There have been no changes to AI for 105100, so any difference in behaviour is to do with track and signals.
Given that it changed on your same build, then clearly something changed on the route. Perhaps additional track was laid, signal changes, content updates (which could mean new scripts).
Something else to check is whether you're starting the "original" session with session rules not run/completed, or running a saved session where some of the rules have already been run. See
http://game.ts2009.com/mediaWiki/index.php/Unified_Driver_Surveyor
Given the nature of code, the AI does exactly the same every time in the same circumstances. That doesn't mean 100 tests will have identical results as "timing" could play a factor due to your PC running faster or slower (e.g. more/less background apps running) triggering a script a split second before or after the previous run. (That's a layman explanation for one possible difference, not saying it is the only variable).
Fwiw, usually when a loco gets "stuck" you can try Stop Schedule > Continue Schedule and that will often clear things. Another option is to Stop, manually intervene then Continue.
And one last thing - yes, we do need to improve the AI in cases such as when you describe, and to provide more feedback as to why it is doing what it is doing (which may appear to be "nothing" but is in fact "something").
Hope that helps.