Is it possible to construct a tx such that bitcoin nodes believe total signature operation costs is low, but in actuality nodes end up wasting CPU cycles and fail to catch up with the tip?
Software doesn't think or believe. What you're asking is: is it possible to construct a transaction that has an operation cost unexpectedly higher than it should? Right?
Very interesting. There is a thread about the transaction that takes 3 minutes to validate by Sergio_Demian_Lerner (according to his hardware's specs at that time) here:
https://bitcointalk.org/index.php?topic=140078.0. Following the old, non-segwit-upgraded rules, a 1MB transaction can hold 10,000 signatures, given that transaction inputs require at least 100 bytes to hold a signature. Provided that validation of a signature takes about 2 milliseconds, our transaction would require 20 seconds to validate.
I presume the way the network defenses this attack is by charging extraordinarily high for 1MB transactions?