I'd suggest that you go through the bug report database and the test suite
(and maybe old emails) looking for both limit and tlimit, and put together a
table whose columns are: problem, limit, tlimit. For each row, indicate
whether the result for (t)limit is (a) correct; (b) correct but not the best
form (e.g. some complicated complex-exponential expression which simplifies
to 0; (c) noun form. (d) blows up (internal error, infinite loop, etc.); (e)
incorrect. I think this is the order of preference, strangely enough: that
is, it is better for the system to blow up (so the user knows that it is not
giving a correct answer) than that it gives a incorrect answer -- otherwise,
we could just have Maxima return an arbitrary number (42 for cosmological
problems, 137 for physical problems, 666 for metaphysical problems) whenever
it catches an error.
I suspect that neither limit nor tlimit will dominate the comparison, and
the set of problems is certainly not a "fair sample" -- since limit is the
default, it is being tested much more heavily in actual use -- but at least
it will give us an idea of the landscape.
-s