The possible explanation put forth by aos and 10SNE1 is the most common reason a three-terminal regulator, but particularly the 317, gives up the ghost so spectacularly. However, a couple of clues suggest that another mechanism is more likely the culprit here: avalanche failure. I am basing this on two tidbits you reported: 1) that the amp was running when this occurred, apropos of nothing obvious; 2) that the input voltage to the regulator was measured at 33VDC.
The absolute maximum input voltage specified for most manufacturers' 317s is 37v. Believe me, they aren't kidding about it, either! Exceeding the maximum Vce of a bipolar transistor results in what is affectionately, and appropriately, known as avalanche. I'd be happy to explain how the avalanche breakdown process, but let's just leave it for now that it results in the transistor turning on much more rapidly than it would if the base-emitter junction were forward biased in the "proper" way. In fact, there are transistors optimized for avalanche operation - the most common application is the extremely fast pulsing of laser diodes - but the pass transistor in any manufacturer's 317 is assuredly not. What this extremely rapid turn on does is allow current to rise at a rate limited only by whatever stray inductance is in the way. Since good circuit design seeks to minimize stray inductance, the rise time can be very fast, indeed. Perhaps on the order of 5-10nS (that's nanoseconds, folks). Even a relatively small capacitor, especially a tantalum one, can dump a peak current of several thousand, that's right, thousand amps in so short a time. In this case, it is likely that a fast-rise time spike from the line got coupled through the transformer's interwinding capacitance (what Faraday, or electrostatic, shields are there to stop), sailed right through the ample junction capacitance of the slow-***** rectifiers meant for 50/60Hz operation, ignored the filter capacitors because of their high parasitic inductance and resistance, and triggered your poor 317 into avalanche.
Alas, poor 317. I recommend you put a small tantalum capacitor - say 6.8 to 10uF, max - right at the input of the next 317 so that it will present a low enough impedance to slow down fast rise-time spikes.
edited for two grammar mistakes
