Stop derived fields from inheriting their parent's metadata

[xylar]

Problem

In every p-star product, minLevelCell, maxLevelCell and cellMask were labelled long_name = seafloor pressure, units = Pa and, simultaneously, unit = meters, while BottomPressure — the one field that really is a seafloor pressure — had lost every attribute it was given, along with RefPseudoThickness and vertCoordMovementWeights. Values are correct throughout; only metadata is wrong. Nothing in Polaris, MPAS-Ocean or Omega reads these attributes, so no results change — what changes is what a file staged to E3SM's inputdata server tells the next reader about itself.

The cause is that xarray propagates attrs through arithmetic, comparisons, where(), zeros_like() and NumPy ufuncs, so a level index derived from a pressure arrives claiming to be that pressure. Setting individual keys (da.attrs['units'] = ...) leaves everything else the parent carried in place. Note that xr.set_options(keep_attrs=False) is not a fix: it suppresses propagation through arithmetic and comparisons, but zeros_like() and where() keep attributes regardless, and both are on the failing path.

The damage was partial_cell_type-dependent and worst where nobody was looking: with partial_cell_type = none (which horiz_press_grad uses) every p-star field was mislabelled, not just the three index and mask fields.

What changed

• New polaris/attrs.py with set_attrs(), which replaces a variable's attributes wholesale rather than merging, plus a developer-guide page on why that is necessary.
• pstar.py now labels every field it creates, and preserves BottomPressure's caller-supplied attributes across the post-snap overwrite so horiz_press_grad keeps saying "seafloor gauge pressure".
• pstar_init.py, the init-state helpers, ztilde.py, eos and the overflow p-star step all set attributes wholesale; a dead vertCoordMovementWeights branch in pstar_init.py is removed.
• The whole topography chain is labelled: the combined fields (where base_elevation came out of the latitude-weighted blend calling itself Latitude of Grid Cell Centers, and ocean_mask claimed standard_name = latitude), and the masked and remapped fields downstream (where land_frac and ocean_frac — fractions in [0, 1] — claimed to be a bedrock elevation in metres). Dangling cell_measures, coordinates and grid_mapping attributes, all naming variables the remap step drops, are removed. ice_thickness keeps its CF standard_name; the other fields get none, since no verified CF name applies.
• Tracer conversion no longer carries the source convention's metadata onto the converted tracer, where for example a standard_name of sea_water_conservative_temperature would survive conversion to potential temperature.
• Two fixes found while verifying the suite output: RefPseudoThickness is no longer rebuilt into init.nc after write_initial_state_dataset drops it (that duplicate conversion also used the wrong surface pressure), and Density is consistently in-situ density, distinguishing it from a potential density.
• One unrelated fix found along the way and bundled to save a second review: polaris cache computed its output json filename outside the per-component loop, so caching two components at once wrote both into one file named after whichever component came last, and read the wrong one back in.

Testing

pytest tests/ passes at 322 tests, up 30. The new tests seed their inputs with attributes that must not survive — a BottomPressure labelled as a pressure in Pascals, a geom_z_bot carrying the topography's unit and cell_measures — and assert nothing downstream keeps them; they fail on the pre-fix code and would fail again if a future xarray changed propagation in either direction. test_pstar.py is parametrized over all three partial_cell_type settings, which previously disagreed with one another.

Cached topography

The combine step is default_cached, so its labelling reaches no one until the combined topography is regenerated; the remap and cull steps are not cached and need nothing. All seven resolutions were rerun and re-uploaded, and cached_files.json points at them. Data is unchanged to remapping round-off (at most 6e-16 relative). Regenerating ne3000 also restored ocean_mask, which the previous cached file was missing entirely. One late change — restoring ice_thickness's CF standard_name — landed after that regeneration and so is not in the cached files; it will appear the next time the topography is rebuilt.

Checklist

• User's Guide has been updated
• Developer's Guide has been updated
• API documentation in the Developer's Guide (api.md) has any new or modified class, method and/or functions listed
• Documentation has been built locally and changes look as expected
• Testing comment in the PR documents testing used to verify the changes
• New tests have been added to a test suite

[xylar]

Testing

Polaris omega_pr suite

• Baseline workdir: /lcrc/group/e3sm/ac.xylar/polaris_1.1/chrysalis/test_20260808/omega-pr-ers-fix
• Baseline build: /lcrc/group/e3sm/ac.xylar/polaris_1.1/chrysalis/test_20260808/omega-pr-ers-fix/build
• PR build: /lcrc/group/e3sm/ac.xylar/polaris_1.1/chrysalis/test_20260808/omega-pr-ers-fix/build
• PR workdir: /lcrc/group/e3sm/ac.xylar/polaris_1.1/chrysalis/test_20260808/omega-pr-fix-attrs
• Machine: chrysalis
• Partition: debug
• Compiler: intel
• Build type: Release
• Log: /lcrc/group/e3sm/ac.xylar/polaris_1.1/chrysalis/test_20260808/omega-pr-fix-attrs/polaris_omega_pr.o1265884
• Result:
  • Failures (1 of 22):
    • ocean/column/ekman

The primary purpose of this work is to fix attributes with an eye toward component_inputs to E3SM produced by the e3sm/init component. I will also test the full workflow but this branch need not wait for that testing to go in.