ares-novacom: publish a host port on the device with --reverse

[mariotaku]

Stacked on #18 — that PR carries two fixes to --forward as it ships today (throughput, and a forward that can't tell a dead tunnel from an idle one). This one is the feature alone. Merge #18 first and this retargets to main cleanly.

--forward only carries connections one way. This adds --reverse for the other: the device listens, and every connection it accepts is pumped to a port on this machine, so an app under development can reach a dev server on the host at localhost:<device port>.

# TV apps/scripts reach the host's dev server at http://localhost:9977
ares-novacom -d tv --reverse --port 9977:18080

--port keeps its shape — device port first, host port second. Which end listens is what the direction flag decides. The device is asked to bind its own loopback rather than every interface, so a published port stays off the network the device sits on.

The accept loop

libssh holds the session lock for the whole timeout a poll is given, so an accept loop that asks for it again straight away starves the threads pumping the connections already open. The first test carried an HTTP request to the host and logged a 200 there, but the reply never went back. The loop glances for 10 ms and stands back for 40 ms instead.

bridge splits into the channel setup each direction needs and the pump they share. A session that has gone away usually leaves TryAgain behind as its last error, so the loop reports the loss without that meaningless suffix.

Verified on a UP7560

• content matches through a reverse forward
• a 5 MB transfer keeps its checksum
• three concurrent downloads all match
• 5 MB chained host → --forward → device → --reverse → host arrives intact in 5.2 s

Also works against stock OpenSSH, not only a device.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Gde2tcn5SbUHKhesuTNB7a