Indeed when converting to radical or when explicitly setting Normalizer, it works fine for the exact case.
I thought Normalizer was set to normal by default, as it says in the documentation:
Normalizer is set to `normal` initially, and is used by the kernel in series to normalize leading terms of divisors.
Also, why would your suggestion of TestZero using simplify be 'safer' than the default?
Testzero is initialized with proc() evalb(Normalizer(args) = 0) end proc, which is a relatively trivial test for zero.
Furthermore, neither of these two have any affect on the floating-point computation, which still says the condition number is 68.06563146 instead of 120 (.something). I still don't see where this comes from, and this makes me feel totally unconfortable with relying on Maple here, especially since other software (e.g. Matlab) does not have a problem with this (and neither does linalg[cond])...
Finally, about your remark on the "hidden zero" being used as a pivot, such things should be avoided, is it not? Or is this the responsibility of the user? Try explaining this to a user who just wants to compute a condition number...
In the LinearAlgebra package, the only information about TestZero and Normalizer is in the documentation of IsSimilar:
Assignment to Normalizer of a stronger normalization routine can provide better zero recognition during this calculation than is provided by the default normal routine. Such an assignment to Normalizer or Testzero may be necessary for input which is not of type ratpoly.
I can't say this explanation helps a lot, instead it only confuses the users even more.
Thanks for your reply,