fix(tree): stop focus snapping to open file (#17); never strand focus on hidden tree (#16)

[leboiko]

Summary

Two related file-tree focus-correctness bugs.

#17 — Tree cursor snaps back to the open file (user-reported)

A user on Chrostini reported that with a file open, the left tree always shifts back to the opened file's index after ~1 second, making it impossible to navigate to or open a different file.

Root cause: the filesystem watcher fires Action::TreeDiscovered on every change — and on noisy filesystems (Chrostini inotify emits spurious IN_ACCESS) that's roughly once a second. The handler then unconditionally called reveal_path(active_file), yanking the user's tree selection back to the open file on every refresh.

Fix:

• FileTreeState::rebuild now preserves the selection by path across a refresh (robust to sibling rows being added/removed by whatever change triggered the watcher), clamping to the old index when the path is gone.
• TreeDiscovered only aligns the tree to the open file on the first discovery. Intentional reveals (open file, search jump, link pick) still call reveal_path at their own call sites.

#16 — Focus could be stranded on a hidden file tree

Several handlers set Focus::Tree without checking tree_hidden, landing focus on a pane that isn't rendered. Added a focus_tree_or_viewer() helper and routed the 8 affected sites through it: Tab, last-tab close (key + mouse), search-Esc, copy-menu Enter/Esc, FocusLeft, ExitSearch. The tree-click mouse site is intentionally left alone — it's only reachable when the tree is rendered.

Testing

• [Bug] #17: a failing reproduction test written first (confirmed it snapped back to /fake/test.md), now asserts the user's selection survives a second TreeDiscovered while the first still aligns to the open file — guarding both directions so a no-op can't pass.
• Focus can be stranded on a hidden file tree (missing tree_hidden guards) #16: per-entry-point redirect tests, each asserting both visible→Tree and hidden→Viewer.
• Full workspace suite green (0 failures); cargo fmt --check, cargo clippy -D warnings, cargo deny check all clean.

Closes #17
Closes #16

🤖 Generated with Claude Code

[leboiko]

🔍 Senior-engineer review (3 parallel agents: architecture / quality / testing)

Reviewed 02b1a0e. No 🔴 critical issues. Findings below, with disposition — all actionable items fixed in the follow-up commit.

Verified NOT a problem (false alarm)

• Two agents flagged the first_discovery = selected_path().is_none() heuristic as breaking startup alignment. Traced the actual flow: with a visible tree, App::new populates synchronously and TreeDiscovered never fires at startup — alignment to the open file comes from run() → expand_and_select() → reveal_path() (mod.rs:838), not from TreeDiscovered. So no startup regression. ✅

🟠 Major — fixed

• Heuristic was non-obvious and misfired on a real edge. Replaced selected_path().is_none() with an explicit FileTreeState.aligned latch that reveal_path sets. This is clearer and closes the "all tree files deleted then recreated" edge, where the transient empty selection would have re-triggered a forced realign (a mini-[Bug] #17).
• Mouse tab-close path (close_hit) was changed but untested. Added mouse_last_tab_close_respects_hidden_tree.

🟡 Minor — fixed

• Stale-tree_area_rect mouse path: the tree_hidden render branch never clears tree_area_rect, so a click at stale coords could strand focus on the hidden tree (Focus can be stranded on a hidden file tree (missing tree_hidden guards) #16 via mouse). Added a !tree_hidden guard to the tree-click handler and routed it through focus_tree_or_viewer(). Now all literal Focus::Tree assignments route through the helper, making its doc-comment invariant true.
• Missing branch coverage in rebuild: added tests for the index-clamp branch (selected path deleted → clamp to last row), the empty-tree branch (→ clears selection), selection-follows-path-when-siblings-shift, expanded-state preserved across rebuild, and reveal_path latching aligned only on a real match.
• Doc-comment clarifications on rebuild (first-populate behavior) + clamp-underflow-safety comment; |p| p.to_path_buf() closure form.

Out of scope (intentionally not changed)

• key_handlers.rs:34 config-popup focus restore uses pre_config_focus (restore-previous semantics, may be Editor/Search) — correctly not routed through focus_tree_or_viewer().

Test delta: 6 → 13 new tests. Full workspace suite green (477 lib tests), fmt/clippy -D warnings/deny all clean.