Thanks 8cup, much appreciated!
I've just tagged up v0.10-pre, which improves auto-detection of tokenised BASIC programs. This should mean it's hardly ever necessary to use the --input-tokenised option now, even if the program you're trying to work with has machine code attached after the end of the BASIC. Thanks to Tom Seddon for the bug report and auto-detection algorithm (which is taken from BeebLink, and apparently traces its roots back to Matt Godbolt's BBCBasicToText.py, so thanks to Matt too...)
Testing/feedback welcome, if no one reports problems with this after a few days or so I'll tag this up as 0.10 proper.
Edit: Just to be clear, this doesn't fix any of the issues lurkio reported. They're still on the list of things to look at...