[56steve]

The tunnel manager keeps the last few KB of each ssh process's stderr. Once that buffer passed 8192 bytes it cut it down with drain(..len - 4096), a byte offset, so when the cut landed inside a multibyte character (any non-ASCII ssh message, a localized error or a hostname) String::drain panicked in the stderr reader thread.

I moved the append and trim into push_stderr_line and made it step back to a character boundary before draining, the same way push_capped in session/mod.rs already does. The new test pushes lines of two-byte characters past the cap repeatedly; it panicked before the change and now keeps the buffer under the cap. I also added a CHANGELOG entry.