Record a displacement edit as the brush it edits

[saworbit]

Part of #761. The displacement family, which was the biggest group of commands still recording the whole level on every edit.

What changed:

• dock._try_undoable_action() was a hand-rolled capture/call/register-a-restore-pair. It now goes through HFUndoHelper.commit_completed() and takes an optional brush scope. Its five displacement callers pass one.
• The elevation spin and the sew group spin already went through HFUndoHelper.commit() and only lacked the scope argument.
• The sculpt drag took a whole-level snapshot on mouse-down and another on mouse-up, on a gesture that is held down. It was already holding the brush id it is sculpting. It now records that brush.
• Painting a material onto a brush now scopes too. Only four commands can claim a brush scope, and the rest never reach the helper that offers it #761 left it open because a paint that wrote the palette could not claim a brush scope. It does not: the paint sets material_override on the node, and materials is unchanged even for a material no palette slot holds. There is a test for exactly that.
• commit_completed() and register_action() now share one _after_state() for the do side, so the fallback when a scope did not survive the command cannot drift between them.

Sewing keeps the whole snapshot. sew_all() pairs displacements by sew group across the level, so it changes brushes the caller never named. There is a test that pins that too, so a later change cannot quietly scope it.

One bug this found in itself: a stroke's pre-state is a brush scope now, and the Escape-cancels-a-sculpt path was handing it to restore_state(). A scope holds two keys and a level state holds twenty-five, and restore_state() clears what it is not given, so that would have wiped every entity, material and visgroup in the level. The stroke now carries the scope ids beside the state and one function owns which restore matches, used by the cancel, the recovery and the normal end.

Tests: tests/test_scoped_undo_step.gd runs each newly scoped command between two whole-level captures and asserts brushes is the only key that differs, plus a round trip per command, plus the Escape regression in tests/test_plugin_shortcuts.gd.

Still open on #761: bevel and inset are the same hand-rolled shape and are scopeable; the decal, path, polygon, drop and prefab sites change which brushes exist, so most of them cannot scope at all; and the entity scope is still the design question the issue parked.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.