It occurred to me that Andy Ruhl wrote in gmane.os.netbsd.general:
> Yeah but all bets are off if you have to sit next to your access point
> for the thing to actually work. There's only so much the protocol can
> do if the hardware doesn't cooperate.

The hardware was cooperating enough; just at a low speed. But NFS dying
on me, even when using nfs-over-tcp, indicates a more fundamental problem.
Lowering the blocksizes helped, which indicates that there's something
fishy with large packets.