We spawn the daemon detached with its stdio discarded, so a fatal
startup error was just an exit code. A denied listen() even exited 0,
since the server error handler routed through shutdown().
This commit adds a reportFatalError() that writes the reason to a
<sessionId>.error file, appends it to the log, and exits 3, plus a
diagnostics.ts that turns the errnos behind these failures into
explanations. The following commits are going to read this file.
Also it fixes the log stream being created before the session directory
exists.

Previously we only output "Daemon process exited unexpectedly during
startup" on daemon startup failure. Now we report the reason the daemon
records, which is the startup error file first, then the tail of its
log.
When neither exists, report what is known instead of guessing. If the
daemon exited, print the command to rerun the spawn in the foreground.
If it is still running and simply has not published its metadata, say
that and name the pid, rather than declaring it dead and blaming the
sandbox.

An unusable session directory is the one failure the daemon cannot
report, since the error file it would report through lives in that
directory. Check it in the client instead: stat, mkdir, and a write
probe, because mkdirSync() succeeds on an existing directory that cannot
be written to.
Check the socket path length there too. The kernel rejects an over-long
sockaddr_un with a bare EINVAL that names neither the limit nor the
path.
Also stop "session list" from creating the session directory just to find no
sessions in it.

A dead daemon and a sandbox that forbids connect() look the same to
validateSession(), and both ended in cleanupSession(). Deleting the socket file
of a live daemon makes it unreachable for good, and deleting the metadata hides
it from "session list" and "stop --all", leaving a process holding a profile that
no command can find or stop.
This patch gates cleanup on the errnos that prove nothing is listening:
ENOENT, ECONNREFUSED, and ENOTSOCK. On EACCES, EPERM or ETIMEDOUT, keep
the files and report the session as unreachable with its pid.