[saworbit]

Update on a structure nobody changed rebuilt every piece under a new id and recorded a step. That broke anything that named the old ids (a wire, a cordon round them, a prefab record) and left an undo step that undid nothing.

• HFGeneratorSystem.update_changes_nothing(generator_id, settings) is the one check. It is true when the settings build the same structure as the record's and every piece the record names is still there, still this structure's, and still the shape and place it was built in (edited_brush_ids() empty). A piece with no signature, or whose id now names a brush that is not ours ([Review] Rebuilds: prefab and generator rebuilds each re-carry per-copy state field by field, so every new property is lost until reported #1221), counts as a change, so the rebuild goes ahead as before.
• same_settings() reads both sides through the schema's merge and compares floats with is_equal_approx. A plain == fails on 4 against 4.0, and a SpinBox hands back some on-step values a rounding away (0.4, 1.2, 1.7 on a 0.05 step).
• The dock's Update (the section's button, the toolbar's Upd, the palette) says "Nothing to update" and records no step. Over the bridge, update_structure keeps every id and answers no_change: true. With Record no bridge step for a call that changed nothing, and replace the default cordon #1378 the dispatcher then records no step either; tested together locally.
• A structure moved or turned as a whole still counts as unchanged, because a rebuild puts it back where it stands. regenerate() itself still always rebuilds.

Tests: seven in test_structure_dock_commands.gd (unchanged Update from the toolbar and the section keeps ids and records no step; every type at its defaults; settings from outside the dock with 1.7/0.4/1.2; a setting change, an edited piece, a piece moved on its own and a deleted piece still rebuild with one step). Two in test_mcp_dispatcher_structures.gd, fifteen predicate tests in test_generator_system.gd, one line in the MCP structures vibe scenario. Mutation checked on each guard. Local: test_generator_system 104, test_structure_dock_commands 29, test_mcp_dispatcher_structures 16, test_live_generator 45, test_mcp_command_sweep, and more, all green. With #1378 merged in on a scratch branch: the sweep, the structure and command dispatcher tests and the structures scenario pass.

Also folded in: the dock committed Update without the absolute redo Create got in #1275, so a redo of an Update minted new piece ids and a later step that named one found nothing. One argument in dock_brush_handler.gd, pinned by test_redo_of_an_update_gives_back_the_pieces_it_made (failed on the old call).

Fixes #1368