I guess the assumption was that seeking in the gossmap for the node would be more expensive than the signature check.

A few syscalls to read from an open file descriptor (especially when the requested pages are already hot in cache) constitute much less work than checking an ECDSA signature. lightning_gossipd was one of the most CPU-hungry processes on my system previously, and now it's barely on the radar.

EDIT: Actually, looking at the perf report, it looks as though gossipd mmaps the gossip store, so that's even better. Instead of a few syscalls to read from an open file descriptor, you only have potentially a few page faults to fetch pages of a mapped file in the worst case and only hot memory reads in the best case. Not even close to the same order of magnitude as ECDSA signature verification.

Note: gossipd really ought to be using asyncio to service gossmap queries since it's a single-threaded process that services a queue of requests by consulting the disk. Modern storage subsystems can support hundreds of overlapped requests, but gossipd presently is only ever submitting one at a time.