> Each time a transaction modifies rows in a PG database, it increments the transaction ID counter.
It's a bit more subtle than that: each transaction that modifies, deletes or locks rows will update the txID counter. Row updates don't get their own txID assigned.
> It looks like the new vacuum "emergency mode" functionality starts vacuuming more aggressively when getting closer to the wraparound
When close to wraparound, the autovacuum daemon stops cleaning up the vacuumed tables' indexes, yes. That saves time and IO, at the cost of index and some table bloat, but both are generally preferred over a system-blocking wraparound vacuum.
It's a bit more subtle than that: each transaction that modifies, deletes or locks rows will update the txID counter. Row updates don't get their own txID assigned.
> It looks like the new vacuum "emergency mode" functionality starts vacuuming more aggressively when getting closer to the wraparound
When close to wraparound, the autovacuum daemon stops cleaning up the vacuumed tables' indexes, yes. That saves time and IO, at the cost of index and some table bloat, but both are generally preferred over a system-blocking wraparound vacuum.