Rename filter not consistent with rename-action in `open_dataset`

[mpvginde]

Describe the bug
Let's say I want to build a dataset with 6h- and 3h accumulated precipitation

# 3h accumulations
       - dates:
            start: 2020-01-01 00:00:00
            end: 2021-01-01 00:00:00
            frequency: 3h
          accumulations:
            <<: *mars_request
            time: [0]
            accumulation_period: [0, 3]
            param:
            - tp    # total precipitation
# rename to allow 2 total precipitation fields    
      - rename:
          param: "{param}_3h"
# 6h accumulations
       - dates:
            start: 2020-01-01 00:00:00
            end: 2021-01-01 00:00:00
            frequency: 3h
          accumulations:
            <<: *mars_request
            time: [0]
            accumulation_period: [0, 6]
            param:
            - tp    # total precipitation
# rename to allow 2 total precipitation fields
      - rename:
          param: "{param}_6h"

The rename filter in the recipe updates the param key in the metadata, which I guess is then used as the variable name in the dataset.

As far as I know there is currently no way of renaming a single variable (during dataset creation) without also changing the param metadata.

When changing the variable name structure of all variables with remapping, this is possible

output:
  remapping:
    param_level: "{param}_{levelist}"

I use the above snippet to get rid of the _2 or _10 in the name of surface levels fields like 2t_2 or 10v_10.
Here only the variable name is changed.

Now I want to combine this dataset with another dataset using the cutout functionality

Now the combined dataset has variable name tp with param metadata tp_6h.
This (currently) trips up the scalers during training.

** Version number **
building of cerra
anemoi-datasets branch abstracting-accumulation (8fb0a16)

opening datasets with cutout:
anemoi-datasets: current main (c26a9d8)
anemoi-transform: current main (ecmwf/anemoi-transform@548e2fa)

Additional context
I see two possible solutions:

1. Add functionality to specify the name of a variable during building without changing the param metadata
2. Let the rename action in open_dataset also rename the param metadata, but this might lead to other problems with pressure level fields as these typically have name: t_600, param: t

[passionworkeer]

Great issue! A few thoughts on potential solutions:

1. For option 1 (separate variable naming): Consider adding a �ariable_name or output_name key in the rename filter that only affects the dataset variable name without touching param metadata. This mirrors how
   emapping works for output.

2. For option 2 (consistent rename): You could add a preserve_metadata flag (default True) to control behavior. This way users can opt-in to metadata renaming when they need it.

3. Quick workaround: In the meantime, you could post-process the combined dataset with xarray to rename variables before passing to transformers:
   python ds = ds.rename({'tp': 'tp_6h'})

4. Another approach: Add a strict mode that validates variable name consistency between datasets during cutout operations.

Would love to see this fixed as it affects the scalers!

[mc4117]

I think this is related to this issue #332?

[mpvginde]

Yes, should we close this issue, to prevent duplication?

[mc4117]

Yes, should we close this issue, to prevent duplication?

I think it would be good to add your example above to the original issue (just to keep track of where the issue is occurring) and then close this one as you suggest. This has been a known-issue for a while so could be good to get a group working on this