Large parallel uploads (e.g. ClaudeSync dispatching many files at once)
could trigger the server's sshd MaxStartups limit, dropping connections with
'kex_exchange_identification: Connection reset' or 'Connection closed'.
These are transient but the batch worker never retried them.

- Classify kex_exchange_identification / connection closed as retryable
  (ConnectionLostError), with a defensive string match in is_retryable.
- parallel_operation now retries each file on transient errors with
  exponential backoff + jitter, reconnecting between attempts. The jitter
  desyncs workers so they don't all re-handshake at once.
- Show '(after N attempts)' on recovered uploads; document the cause and the
  parallel_connections knob in the config template.