OP here. Thought I would followup now that I have the OSHPark boards in hand.
First, due to distributor stock issues, I ended up using a slightly different BGA part, one with 12 balls in a 4x3 array at 0.5mm pitch, rather than the original plan of a 9-ball BGA with 3x3 array at 0.5mm-pitch. The 12-ball BGA requires the two inner balls each to be escaped with a 75um/3mil trace that routes between two 0.5mm-spaced balls on the perimeter of the BGA.
In terms of the OSHPark boards, there's good news and bad news.
The good news is that both 3mil traces seem to have been successfully fabbed on all 3 copies of the board.
The bad news is that each 3mil trace is shorted to one or both of the two perimeter BGA balls as the trace passes between the perimeter balls. The short is present for all 6 traces (2 traces on each board, 3 boards).
Unfortunately I don't have a camera on my stereo microscope, so I'm not able to post a picture of the board. The shorts are reasonably apparent visually, and I confirmed them with a multimeter using testpoints on the board.
Oh well, it was worth a try. Out of sheer stubbornness/stupidity, I may try giving Aisler a try with the same board design just to see what happens.