Narrow the four cross-cutting domain mixins, and put mypy on the package

[dovvnloading]

Problem

After #411, backend/domain/ still reported 124 mypy errors, all in
graph.py and the four cross-cutting mixins:

file	errors
groups.py	71
branches.py	26
commands.py	10
graph.py	9
layout.py	8

Every one is the same union-attr shape the per-kind mixins had:
SceneNode.state is typed NodeState | None against a field-less marker,
so node.state.group_width cannot be verified.

require_node (#410) could not reach them. It looks a node up by id and
narrows on its kind, and this code mostly does neither — it iterates every
node in the scene and skips what it does not want, or narrows a node handed
to it as a parameter.

Change

Three additions to the narrowing vocabulary

is_node_of(node, kind, state_cls) — a TypeGuard for a node already
in hand, for the iterate-and-skip shape. TypeGuard rather than TypeIs, so
the repo's declared 3.10 floor is enough and nothing has to import
typing_extensions at runtime. It narrows only the positive branch, so a
continue-on-wrong-kind loop reads as if is_node_of(...) instead of
if not ...: continue. It also requires a non-None state — matching the
and node.state is not None conjunct that every hand-written check it
replaces already carried.

require_node/optional_node take a tuple of kinds as well as a single
one, because group geometry is frame-or-container throughout.
" or ".join of a one-element tuple is that element, so the single-kind
error message is byte-identical.

GroupSizedState — a base holding group_width/group_height, which
FrameState and ContainerState now inherit instead of each declaring.

FrameState's docstring explains at length why those two kinds are not one
class: is_locked and the group_manual_* quartet are meaningless for a
container, and forcing them onto a shared base would resurrect the "every
kind carries fields it never uses" problem the migration exists to remove.
That reasoning is about those five fields and is unchanged.
group_width/group_height are common to both, are named for both by the
wire contract, and are read for both by every piece of group geometry —
which is exactly the code no single per-kind class could describe.

Neither kind's field set changes; the two shared fields are declared one
level up. Nothing serializes by field order (there is no asdict or
fields() call anywhere in the domain package) and both construction sites
use keyword arguments.

Declarations

SceneDocumentParts gains last_chat_node_id, final_deliverable_node_id
and current_chat_id. Without them BranchOps inferred str from its own
first assignment and then rejected the = None two lines later — which is
how SceneDocument's own final_deliverable_node_id: str | None = None
declaration ended up reported as an error against itself.

Four non-narrowing fixes, none behavioural

• _bbox_of_members initialised four accumulators to None on one line and
  tested only the first before using all four. They are set together or not
  at all; the condition now says so.
• layout.py rebound parent — a str elsewhere in the module — to a
  SceneNode, and passed fid if fid is not None else cid to a function
  taking str after a guard that proves it is not None. Renamed, and
  hoisted into a variable the guard can narrow.
• branches.py reused the edge loop variable for a list that also holds an
  optional edge.
• Command.invert/apply took document: object and then read four
  attributes off it. They now take SceneDocumentParts, which declares
  exactly those four.

Scope

[tool.mypy].files now lists backend/domain rather than individual
modules: all 18, up from 3 at the start of the day. 48 source files check
clean.

Test plan

• Full suite: 3,245 passed, 19 skipped.
• The one failure, test_cancel_during_backoff_aborts_promptly, asserts a
  cancel completes within a 2.0s wall-clock budget. It fails the same way on
  a clean checkout of main with no local changes, and passes on the same
  checkout when the machine is idle — it measures the machine, not the code,
  and is unrelated to this change.
• 117 frame/container/group/bbox/measured-size tests pass. That is the real
  check on the _recompute_group_bounds restructure, which is the one piece
  of load-bearing geometry this touches.
• ruff check . clean. mypy clean across 48 source files.

🤖 Generated with Claude Code