### System details

AMD Raphael iGPU + RTX 5070, Omarchy 4.0.0.alpha, bash 5.x, foot and Herdr 0.8.2

### What's wrong?

`_ssh_disarm` in `default/bash/fns/ssh-reconnect` runs after **every** `ssh`
invocation, and it ends with `\e[?1049l` — leave the alternate screen.

Leaving the alternate screen also restores the cursor, the way DECRC does. When
nothing ever entered the alternate screen there is no saved cursor to restore,
and DEC semantics put the cursor at home. The shell then draws its next prompt
and output over lines already on screen.

No remote host is needed — `ssh -V` is enough, and so is an ssh to a name that
does not resolve.

**Reproduce**

```bash
clear
for i in 1 2 3 4 5; do echo "BEFORE-$i"; done
ssh -V
for i in 1 2 3; do echo "AFTER-$i"; done
```

Expected: `AFTER-1..3` below `BEFORE-1..5`.

Actual: the cursor is back at the top, so `AFTER-1` lands on `BEFORE-1`. Since
`BEFORE-1` is one character longer, the row reads `AFTER-11`, then `AFTER-22`,
`AFTER-33`, with `BEFORE-4`/`BEFORE-5` still below — the overlapping text that
gets cleared by hand.

**Isolating it**

Bypassing the wrapper is clean, so it is the wrapper and not ssh:

```bash
command ssh -V     # clean
ssh -V             # scrambled
```

And within the disarm, it is specifically `1049l`:

```bash
printf '\e[?1000l\e[?1002l\e[?1003l\e[?1006l\e[?1004l\e[?25h'   # clean
printf '\e[?1049l'                                              # scrambled
```

**Where it shows up**

| Terminal | Result |
| --- | --- |
| foot | not affected — ignores the reset when not in the alternate screen |
| tmux | not affected — same |
| Herdr | **affected** — applies the reset, cursor goes home |

Herdr is one of the environments this wrapper's own comment names ("A remote
tmux, herdr, or editor arms terminal modes over the SSH pipe"), so the cleanup
misfires in an environment it was written for. Emulators that ignore the reset
are being lenient — restoring the cursor with nothing saved is the documented
DEC behaviour, so the unconditional send is the fragile part.

**Suggested fix**

Keep the mouse-tracking (1000/1002/1003/1006), focus-reporting (1004) and cursor
(25h) resets unconditional — none of them can move the cursor. Send `1049l` only
for the case the wrapper exists for: a session that was established and then
dropped. That is the same `rc == 255` plus elapsed-time pair the reconnect path
already uses to tell a real drop from a connect or auth failure; a fast 255
never had a session and cannot have armed anything.

PR to follow.