Some suggestions for the colormap scaling

[sadamov]

This PR replaces the implicit 3-level fallback chain in _get_heatmap_color_values with an explicit normalization parameter on plot_error_heatmap. I think this should be a clear choice. But would argue that state_std is the better default than diff_std. There is no need however to expose this flag to the user I don't think.

plot_error_heatmap(errors, datastore, title=None, normalization="state_std")  # or "diff_std"

normalization	Color value	Required stat
"state_std" (default)	Error / state_std	state_std
"diff_std"	Error / diff_std	state_diff_std_standardized and state_std

Both modes fall back to per-variable max normalization when their required stat is missing (legacy behavior). Here I think the absolute error values between variables are just too different and the heatmap is not really meaningful at that point.

I had to update the docstrings and the tests accordingly.

[kshirajahere]

Thanks @sadamov. I reviewed this locally and I like the direction overall. Making the normalization explicit is definitely clearer than the previous silent fallback chain, and from the COSMO examples I also agree that state_std is the better default for cross-variable comparison, while diff_std still remains available explicitly when that view is wanted.

I reran the focused heatmap tests locally on this branch and they pass on my side.
Just two things I wanted to confirm before I merge:

1. are you happy with diff_std remaining an explicit plot_error_heatmap(...) option for now, without plumbing it further through the current eval/logging path?
2. and are you happy with the missing-stats fallback now being per-variable max normalization rather than the previous raw absolute-scale behavior?

If yes, I m happy to merge this into mllam#376.

[sadamov]

2 * yes. :)

I think having a CLI flag just for this is too much. But having the explicit option in the code makes it easy to understand and change in research applications.

Absolute RMSE across variables (Pa vs K vs m/s) produces an unreadable colormap dominated by whichever variable has the largest physical units. Per-variable max at least preserves the relative evolution across lead times per variable.

[kshirajahere]

I agree with keeping the normalization choice as an explicit code-level option rather than a dedicated CLI flag, and I also agree that per-variable max is the better fallback than raw absolute RMSE across mixed-unit variables.
Also, I reran the focused heatmap tests locally on this branch and they pass on my side, so I m happy to merge this into mllam#376 ;)