Split each channel's backlog scan into a find phase (settles the
found count first) and a react phase, so the progress display no
longer grows both numbers in lockstep. Found messages are logged in
a dedicated, message_id-deduped table and recovered across restarts
independent of the resume checkpoint, so nothing found is silently
abandoned if the script stops mid-react.
Also drops the separately-maintained messages_found counter, which
proved prone to drift under repeated interruptions/rescans, in favor
of computing it live from the dedup table - self-healing regardless
of how many times a channel gets rescanned.
