I'm OK with this, but I would prefer not to.

When I see:

class COOStructure:
   ...

I know that everything I need to read is in front of me. When I see:

class COOStructure(Structure):
    ...

I have to cross-reference. The cross-referencing is worthwhile if we can avoid significant repetition. But given that the among of DRY-ness that this achieves is quite small, I would be happier repeating this out in the specific classes.

However, YMMV, and if this seems much better to Team DLS, I do not object.

After #1036 I want to change e.g.

STRUCTURE_TYPES = OneShotCachedMap(...)
# to
STRUCTURE_TYPES = OneShotCachedMap[StructureFamily, type[Structure]](...)

When we then access the types inside, I want typing and my IDE to help me out with

            structure_type = STRUCTURE_TYPES[attributes["structure_family"]]
            self._structure = structure_type.from_json(attributes["structure"])

Structure could become a Protocol here, since it's just a contract that the Map will only contain types that can be deserialised.

From trying this out, it seems Pydantic is not happy if you try and use protocols with a generic BaseModel with bounds, so I've had to revert to an ABC.

class Proto(Protocol):
    ...

P = TypeVar("P", bound=Proto)

class DataSource(BaseModel, Generic[P]):
    t: Optional[P]

        form_keys_touched = set(report.data_touched)

Looks like it's a set in the only instance it's used. I'll make it an Iterable to prevent issues.