I started a thread on perl5-porters about this, SvUTF8 predictability, and it also lead to a fair bit of discussion. However I think this snippet from dave_the_m probably best sums up the likely consensus:
I don't think its reasonable to document perl's behaviour vis-a-vis UTF8 flag behaviour. It will vary between
releases, and it may well vary between different code paths (for example hypothetically rvalue and lvalue substr() might differ). It would also constrict any future bug fixes or optimisations.
.. with an assumption that mechanisms such as The 'unicode_strings' feature will continue to be added and refined to reduce as far as possible any need to know.