Farss, isolating subsystems can be differcult with bloated OSs, but I think JohnnyRoy has built a fairly clean system for what one can commercially buy. The decoding of M2t stream, little bit of data = a lot of data, probably is forcing additional write cycles from the cache to main memory,(idling cpus demanding read new read cycles). Intel is using a delay-write-thru-read-demand cache design, giving read cycles high pr. ie. 10-1.

But I think your ideal about refining a series of tests for HD/HDV is a good ideal.

Jumping subject, Farss, have you looked at Cheroxy's veg? This is an interesting problem concerning 3d and frame size(I think). Veg runs in draft or preview settings, but goes bye-bye- at good or best settings. It looks like an internal problem, but knowing what set of parameters trigger this behavior could be useful.
I've change some of parameters defining images with a little success for an few extra frames, but no real solution. Maybe you could investigate this problem? Thanks in advance.