[dnerini]

#270 added cloud cover (tcc, lcc, mcc, hcc) and skt to the shared -oper forecaster and downscaler configs, including a cloud cover rescale on the ICON-CH1 cutout. Varda-single 1.0 does not use these variables, and the rescale fails (AssertionError: tcc) with the 2024-2026 v2.0 dataset, which names them CLCT, CLCL, CLCM, CLCH and SKT.

• Move the cloud cover setup of Support cloud cover (CLCT, CLCL, CLCM, CLCH) and skin temperature (SKT) for the Varda-single release candidate #270 unchanged to new -clouds configs and metadata patches (still on the v1.0 dataset).
• Restore the -oper configs and patches to their pre-Support cloud cover (CLCT, CLCL, CLCM, CLCH) and skin temperature (SKT) for the Varda-single release candidate #270 state, keeping the SKT fixes (downscaler input namer, IFS GRIB encoding of skt).
• Point the -oper patches at mch-ich1-1km-2024-2026-1h-pl13-ifsnames-v2.0 restricted to 2024-05-01..2026-05-01, so 2026 reference times can be evaluated.

[Louis-Frey]

Thanks for spotting this Daniele, I'll check it.

[cosunae]

I will insist on my comment here:
#282 (comment)

to avoid these kind of issues, inconsistencies of configs with checkpoints and a high maintenance in evalml, we could try to make the effort to move them to the workspaces.

[Louis-Frey]

Thanks Daniele, and sorry for introducing this in #270. I tested the PR and it looks good.

Test: one config with two init times, 2025-06-01 (in both KENDA-CH1 v1.0 and v2.0) and 2026-02-01 (v2.0 only), short lead times. I ran it on this PR and, as a reference, on my branch (which reads v2.0 and renames CLCT etc. back to tcc etc.), then compared the GRIB outputs byte by byte:

Run	Configs	2025-06-01	2026-02-01
Varda-single 1.0 (forecaster + downscaler)	-oper	runs, identical to reference	runs, identical to reference
9f4626a7 (forecasters-ich1-oper-fixed.yaml)	-oper	runs, identical to reference	runs, identical to reference
RC (forecaster 07f64210 + downscaler d642c4ba)	-clouds	runs, identical to reference	fails: date not in v1.0

So the split doesn't change any output, and Varda-single 1.0 now works on v2.0 including 2026 dates.

One follow-up (doesn't block this PR): since the -clouds patches still read v1.0, the RC can't be evaluated for 2026 dates. Pointing them at v2.0 with a rename would fix that, as on my branch:

- dataset: /store_new/mch/msopr/ml/datasets/mch-ich1-1km-2024-2026-1h-pl13-ifsnames-v2.0.zarr
  rename: {CLCT: tcc, CLCL: lcc, CLCM: mcc, CLCH: hcc, SKT: skt, SSO_THETA: anor, SSO_GAMMA: isor}
  rescale: {tcc: [0.01, 0.0], lcc: [0.01, 0.0], mcc: [0.01, 0.0], hcc: [0.01, 0.0]}

Maybe worth doing at some point.

@cosunae I agree that moving the inference configs into the workspaces would avoid this kind of issue. I don't have time for it right now, but I'm happy to help with it later.