Given many payments are only a few hops, I assume most of them, honestly.
I assume that in many payments an attacker can see network messages for much of the path, or at a minimum the first hop and last hop of a path, which suffices to figure out a payment based on timing (assuming we don’t have any delays and only a moderate flow of payments along the path, which is probably common-ish, at least today, but if you only add one or two intermediate hops I assume its still pretty doable).
My point was that being an on-path attacker is pretty similar in principle to a network attacker who can see only a subset of the hops, but with additional information, that lets you remove some false-positives in the classifier. Obviously pre-PTLCs you just know from the payment hash, but post-PTLCs the amount is pretty valuable, just not perfect - at that point we really want some delays so that the amount stays a bad classifier rather than a perfect one when combined with time.