Simplify context models

[blakeNaccarato]

Please check for similar discussions first

• I searched and couldn't find an appropriate discussion to participate in instead.

Description

• Assume instances of ContextBase to be leaves. Don't touch their context, because it represents the actual context used to construct that instance
• Instances of BaseModel are also leaves
• The above two conditions simplify to stopping traversal at any model instance, assuming they were built bottom-up
• Don't model dump in init/validate base methods
• Maybe don't encode keyword arguments context, just handle them in their own methods and bake in context only? Model validate will share some logic with init, have both call a common method
• String validate should load json strings and then defer to model_validate path
• ContextBase propagates context, ContextModel (ContextStore?) also saves context as an instance variable
• Don't sync/merge context down at end of init, the contexts used at each layer actually represent the contexts used for construction
• ContextShare unifies context when possible, but not for instances already created? And warns?
• Should dumping roll up all contained context? This is important in piping JSON dumps to notebooks, as the JSON blob has only one root context that must reflect all inner contexts. ContextShare should represent this, and maybe it actually warns on mismatched inner contexts (due to bottom-up model construction).
• Rename root model to ContextRoot. Note that root models are incompatible with storing context unless they have a TypedDict or similar.
• ContextBase could have a simple config to avoid passing context down, or even operate in "share" mode? Maybe ContextBase AND ContextStore can config to run in isolate/share modes.

[blakeNaccarato]

Mostly done with this now.