Addresses five divergences from OpenSSH sshd surfaced during review.

Channel-close hook (closes #1 of review)
  Previously openChannelSet.remove fired only on a SendRequest failure,
  so a connection that opened and cleanly closed many channels
  accumulated dead entries forever. any() always returned slot 0 — the
  oldest, most-likely-dead channel — degrading the OpenSSH-mirror mode
  into "always probe a dead channel, then fall back to global" within
  one tick. trackingNewChannel.Accept now proxies the per-channel
  request stream through a buffered (16, matching gossh's chanSize)
  forwarder goroutine; when the upstream reqs channel closes, an
  onClose callback prunes the set.

Remove the time.After race (closes #2 of review)
  connectionKeepAlive used a replyCh + time.After(interval) race
  around SendRequest. A reply arriving at interval+epsilon left
  keepAlive.Reset() un-called, ticking the false-disconnect counter
  for live-but-slow clients. Removed the outer timeout entirely. The
  inFlight semaphore already prevents overlapping probes; TimeIsUp
  enforces the deadline at the next tick; sshConn.Close on TimeIsUp
  unblocks any hung SendRequest.

Reset on any inbound traffic (closes #3 of review)
  OpenSSH resets its client-alive counter on every received packet.
  Added SessionKeepAlive.NotePeerActivity — like Reset but without
  bumping the KeepAliveReplyReceived metric. Called from
  Server.handleRequests (global requests), session.handleRequests
  (per-session requests), and the per-channel request forwarder in
  trackingNewChannel. Resets the ticker so the next probe fires
  ClientAliveInterval after the most recent activity, which also
  closes #4 (probe suppression on recent traffic) for free.

Unwrap helper + ChannelHandlers doc (closes #5 of review)
  Added Unwrap() gossh.NewChannel on trackingNewChannel so downstream
  handlers that need the underlying type can recover it. Documented
  the wrapping on Server.ChannelHandlers.

Test added: TestConnectionKeepAlivePrunesClosedChannels opens three
channels, closes them, then asserts subsequent keepalives over a 1s
window are global-only (0 channel-typed, >=1 global). Without the
close hook, the dead channels in the set would force channel-typed
probes that fail and prune one-at-a-time over many intervals.