Interface editor: typed edits, tree, multiselect, versioning

[EthanRStokes]

Requested by Treetrain1 · project thread

Before: The Interface workspace still rewrote the whole document with set_interface for adding, deleting, duplicating and most inspector fields. It had a flat hierarchy list, single selection, no schema version, and every commit reloaded the world.

After: Those actions are typed, undoable edits in the begin/update/commit transaction used by move and resize, also available from shell and MCP. The hierarchy is a searchable tree with editor-only Lock and Hide. Widgets can be multiselected (Shift-click, marquee), moved as a group, aligned and distributed. The interface document is versioned, and committing with a design preview open no longer reloads the world.

What changed:

• New UiEdits: Create, Delete, Duplicate (remaps parents and scroll targets inside the copy), Batch (all or nothing), SetDocument, SetClass; SetProperty now covers range, style, class, bindings, items, tooltip, scroll target, tab page and transition.
• interface-properties lists every property path with label, group, type and applicable kinds; a test keeps it in step with the edit enum.
• UiDocument.version (1); old documents migrate on load, newer ones are refused rather than rewritten.
• InterfaceDesign.hidden (protocol 36) hides widgets in the preview only.
• docs/interface-editor-plan.md is updated against recent master changes and lists what remains; AGENTS.md has an Interface editor section.

How: core in blockloom-core/src/ui_framework.rs, backend in commands.rs/dispatch.rs/shell.rs, runtime hide in ui_design.rs, UI in UiDesigner.qml.

Tested locally: core and runtime design-session tests, protocol tests, the backend interface transaction tests, clippy -D warnings and fmt. QML lints clean. just qml-test has now run: 169 pass; PhysicsInspector::test_aConflictNeedsTheConversionBeforeApplying fails independently of this PR.

Follow-up on this branch (verified on a Qt 6 + RTX 5080 machine):

• Qt Quick tests run and fixed: two viewport tests toggled checkboxes without firing onToggled, and the hierarchy's Ctrl/Shift multiselect never received modifiers (it read a mouse argument ItemDelegate.onClicked lacks).
• Text input: the caret, selection and IME underline are drawn from the text layout (input_caret.rs, parley geometry), and click/drag places the caret by character index. The | and [..] marks in the text are gone.
• Animation: Opacity, OffsetX, OffsetY, Rotation tracks plus an authored fx property. Offset rides margins/insets and rotation the transform's rotation, so the anchor's translation is untouched; opacity multiplies down the tree (slider tracks and fills are not faded yet). The timeline has Preview and a scrub slider via InterfaceDesign.animation (protocol 37); editing pauses while previewing and nothing is saved.
• Hardware: ignored GPU test interface_design_geometry_follows_viewport_dpi_and_safe_area (six viewport/DPI/safe-area/scale cases) passes through the embedded runtime. Still open: process runtime matrix, real image/font loading, presentation timing, screenshot baselines.
• cargo test --workspace: all pass except screen::shell::tests::keys_type_their_keycode_and_unknown_ones_are_refused, which passes alone and looks load-sensitive. Clippy is clean for this branch's code.

🤖 Generated with Claude Code

https://claude.ai/code/session_013K2j7WKJ3rVme2h7LSfvc6

[EthanRStokes]

CI test failed on one test, android::tests::boot_state_reads_sys_boot_completed (890 passed, 1 failed). It runs a stub adb shell script and does not touch the interface editor code this PR changes, so the failure is not from this PR. Likely a flaky stub-script exec in the test. I re-ran the failed job once; if it fails again I will look at the test itself.

---

Generated by Claude Code