[mtao]

Stacked on #26 (and through it on feature/controls-model). This replaces the earlier draft of this PR: it is rebuilt on the current scene graph, keeps the useful parts, and fixes or drops the rest.

Object

• Style: trailing return types throughout.
• Neither copyable nor movable. Children, features and signal listeners refer to an Object, so moving one was never safe, and nothing in the codebase moved one. This also removes the draft's problem of signal connections and index pointers going stale after a move.

Traversal

for_each_descendant (pre-order; return false to skip a subtree), for_each_ancestor (return false to stop), find_descendant, and is_descendant_of.

These replace hand-written walks in:

• the ImGui panel's drag-drop check and duplicate-then-add-drawables code
• the Qt model's re-parent check
• MeshScene's first-light lookup

Code that iterates direct children (row counts, root-level mesh lists) stays as it is.

Signals

Built on sigslot v1.2.3, added as a wrap.

• feature_added / feature_removing
• child_added / child_removing, also emitted when re-parenting
• destroying, emitted at the start of destruction. The destroyed object's own features and children go with it without per-item signals; each child emits its own destroying.

A small Signal<Owner, Args...> wrapper lets anyone connect/disconnect but only Object emit, so listeners can trust an event really happened. In the draft, the signals were public members anyone could fire.

FeatureIndex

Maps each feature type to the objects in a tracked tree that carry a feature of exactly that type: track(root), objects_with<F>(), count<F>().

Fixes over the draft:

• Destruction: destroyed objects leave the index. The draft kept dangling Object*.
• Re-tracking: tracking an already tracked object does nothing. The draft duplicated its connections and entries.
• Several features of one type: the object is listed once, and stays listed until its last feature of that type is removed. The draft listed it twice and dropped it on the first removal.
• Exact types only: queries match the exact feature type, and the documentation says so.

Dropped from the draft

• The per-object type map behind find_feature. It returned the last feature of a type instead of the first, and missed features after a removal. Objects have a handful of features, so the linear scan is already cheap.
• AbstractFeature::on_modified / mark_modified(). Nothing called them. MeshGeometry and BasicShading use version counters for change tracking.

Tests

• Traversal order, skipping subtrees, stopping early, and is_descendant_of.
• Signal sequences, including re-parenting and destruction; that emit isn't accessible outside Object; that Object isn't copyable or movable.
• FeatureIndex: existing and later features, several features of one type, new, detached and re-attached children, exact-type queries, and either side being destroyed first.
• meson test: 13/13 suites pass. test_scene_graph: 591 assertions in 67 test cases.