SSH exec over ControlMaster still costs ~250 ms; the network is ~10 ms
Reusing an SSH ControlMaster saves the TCP and auth handshake, but every ssh host cmd still opens a new channel, makes sshd fork a session (PAM), and runs the remote login shell's rc files. It also spawns a local ssh process, which is slow on a loaded Mac. Measured: ping RTT 5–13 ms, ssh host true over a live master 240–320 ms, bash -ic exit on the host 135 ms. So an app that runs one ssh exec per file op (list, stat, read, save, git status) pays about 300 ms each, regardless of the network. To get near RTT, keep one long-lived remote process per host speaking a small request/response protocol over its stdio (a fixed op set, no shell interpolation). Keep bulk transfers on a separate channel so they don't head-of-line block small requests.
Data point for the "where does the 250 ms go" part, measured on Linux with OpenSSH 10.0p1 on loopback, no PAM (portable build without it), login shell bash 5.2.21, no ~/.bashrc.
ssh host trueover a live ControlMaster took 2-4 ms, so opening a channel and forking the session costs little by itself. Adding a ~/.bashrc containingsleep 0.1raised every mux exec to 104-105 ms. A cold connect (ControlPath=none) took ~106 ms. So in your 240-320 ms, I'd suspect PAM session modules, rc files and local process spawn on the Mac, not the mux protocol. Checkingssh host 'true'with a stripped rc file would tell you which. The persistent-channel fix holds up here too: one long-livedsh -c 'while read ...'channel did 20 small ops in 13-22 ms.