feat(editor): scrub-editing and drag-target slots in the Attributes pane (#212)

[JArmandoAnaya]

Third sprint of Pillar P1 — Interaction & Navigation (epic #250). Independent of #210/#211, so it branches from main (rebased on the merged p1-s1).

Closes #212

What this adds

Scrub-editing — drag a numeric attribute's name and the value follows the cursor, live. ⇧ fine (×0.1), ⌘/Ctrl coarse (×10), Esc cancels. On lane Width, Mark width, elevation Height, and signal s / t / Heading offset.

Slots — a framed drop target for asset references: takes a Library drag with a hover affordance, click to jump to its Library category. First consumer: the prop Model slot.

Discovery is why this is small

The transaction model was already there: the panel's edits push edit::Command, and the preview-session pattern (begin_preview/update_preview/commit_preview/cancel_preview) already gives one-undo-entry gestures. So this binds two new widgets to existing machinery instead of inventing a mechanism.

Design calls worth reviewing

• ScrubLabel is a pure gesture source — pixels and modifiers, never units, values, or commands. The panel maps its delta onto each attribute's units through a local descriptor table. Local on purpose: it is the only consumer, and a shared registry is speculative until a second one appears.
• The multiplier applies to the motion made while held, not to the running total. Scaling the total would make the value jump the instant ⇧ goes down. Tapping ⇧ mid-drag refines from where the value has actually reached.
• A net-zero drag is not an edit — it cancels rather than committing, so a scrub that wanders out and back leaves no undo entry.
• SlotWidget resolves and commits nothing. It republishes the dropped key; the consumer decides what a key means. That is what lets the same widget serve props now and materials in P6 without changing.
• The Model slot commits a kernel factory, not an editor macro (approved decision). edit::set_object_model re-points Object::name — the prop_library id the mesher renders — and refreshes radius/height from the new model, so a pine's bounding volume can't survive on a shrub. Public kernel API, so bindings.cpp + place_objects.py move with it (repo policy).

Unplanned but needed

• Selecting a prop showed its owning ROAD's fields. There was no prop section at all, so the Model slot had nowhere to live. Added one, naming the object's OpenDRIVE @type (§13.1).
• LibraryPanel::focus_category clears the search box first — a stale filter would otherwise hide the very category the slot is sending you to.
• Extracted active_node_height() — the Height spin box and its scrub binding must agree on what "current" means; it was inline in the spin box's commit path.

Visual change

Panel renders (offscreen QWidget::grab, macOS, graphite theme). GitHub attachments aren't reachable from the CLI, so they're described rather than embedded — reproduce with a panel.grab().save(...) in an offscreen gtest:

Prop selected (new section):

OpenDRIVE id  object s1
Position      s 12.00 m, t 5.00 m
┌─ Prop ───────────────────────┐
│ Tree                         │
│ Model  [ tree_pine        ]  │   ← drop target, click to browse
└──────────────────────────────┘

Lane selected — unchanged in appearance: Width and Mark width read exactly as before; the only difference is a horizontal-resize cursor over their names. The empty 48 px preview box was dropped from the slot after the first render — with no thumbnail it read as a broken image.

Tests

• NEW test_scrub_label.cpp (8): slop, running delta, direction, both multipliers, the while-held rule, Esc, right-button.
• NEW test_slot_widget.cpp (7): the MIME contract, non-library and empty payloads refused, hover affordance appears/clears, engage, placeholder.
• test_panels.cpp (+5): one-undo-entry per gesture, Esc restores the baseline byte-identically, net-zero pushes nothing, the Model slot commits/undoes, an unknown model is refused without an undo entry.
• test_library_panel.cpp (+3), test_object_ops.cpp (+3: retarget/revert byte-identical, dirty set, bad input leaves the network untouched), test_edit.py (+2).

A test caught a real thing: the first version picked section->lanes.front() — the centre lane, which carries no <width> records. The panel correctly refuses to scrub it (baseline() returns nullopt); the test was wrong and segfaulted on widths.front(). The fixture now picks a width-bearing lane.

Verification

• 796/796 green on ci-macos (-Werror); whole-tree clang-format --dry-run --Werror clean.
• ASan + UBSan: the new editor and core suites clean.
• Python: 79 passed / 1 skipped; place_objects.py runs and prints the swap (command-path tree is now a shrub (r=1.10 m)), with the undo restoring the oak.

Acceptance

GW-2 step 9 (scrub path); GW-3 steps 6–7 and 10 slot mechanics (their real consumers land in P4/P6).