snakemake.input.godeeep_cf and never picks a file itself.
• zenodo_downloader.py — every failure raises ZenodoDownloadError; nothing returns None or leaves partial files. The unregistered wind_125m_rcp45hotter record (20127545, HTTP 404) is declared with an empty year list so requesting it fails honestly instead of via the legacy silent-None path.
• Compression pipeline (nrel_exclusion/compress_godeeep_*) used to build the 172-file sha256-verified mirror (solar + wind 80/100/125 m × 1980–2022 historical), plus 68 unit tests covering the registry, downloader, and compressor.
• Docs — retrieval mechanism and a pre-aggregated data availability matrix (which datasets/scenarios/years exist on the local mirror vs. Zenodo).

Verification

• 81 fast tests pass (68 new).
• Parse-time probes: invalid hub height, out-of-range year, and the unpublished rcp45hotter wind record each fail loudly with actionable messages listing per-source availability.
• End-to-end on Sherlock: historical 2019 / _125m / western / s75 build_renewable_profiles run retrieves via symlink + sha256 from the mirror and produces sane profiles (8760 hourly 2019 snapshots, CF ∈ [0,1], zero NaNs, full bus coverage; onwind's 73/75 buses reflect zero NREL wind capacity in LA + Inyo counties, pre-existing behavior).

Notes for reviewers

• resolve_cf validates from config declarations alone — it does not stat the local root. An off-Sherlock user keeping the default config passes parse-time validation and fails at retrieve time with a missing-file error; docs tell them to remove/repoint the local source. Guidance, not enforcement, in this PR.
• The mem_mb=32000 request for godeeep build_renewable_profiles measured ~17× actual peak (1.9 GB) at s75/western — left as-is since it may be sized for finer resolutions; happy to right-size.
• 80 m wind data exists in the mirror archive but is deliberately not wired into VALID_WIND_HEIGHTS — the _80m silent default was the original GODEEEP: historical weather years other than 2012 cannot receive NREL land-access screening (unscreened fallback profiles) #803 bug; enabling it explicitly can be a follow-up.

🤖 Generated with Claude Code