This is a solution, not a problem.
I have just spent a couple of days chasing this bug; the debugger was of no use because the part starts perfectly when the JTAG is attached.
The problem occurs with both Debug and Release versions of the firmware.
This was caused by BCL12 in the Silicon Errata for the 2617. I was setting the DCO to the calibrated 8MHz values from the info memory soon after startup.
This causes the DCO to stop in some circumstances.
The solution, suggested by the Errata: clear DCOCTL, increment or decrement the Rsel bits in BCSCTL1 until they match the desired value, finally set DCOCTL to the info memory value.
The system starts perfectly every time now.
I hope that this saves somebody a lot of time.