But this way requires the script to be constantly polling for "triggers" (here am_type_subtype), which will require extra cycles (and hence cost) right? (Or would the extra cost be negligible?)
Indeed the normal thing for an AT to do *is* to poll for triggers, as transactions (which *are* its events) are exactly what drives it, so I don't really see that as "wasting cycles".
The idea of the hibernation is just to stop it polling completely until a given block height (as I think that could be reasonable for some use cases). There might also be a case for being able to send an AT NXT without letting it execute code (will have to think about this as you wouldn't want the AT to be easily starved).
It might also make sense to have an AT "priority" setting (a bit like task priority) so that lower priority ATs might only be run at most every X blocks.
Although obviously a lot of careful measuring will need to be performed at the testnet stage the design itself is aimed at being very minimal in "cost" (current use cases would require no more than 1K of memory and would only need to execute 5..10 instructions per "poll".