Type the remaining 13 mixins, closing the *Ops attribute gap entirely

[dovvnloading]

Problem

#397 did this for settings_store and named the rest as still outstanding: 23 mixin classes across three families, each using attributes that only exist on the class composing it. Correct at runtime, invisible to a type checker reading one mixin on its own — and a large part of why [tool.mypy].files has stayed at four entries while the real error count grew.

This finishes the other two families.

Change

backend/domain/_composed.py declares what SceneDocument provides its four mixins (nodes, edges, command_log, redo_stack, and the handful of methods they call across the composition). backend/agent_dispatch/_composed.py declares what AgentDispatcher provides its nine (_runs, _settings_manager, _dispatch, _run_node_blocking_action).

mypy backend:                  1,059 errors in 68 files  ->  844 in 52
"*Ops has no attribute" errors:                    354  ->  0

Both bases are TYPE_CHECKING-only: at runtime they are empty, add no attributes, no methods and no __init__, and land immediately before object in their composed class's MRO where they can never take precedence over a real mixin.

SceneDocument MRO tail  : ['CommandOps', 'SceneDocumentParts', 'object']
AgentDispatcher MRO tail: ['ContentDispatchOps', 'DispatcherParts', 'object']
SceneDocumentParts adds : []
DispatcherParts adds    : []

What this does not fix

327 of the remaining 666 union-attr errors are the NodeState | None gap — SceneNode.state typed against a field-less marker class. That is not a declaration problem, and it is not tractable here: tests/test_node_state_migration.py requires every migrated field to be read as X.state.<field> and rejects the local alias a narrowing accessor would need. It belongs with the ADR-002 stage 2.5 work.

So mypy's configured scope does not widen in this change. It becomes possible to widen later, which it was not before.

Test plan

• 12 new parametrised gate tests covering all three bases together: no runtime body, no __init__, last in the MRO, and every *Ops mixin in each family actually inheriting its base. That last one catches a new sibling landing without one, which is precisely the hole these close.
• Full suite: 3219 passed, 19 skipped. ruff and mypy both clean.

One process note for the record: the first attempt inserted the new imports by a line heuristic and landed one between a @dataclass decorator and its class, breaking the module on import. Redone with an AST-located insertion point after the real end of each import block, and the 13 files reverted cleanly in between.

🤖 Generated with Claude Code