No coin miner is ever standard. The main fun is that they invalidate all standard design flows: they either lead to failure or vast over-specification and under-performance.
As far as I know several people tried to use the internal self-clocking and reported problems with them: not always starting reliably and humongous jitter. My experience (not with mining, but with somewhat similar digital-but-nearly-analog chips) is that you are better off forcing your own external clocking and very frequently run short test mining jobs with known answers (like once every second feed the chip a job that requires just a few increments to find the golden nonce).
This is where I disagree with you the most. The failure modes of MOS devices are really strange, silicon has negative temperature coefficient and is prone to creating hot spots, where the device fails way below their normal rated specifications when the dI/dt or dV/dt was too fast. Hopefully you'll get your 55nm Bitfurries really cheap so you won't regret skimping on the safeties. But speccing and acknowledging the need for parallel transil is a good start.

My school had a program sponsored by International Rectifier (www.irf.com), we could burn power semiconductors for free (provided that we documented the conditions of the burn) and I still have some fond memories from that time.

I've seen your "back of the envelope" calculations and basically I suggest that you get a bigger envelope and make some more calculations. In particular, don't always assume that Vpeak/Vrms = 1.4, it varies widely with the load.

Or maybe just want want to go for the shock value and use one farad capacitors in your ripple filter, the ones that are used by the bass-heavy car audio enthusiasts.