Open World Patient Ranking Pipeline

[eveenhuis]

Rebased version of !283, includes changes from !273

See the description from those below:

• Fixes unit tests broken by merit focus weight modification from ADEPT 2026-01-21
• We weren't finding that medical urgency/KDMA(s) varied across actions for the same character.So, we are now regressing based on character descriptions only (or non-character actions, but that doesn't really apply due to the available actions). [OWFormatChoicesADMComponent, then CR/DR]
• We use alignment to pick the highest ranked character (or non-character action) [Existing alignment function].
• That character may correspond to multiple actions, so we then have a secondary non-aligned prompt to choose which sub action to do. [OWChoiceToActionADMComponent]
• Once we've picked a single action, we may have details to fill in like tag color or character ID, so we then do that with followups (again, not aligned) [OWActionParameterCompletionADMComponent]
• Then finally, we do a tag adjustment based on a heuristic to "align" tag color. [OWTaggingAdjustmentADMComponent]. I didn't want to name this component alignment because it doesn't directly use the alignment target. Instead it considers the KDMA ranking vs medical only ranking as well as how other patients have been tagged so far

[dmjoy]

Mostly minor comments and overall LGTM. Perhaps most pressing would be to add a bit more to the Changelog, but the other things can be deferred to another PR (or included in a subsequent PR for OW).

[dmjoy]

These should probably be parameterized (i.e. added to the __init__ args)

[dmjoy]

Wondering if it makes more sense to have new adm configs (I guess we would just need one for comp reg and one for direct reg) to base these experiment files off of.

[dmjoy]

I suspect there's a few more things we'd want to add to the changelog (new components etc.)