Ah great, that helps. I had forgotten about the parallel execution. Even if there was some possibility of ordering, it's far less relevant than the dependencies being correct so a task is not executed before a dependency (ad infinitum).
I honestly don't have an effin clue how this thing works, even after buying Josh Suereth's promising start on "SBT In Action", reading what I can of Mark's pretty excellent xsbt docs and wading through what I could stay awake for of the official docs. IIRC, dumping the tree provides a lot of opaque object references for objects that don't have toString serializers. So super relevant stuff like filenames and other important parameters are lost. Then there are the plugins that don't output anything as far as what they are doing. They just work or they don't. And I can't figure out how to get IntelliJ to do the right thing and load source jars (though it doesn't seem to matter as a lot of releases are optimized and without symbolics).
Maven also has this behavior (due to it's declarative nature and early XML parsers having no guarantee of deserialized object order) and it has caused marginally violent consternation for many an Ant lover. I consider myself to have a relatively open mind and am not working against myself in learning SBT, but the multitude of operators and patterns approaches overload, much less understanding the rest of it and getting it to work. What worked for Maven is the augmentation of the "-X" option that prints out the input parameters to every plugin execution in human-readable form. This is probably going to be harder since there's no parameter encapsulation like there is in Maven (a SBT plugin has access to all the property trees), but what other option is there? It's like there needs to be a Swing-based debugger that takes inspect-tree to level of a graphically browsable log archive (not holding my breath as I wouldn't sign up to develop it either).
As it turns out, one of my colleagues found the problem I originally posted yesterday, it was with sbt-osgi silently failing with no debug output or error when it didn't like it's input. We spent something like ten man-days trying to find that. If SBT is the primary ambassador and gateway to the Scala ecosystem, it's seems really inexcusable.
Again, this isn't to denigrate anyone, I'm just observing my/our experiences, not pointing fingers at really smart people that have invested tens of thousands of hours creating this otherwise interesting tool.