* fix(multiselect): off-by-one in redraw causing menu to drift up each keypress

`reposition_and_write` was moving the cursor up by `last_line_count` to
return to the top of the previous render, but the redraw loop emits no
trailing newline after the last line — so the cursor sits at the *end*
of row N-1, not at row N. Moving up by N landed one row above the prior
top, and every keypress shifted the menu up by another row (visible in
`mise up -i` and any other multi-iteration `MultiSelect` redraw). Fix
by moving up `last_line_count - 1`, matching the invariant the pre-#129
code expressed via `height = lines().count() - 1`.

The shrink branch had the same off-by-one in mirrored form; restructure
it so the cursor lands on the last line of the new render, preserving
the same invariant for the next iteration.

Adds two regression tests that drive the redraw path through a
`Term::read_write_pair` whose writer captures bytes into a shared `Vec`,
replays the bytes through a `vt100` emulator, and asserts that the
cursor row is stable across iterations (drift case) and repositions
correctly when the render shrinks. Both tests fail without the fix —
the drift trace `[26, 25, 24, 23, 22, 21]` is the smoking gun.

Drive-by clippy fixes across the workspace (collapsible match in select,
needless `vec!` in spinner test, needless `return`/borrow in examples
and the non-TTY integration test) so `cargo clippy -- -D warnings`
passes again on stable.

* ci: add test-ci aggregator job for branch protection

Mirrors the pattern in jdx/mise — a single one-minute job that needs
the matrix `test` job and re-emits its result. Lets branch protection
gate on one check (`test-ci`) instead of all three matrix instances
(ubuntu/macos/windows), so adding or removing platforms doesn't
require rule updates.

Uses `if: !cancelled()` rather than `always()` so workflow cancellation
doesn't spin up a wasted runner.