(remote): push liveness over a persistent watch channel

[devsuitup]

Closes #240.

A remote session's state could be five minutes stale, and lowering the poll
interval only multiplies handshakes — the OpenSSH shipped with Windows has no
ControlMaster, so every cycle is a full connect plus authentication, measured
at 416–585 ms whatever the command.

This adds a persistent push channel per host alongside the periodic cycle. The
cycle stays exactly as it was: it remains the reconciliation pass, and the only
thing covering a host the channel cannot watch.

Shape

remote-watch.js — one long-lived ssh -tt child per alias running
inotifywait, one line per filesystem event, restarting on exit with the same
backoff remote-index.js already uses for host failures (hoisted and shared,
not a second policy).

Events are told apart at the source: .claude/sessions means a descriptor was
rewritten, .claude/projects means a transcript grew. Paths are checked with
the transport's existing isSafeRelPath; anything else is dropped rather than
guessed at.

remote-index.js gains refreshHostNow(alias) — the narrowest possible entry
point, guarded so an event-driven refresh and the periodic cycle can never run
on the same host at once. The signal lands in the existing notify throttle; no
second path to the renderer.

Verified against the live host, with a real session

The whole chain was run against the declared host — not only unit-tested. A
real Claude session was started on it in tmux and asked a question, with the
shipped buildSshArgs and parseWatchLine driving the channel:

+4s   session  3610858.json                            (descriptor rewritten on submit)
+5s   project  -tmp-sb-live/7eaba321-….jsonl  ×9        (transcript growing)
+11s  project  ×2
+13s  project  ×4
+14s  project  ×4
+14s  session  3610858.json                            (descriptor rewritten on idle)

The liveness signal arrives at once instead of waiting up to five minutes.

Orphans: measured twice, zero left behind. -tt is load-bearing — without a
TTY the remote inotifywait survives the local ssh being killed, one orphan per
restart, forever. A stdin-guard wrapper was tried and still leaked. With two
backgrounded watchers behind a wait, -tt still takes the whole process group
down.

Parser against real traffic: transcript and descriptor paths parse;
inotifywait's own startup chatter and any ../ traversal are dropped. Note
that with -tt the remote stderr merges into stdout, so that chatter does reach
the parser — it is dropped, by design.

One correction on top of the original commit

The coalescing window was one second. Since every coalesced signal opens a full
ssh cycle, a continuously writing session would have driven roughly one
connection per second — 3600 an hour against a host that gets 12 today, the
opposite of the point. It is now 15 s. The leading edge still fires at once, so
an idle-to-active transition is immediate; sustained activity costs at most four
connections a minute.

Degradation

inotifywait missing on a host is not a host failure: the channel marks that
host unwatchable, logs it once, and stops retrying. The periodic cycle already
covers it. inotify-tools was installed on the declared host for this work
(4.23.9.0); no other host is assumed to have it.

Mutation proofs

-tt dropped; the unwatchable guard removed; the coalescing gate disabled; the
path check removed; the backoff replaced with a linear formula; the
in-flight guard removed. Each turns its own test red and nothing else. The
backoff test had to chain three exits — a single exit cannot tell exponential
from linear at the first failure.

task check: 1184 pass, 0 fail, 8 pre-existing skips, lint 0 errors.