Feature/global trends figs: global_trends trend-fields figure + teleconnections HadISST SST reference

[armellereca]

Summary

Two independent diagnostic improvements for the EERIE/CMIP6 evaluation:

1. global_trends: add an absolute trend-fields figure alongside the existing trend-difference figure.
2. teleconnections: use HadISST as the SST reference for the SST-based variability modes (ENSO, IOD, PDO), replacing ESA-CCI whose file only spans 1990–2014.

1. Global trends — absolute trend fields

The global_trends diagnostic previously produced only the observed trend plus model−obs trend difference panels per variable/period. It now also emits a second figure, {var}_{period}_trend_fields, showing the observed trend and each simulated trend field directly on a shared symmetric colorbar (no differencing), so absolute warming/cooling and precipitation-change patterns can be compared side by side.

• Reuses already-computed field data (model annual_regrid/seasonal_regrids, benchmark MMM and individual CMIP6 regrids) and the shared field colorbar range — no extra computation.
• Obs panel label is now derived from var_info.obs_dataset instead of hardcoded ERA5.
• run()'s skip-existing check covers both combined and fields figure ids.
• LLM figure-type prompt updated to describe both trend-map variants so Gemini interprets the new panels correctly.
• The website renders the new figures automatically (it collects PNGs generically and groups by the group metadata field) — no template changes needed.

2. Teleconnections — HadISST SST reference

The SST-based teleconnection indices (ENSO Niño 3.4, IOD, PDO) loaded their obs reference from tos → ESA-CCI, whose configured file only spans 1990–2014. Against models running from 1980 this left the observed ENSO index empty before ~1990, and the panel — hardcoded as "ERA5" — was actually ESA-CCI.

• Added an optional obs_dataset/obs_variable override to ModeDefinition and pointed the three SST modes at HadISST (HADISST/sst, full 1870–present record). Non-SST modes (SLP/wind) already resolve to ERA5 and are unchanged.
• Added a per-mode obs_label so the observation panel, spectrum legend, seasonal-variance legend, pattern panel, and metadata stats keys all read "HadISST" for SST modes and "ERA5" for the rest.
• Fill-value masking: HadISST flags land/sea-ice cells with sentinel values (-1000.0, -1e30). These are masked to NaN right after loading the override so they cannot corrupt Niño/IOD box means or the PDO North-Pacific EOF. Both EOF paths already drop non-finite columns and box means skip NaN, so the change is safe downstream.
• Deseasonalised anomalies are unit-invariant, so K-vs-K needs no conversion.

Tests

• global_trends: updated figure-count assertions (5 → 10 per variable), skip-existing pre-creation, and the summary-stats filter.
• teleconnections: added a load() method to the mock obs loader, registry assertions for the HadISST override + labels, and a fill-value masking test.
• All tests/test_global_trends.py (107) and tests/test_teleconnections.py (69) pass.

Regeneration

Full-resolution runs — execute in a compute environment, not the login node:

feather --config configs/eerie_all_members_cmip6_highresmip.yaml \
  --diagnostics global_trends teleconnections --benchmarks CMIP6 \
  --steps diagnostics -v

🤖 Generated with Claude Code