refactor: lift pointer hit-testing out of Editor into a pure layer

[sinelaw]

Follow-on to #2943, applying the same decision-core/imperative-shell split to the mouse side.

The problem

mouse_input.rs carried the "what is under the pointer?" ladder as impl Editor methods reading self.active_layout() / self.active_chrome() directly. The precedence rules — which overlay wins where two of them cover the same cell — could only be exercised by constructing a whole editor, so they never were. These are the rules that surface to users as "the button doesn't work".

There were also two implementations of the same probes: compute_menu_dropdown_hover (menu_actions.rs) and compute_file_browser_hover (file_open_input.rs) each duplicated hit-tests the hover ladder also performed.

The change

app/hit_test.rs (new, no impl Editor) holds the ladder as free functions:

• hover_target = floating_overlay_target (context menus → suggestions → popups → file browser) then chrome_target (menu bar → open dropdown → explorer → separators → split controls → tabs → scrollbars → status bar → search options).
• menu_dropdown_target and file_browser_target live here too; their former impl Editor wrappers now delegate, so each probe has one implementation instead of two.

The view borrows exactly the rects and slices the decision reads — deliberately not the layout caches themselves. Naming each field means the shell has to state what it hands over, a newly-added layout field cannot silently drift into the hit-test, and tests construct precise values rather than blanking a cache and trusting that the zeros are meaningful.

One probe cannot be pure: the file-explorer status indicator resolves its columns through the theme, the plugin decoration/slot caches and the explorer renderer's slot layout. It stays on Editor as explorer_status_indicator_at and is passed in pre-resolved — so the ordering still lives in the ladder even though that probe's geometry does not.

Notes for review

• Behaviour-preserving: the ladder is the same sequence of the same comparisons, in the same order.
• mouse_input.rs drops ~400 lines net; PopupAreaLayout is re-exported pub(crate) for the view's popup slice.
• 10 new unit tests cover precedence rules that had no coverage at all before, because reaching them previously required a whole editor: overlay-beats-chrome-underneath, topmost-popup-wins, scroll-offset suggestion rows (and past-the-end fall-through), explorer close button vs status indicator vs resize border, split controls on the tab row, scrollbar thumb vs track, separator orientation, status-bar segment span.

Still ahead (not in this PR)

handle_mouse_click is a second ~1,100-line ladder over the same geometry — menu bar, tabs, scrollbars, status bar, explorer, split controls — with hit-test and effect fused together. Two hand-maintained ladders over identical rects is the drift risk the overlay Layer work was created to eliminate. With the pure ladder in place, the click path can become resolve target → match → apply effect, the same shape handle_key now has.

Testing

• cargo check / cargo clippy --all-targets / cargo fmt --check (--all-features) clean — no warnings in touched files
• cargo test --all-features --lib — 3,313 passed, 0 failed

🤖 Generated with Claude Code

https://claude.ai/code/session_014cyMxXzFLeS94CrnXuXbJf

---

Generated by Claude Code