I'm not surprised that a look-behing assertion takes longer than something like (?:\A|\n), especially by this small of a margin. What wouldn't make sense to me would be that /^.../m would be 15x slower than either of the regexes I proposed unless I was missing some feature of ^ that isn't captured by my proposals.

Perhaps ^ needs to be re-implemented as using the regex operations used to implement either of my proposals (since look-behind assertions weren't available when the ^ code was written and so it might use something much worse and your testing seems to show that it does).