I'd drop the hardware SPI "machine" and bit-bang those few bytes "manually". This would exclude any other hardware or logical error

I know this is an Atmel chip, but even then, you have to assume that (most of) the chip works. Of course, re Atmel if the SPI does not work, it would not be in the errata for a year or two.

Since the SPI engine in these chips have been around for a while I would be surprised if there was an error in it and nothing had been heard of it before

the suggestion "bit-bang those few bytes "manually"" ignores the likelyhood that the OP is 'experimeting' with SPI to implement it for more than "those few bytes".

Using the HW SPI has oodles and bunches of advantages over using bit-bang.

To the OP:
Have a quick look at the SPI in the Philips LPC chips, it would be very likely it is the same. If so, use CodeArchitect, free form http://www.esacademy.com, to generate WORKING code and develop your own style of code based on that.

Erik