[patnr]

Distance-based localization (distance_loc) could not be used on a state that holds scalars besides a field, e.g. state = ["permx", "sor", "vo"]:

• A parameter with no localization entry gets an all-zero mask, sized from its prior's nx/ny/nz. A prior without a grid (Draw a prior without a grid as scalars, instead of dropping it #180) has none, so the analysis raised KeyError: 'nx'.
• Even with a grid, such a parameter could only be frozen: an entry places a kernel on the field, so there was no way to have a scalar updated.

Change

• A parameter without a grid is sized from its mean, as StateLayout.from_prior_info draws it (and has no inactive cells to drop).

• A new unlocalized key lists parameters that every datum updates with weight one:

  [dataassim.localization]
  name        = "distance_loc"
  taper_func  = "gc"
  unlocalized = ["sor", "vo"]
  ...

  It must name parameters of the state; anything else is an error. A parameter that is neither in unlocalized nor in an entry keeps getting weight zero, as before.

Verified

• Three new tests in tests/assimilation/test_distance_loc.py (gridless parameter without an entry, unlocalized parameters of both kinds, a misspelt name); all three fail without the change.
• End to end, ES-MDA with distance localization runs on the MiniRes benchmark case of Add a benchmark of the update schemes on MiniRes twin experiments #184 (log-permeability field plus sor and vo), with Gaspari-Cohn tapers at the wells for the field and the two scalars unlocalized.
• Full suite: 626 passed (with the minires extra). ruff clean.

🤖 Generated with Claude Code