fix(undo): restore deleted entities under their original parent (#75)

[dreamersilly]

Fixes #75.

Both reported symptoms are one bug, and it's in the delete-undo restore path rather than in anything UI-specific.

What happens

snapshot_entity_subtrees deliberately keeps ChildOf — its doc comment says so — "so restore puts each entity back under its original parent". But the parent is outside the snapshot: only the deleted roots and their descendants are captured. So the parent's id is never in the EntityHashMap that write_to_world remaps references through, and SceneEntityMapper::get_mapped maps anything absent to a freshly reserved dead id:

/// Returns the corresponding mapped entity or reserves a new dead entity ID in the current world if it is absent.
fn get_mapped(&mut self, source: Entity) -> Entity {

(bevy_ecs-0.19.0/src/entity/map_entities.rs:273)

The relationship hook then sees a ChildOf pointing at an entity that doesn't exist and discards it. The restored entity has no parent at all — which is symptom 1, hierarchy not respected.

Symptom 2 follows from the same thing. A UI node that loses its ancestor chain is no longer under its canvas, and UiTargetCamera / ComputedUiTargetCamera are denied at snapshot time on the assumption that they get re-derived from that ancestor. So the modal comes back as an independent root node with no target camera and lands wherever the default layout puts it — the box at a random position in the viewport.

Worth noting this was never UI-only. Any nested entity had its parent link dropped on undo; UI just makes it obvious because the misplacement is visible rather than a silent change in the hierarchy panel.

The fix

Identity-map every live entity that isn't itself part of the snapshot before the write, so references pointing outside it resolve to the real entity instead of a dead placeholder. Entities that are in the snapshot are excluded from the seeding, so they still receive fresh ids and the links between them follow the remap rather than the stale originals.

The returned map is narrowed to the snapshot's own old→new pairs afterwards — the identity seeds are scaffolding for the remap, not results. The only caller, DeleteEntitiesCmd::undo, looks up restored roots in it, so it's unaffected.

This also covers any other component holding a reference to a still-live entity outside the snapshot, not just ChildOf.

Verification

The first test fails on current main with exactly the reported behaviour:

assertion `left == right` failed: restored entity must point at the original live parent
  left: None
 right: Some(2v0)

left: None is the whole bug — not the wrong parent, no parent.

Both tests pass with the change. Full suite: 695 passed, 0 failed (693 before, plus these two). Clippy clean on renzora_engine under the workflow's flags.

Tests run with --profile dist; the debug profile can't link on Windows (too many exported symbols (got 878424, max 65535)).

One caveat

CI on this branch will be red, and not because of this change — main has been failing since 2026-06-14. #76 fixes those two causes. This branch is cut from main so it inherits them; it should go green once that lands, or if you'd rather I rebase this on top of #76 to demonstrate that, say the word.

I haven't driven this in a running editor — the verification above is the test, not a repro of the screenshots. If the reporter can confirm the modal comes back in place, that closes the loop.