Something in 5e86746 is going haywire --- 4x longer to precompile Plots, +5% sysimg.

Ah, I see what happened. The new lazy-ambiguity logic assumes that subtyping is transitive over specificity, and I accidentally formed a contradiction of that premise on the definition of == and returned too many matches. In particular, I observe the set of facts that:

!(a <: b)
a <: c
c morespecific b # ie. c < b

where a == c == Tuple{Type,Type}
and b == Tuple{T,T} where T

And I wasn't checking for this case, so in the results, it appeared that b was also a possible method match, which complicated the caching logic (made it slow).