Thanks for the reply!
I didn't actually remove the switch from the stack for my 2nd test - I carved out a new interface & iscsi target on the array and connected it via a single link to to one of the member switches. Then I literally plugged my laptop into another port on the same switch and ran IOmeter from that. Switches are stacked using their 10Gb interfaces via DAC cables.
Path selection is round robin, iscsi bound vmks etc etc. I have another test I ran to this same array via a host that again, lives on the same switch as the array. The results were similar (better actually, as my ESXi guests are using multipathing and whatnot). It seems to be only traffic that traverses the core stack that get screwed up, yet the stack ports themselves show no errors / drops.
I wonder if maybe we've hit a firmware bug of some sort. We're running a fairly old firmware revision, in fact the one that first supported ethernet port stacking, 4.2.2.3. Problem is the upgrade process would be site-wide downtime - so it'd be nice to know it's not a config issue of some sort.