Socket.read with no count, or with a huge one, filled the buffer until the
peer closed the stream or the timeout expired. A peer that never ends its
stream grew the buffer without limit, and a failed extend_from_slice aborts
the process, which breaks the never-crash rule.
One read now holds at most 64 MiB and reports an error past that, the count
is clamped to the same bound, and buffer growth uses try_reserve so an
allocation failure is returned to the script instead of aborting. Socket
paths that had reported every fill failure as a timeout now say what failed.