Hi Bassam,

This is actually the second report about this we have in the tracker ;)

We have a bit of a problem when it comes to reliably decomposing negative scale values - we can either say that there are negative values present, or no negative values present. The downside to this that we'd have to say that all values are negative, potentially causing some other problems.

For a while now, I've been starting to wonder whether we need a slightly different approach to passing transforms down the constraint stack. Namely, I may need to put in a system that ends up passing down "managed" transforms alongside the matrices so that we can resolve a good number of the ambiguities: in particular, euler rotation and scaling. It would end up complicating things a lot, especially as far as some constraints (e.g. track to's) where the exact transforms applied are not that simple to map back to individual components, but at least we'd finally be able to solve a few other limitations...

Aligorith, would love to chat with you (this is an entirely different idea with a different scope) about stackable transforms (an idea used by other software in different ways)
the idea is you could add a transform on the 'stack' (or possibly in a node tree) even after constraints.
for limit constraint this probably has no impact, but it could alleviate some of the other weird issues for constraints with offsets (e.g:
base transform
copy location
extra transform (used to be offset)

anyway, this one is not a show stopper, I should start using the search in our spanking new bugtracker (got too used to the old broken one, sorry)
cheers
Bassam