tabs: mark a pane here, pull it into any tab there (C40) — and a rename caret that shoved the name around

[navbytes]

Two independent changes, one commit each.

1. C40 — mark and pull (Alt+Shift+x / Alt+Shift+v)

C28's Alt+Shift+i/Alt+Shift+m carry the focused pane one tab at a time. On a three-tab workspace that is the whole feature; on a ten-tab one it is a walk, and the walk re-parents the pane into every tab it passes through. C40 is the same verb with an arbitrary destination:

• Alt+Shift+x marks the focused pane (marking the marked one unmarks).
• Alt+Shift+v pulls it into whatever tab is active then — same host, same widest-way split, same refusal below C25's floor, focus following the pane, source tab removed if it was its last.

Nothing moves on the mark, and that is deliberate. A pane is a live process; it cannot sit in a clipboard between two tabs. This is tmux's select-pane -m, not a cut.

carry_pane_to_tab is now shared by both verbs rather than copied — the interesting half is the refusal (fit check before anything is removed, emptied-source rule, destination index fixup), and a second copy would be a second set of edge cases.

Chrome: while a mark is pending, C9's hint bar leads with Alt+Shift+v pull marked pane — the marked pane is usually in another tab, so without it this is a mode you cannot see. Pairs drop whole from the right, hence leading.

Mark lifecycle: cleared by the pull, by "already in this tab", by a pane that has since gone; kept on a "no room" refusal, the one outcome you can fix and retry. Also cleared when the marked pane closes — next_pane_id is max+1, so ids are recycled and a stale mark would eventually name a stranger.

Chord choice: x/v come from §8's free pool, reserved there for clipboard vocabulary — this is that vocabulary, on panes rather than text. Only the shifted forms are taken, so bare Alt+x/Alt+v keep reaching the pane through U5 (emacs M-x/M-v) — the same argument that struck b/d out of the pool.

One bug found in the shared path while testing: pulling into the tab you are standing on picked its host through tab_focus_target, which remember_tab_focus only stamps on the way out of a tab — so the pull split whatever pane the tree listed first, not the one under your eyes. pull_pane stamps live focus first.

Docs: C40 contract + key-table row 13d + the spent v/x reservation in DESIGN-ui.md, one row in the README keymap.

2. The rename caret displaced the text it moved through

Reported from live use: moving the cursor back in the rename dialog "starts to displace the characters". It did. U16's ▏ caret was inserted into the string, so it cost a real cell — the field was one column wider than its text, and everything after the point sat one column right. Walking the point left walked that displacement with it.

The caret now rides the character at the point as a reversed cell (theme::attention() — what a terminal's own block cursor is), so it costs no columns and the text cannot move. At end-of-buffer it is a reversed space, appended, displacing nothing.

rename_field returns spans rather than a String, which fixes its three other callers for free — they had the same bug: the C32 pane editor's name row and note rows, and C36's broadcast composer. The append-only carets (P21 search prompt, picker/roster type-aheads) keep the glyph on purpose: no point to move, nothing to displace.

Review notes

• The intermediate commit builds and tests clean on its own (checked, not assumed).
• hint_pairs gained a marked: bool parameter; the render test shim passes false, so every existing expectation is unchanged.
• Both verbs are in the layout fuzzer's action table now (200 seeds x 400 steps, invariants clean) — the mark is the one pane-carrying verb whose source tab is arbitrary rather than active, so it reaches tree/map states the walk-based move cannot.
• Outstanding: the src/ui/** design-supervisor audit CLAUDE.md asks for has not been run on this branch.

Verification

981 unit tests + all 35 integration binaries green; cargo +1.96.1 clippy --all-targets -- -D warnings and fmt --check clean.

The caret fix was verified through the real binary, not just the unit seam: typing abcdef into the rename dialog and pressing Left three times leaves mainabcdef in place where it used to become mainabc▏def, and typing Z there gives mainabcZdef.

A third reported bug — text escaping the floating pane's border — is not in this PR. Driven through the real binary at 120x40 and at the 80x24 floor, filled edge to edge, with ZWJ emoji / CJK / powerline glyphs, and across hide-resize-show, the float's output stayed inside its border every time; blit_screen clips to inner.width.min(cols)/inner.height.min(rows), so roost cannot paint a pane's text outside its own box. Parked until it reproduces.