From the comments on 2893, I gather the reason for this is that vaapidecodebin appends a queue internally, and we don't want that.
I'm not thrilled about this, because this is fundamentally a problem of us wanting to control exactly how samples are buffered, and there's no guarantees on the GStreamer side. I'd rather be able to just force GStreamer, or some other library, to provide those guarantees to us. But in the absence of that, I'm willing to accept a patch that blacklists vaapidecodebin.
However, it will need a better comment than this. "Too much abstraction" is vague if not meaningless. "It introduces an internal queue object, whereas Windows programs depend on samples being available synchronously" actually explains the problem and why we can't easily work around it, for instance.