MRB-860 multi panel lead time plots

[Louis-Frey]

Summary

Adds a configurable multi-panel metric-vs-lead-time plot, driven by a new multipanel_plots: section in the experiment YAML, plus quality-of-life improvements that also benefit the existing single-panel plots and the dashboard.

What you get

• New multipanel_plots: config section. Each named entry declares rows, cols, optional figsize/title, and a row-major list of panels. Each panel is {metric, param} plus optional region, season, init_hour, title, ylim. Validated by MultipanelPlotSpec (Pydantic): rows*cols must equal len(panels), unknown fields are forbidden.
• New workflow rule verification_metrics_multipanel_plot producing output/results/<experiment>/multipanel/<name>.png. Aggregator rule verification_metrics_multipanel_plot_all expands over every named layout. Wired into experiment_all, so plain evalml experiment builds the layouts when the section is present and is a no-op otherwise.
• Units on y-axis labels. Auto-derived from (metric, param) for matplotlib plots (single-panel and multipanel). Backed by a small PARAM_UNITS map in src/plotting/units.py. Known limitation: the verification netCDFs don't carry units attrs, so this is a hard-coded dict — see follow-ups below.
• Panel labels ( a), b), c), ...) rendered as bold, left-aligned titles at the same height as the centred title.
• Stable, bijective source → color mapping shared with the dashboard. Both surfaces use Vega-Lite's tableau10 over the alphabetically-sorted full source list. Pinning Vega's color.scale.domain to that list also fixes a longstanding dashboard glitch where toggling sources reshuffled the remaining colors.
• Refactor: shared verification-plotting building blocks. src/verification/loading.py (load_long_df, subset_df, _ensure_unique_lead_time, _select_best_sources) and src/plotting/metric_lead_time_panel.py (plot_panel). The existing single-panel script and the dashboard script were updated to consume these; no behaviour change.
• Tests. 5 new tests in tests/unit/test_config.py covering panel-level defaults, rows*cols validation, extra: "forbid", and a ConfigModel round-trip with multipanel_plots populated.

Example

config/multipanel_example.yaml is a working config with two layouts: bias_overview (BIAS by season — all vs JJA — for T_2M and PMSL) and rmse_overview (RMSE by init hour — 0, 12 — for T_2M and PMSL). Run with evalml experiment config/multipanel_example.yaml, which produces single-panel plots, dashboard, and multipanel PNGs all in one go.

Test plan

• pytest tests/unit — 44 passing (5 new).
• Pre-commit pydantic-schema hook clean (schema regenerated).
• Dry-run with --rerun-triggers mtime against a fresh multipanel_plots config: only multipanel/plot/dashboard jobs planned, no spurious inference reruns.
• Visual check of generated PNGs and dashboard.html; colors agree across the three surfaces; dashboard source toggle preserves the remaining sources' colors.

Caveats / follow-ups

• PARAM_UNITS in src/plotting/units.py is a hand-maintained dict because the metric computation in src/verification/__init__.py:verify drops the units attrs that the input data carries. A cleaner fix is to propagate units through the verify → aggregate chain so the netCDF carries them; left as a follow-up.
• Color palette caps at 10 sources (Vega's tableau10); beyond that we wrap and lose bijectivity. Easy to bump to tableau20 later, or to a deterministic HSV ramp, if we ever have >10 sources.
• Suptitle and legend placement on the multipanel figure are minimally tuned (left-aligned panel labels can crowd the suptitle on wide layouts). Worth revisiting if we settle on a publication-ready layout.
• I tested locally by symlinking pre-MRB-820 aggregates into the new run-id paths to avoid re-running inference for a large pre-existing experiment; that's a local hack — output/ is gitignored, so a fresh clone needs a full evalml experiment run. Fine to do a small experiment for development, can still see about a full experiment on truly independent test data (2026) once we polish for the paper.