fix: tmux pane geometry, palette debug snapshot, and stale socket tests

[arzafran]

Follow-up to the test triage in #437. These socket and palette tests were failing in the nightly. Most of the failures were outdated tests, and three were real bugs, fixed here.

App fixes

• tmux compatibility reported the wrong pane geometry. pane_top came out 3 rows too high because pane frames are in window coordinates while the container starts below the titlebar. The window size was also roughly halved on Retina, because it was divided by the pixel cell size instead of the point size. pane.list now includes the container's x/y, and the tmux shim subtracts that origin and divides by the point cell size.
• The debug palette snapshot only showed the first 20 rows, while the RPC allows 100. Tests asking for 80 rows never saw commands lower in the list, such as Rename Tab.
• debug.shortcut.set (debug builds only) now accepts new_window, close_window and rename_tab. Without them the shortcut hint sync test could never pass.

Test updates

Each of these follows a change that was intentional, and each has its commit cited in a comment:

• test_notifications.py:
  • The test now refers to surfaces by id; bare indexes were deliberately rejected.
  • It now selects new workspaces explicitly, since workspace.create no longer steals focus over the socket.
  • A notification for the focused panel is still recorded; only the system banner is suppressed. A unit test already covers that behavior.
• test_review_diff_panel.py:
  • Untracked files are listed but don't count as diffable, so the test now checks diffable >= 1 and file_count >= 2.
  • The test now switches the shell to zsh first. On the CI runner's bash 3.2, the pasted comment text would otherwise run as commands.
• Codex hook install: the stale entries now use the real older command form. Since fix: trust installed Codex hooks #326, ownership is an exact match, so a made-up line is correctly left alone.
• Palette search tests:
  • open can rank either Open Folder or one of the per-app open commands first, and the test accepts either.
  • The palette now opens with > already typed.
  • The test now waits for the rows to fill in.
• Agent screen detection: the fake agent now clears its approval prompt, the way a real one does.
• CommandPaletteAllSurfacesUITests:
  • The test socket now lives in the runner's own temp dir; the sandboxed runner gets EPERM connecting to a socket under /tmp.
  • The backspace test now finds the palette row by its identifier instead of by title, so it no longer matches the sidebar row with the same title.

Seven tests_v2 files join the nightly list, and the backspace UI test comes off the skip list.

Still open

• The Kitty OSC 99 checks in test_notifications.py still fail, because the pinned ghostty fork never dispatches OSC 99. That needs a fork commit.
• Three palette UI tests and the workspace description UI test stay skipped.

All of the above passed in manual nightly runs on this branch: 39292783144 and 39294950284 for v2, 39297584258 and 39298963225 for UI.