I have a strange feeling they're not exactly at 50°C...

Anything above 80°C is going to impact the life span of those cards.
What is the "lowest setting"??

Suggestions...
(1) Increase case air flow.
(2) Undervolt those cards as hard as you are able to. This will prolong their life, increase power efficiency, and decrease heat.
     Undervolting is best done one small step at a time, say 0.020 V. If the card is still stable a few hours later, repeat until you've found minimal stable voltage. It will likely differ from card to card.
     I tested each undervolt level for a whole day before going further though it may have been going overboard.
     On my 6770s I've been able to decrease power by about 0.090 (from stock 1.2 downto ±1.110) while keeping the high OC levels (995/300, stock was 800). That shaved off 4 to 6°C per card.
  

EDIT::Today is your lucky day. Phoenix 1.7.4 has just been released like half hour ago.
EDIT::Grab it and see whether those nasty stales have gone away.

Phoenix is currently flawed. Due to some glitch in the code it doesn't work correctly with low hash rates.
Nothing can be done until the Jedi95 locates and fixes that bug.

My previous suggestions were aimed at helping you to improve your hash rate, hopefully to the sane level.
I have no experience on running miners on Nvidia cards, hence the suggestions to try VECTORS and BFI_INT.
If you're thinking about some serious mining your best option is to sell that card and replace it with an AMD card.
35 MHash/s is almost negligible. An inexpensive hd6770 is capable of pulling 231 MHash/s.
A single overclocked hd6950 will do better than 360MHash/s - that's an order of magnitude faster.

That subject has been discussed to death so many times it's not even funny.
The search function doesn't hurt.

If not for pool hoppers, there should be no difference in the long run.
However, on traditional prop pools the hoppers will take advantage of your steady work thus lowering your profitability.
PPS, SMPPS, DGM, and PPLNS payout schemes are immune to hopping.
Find a pool with low fees and preferably low invalid/reject count.

Give BitMinter or Eligius a try, they are both 0% fee.

One of the databases⁽¹⁾ (either the block chain or the wallet) got corrupt.

Move the block chain data (addr.dat, blk0001.dat, blkindex.dat) into a temporary sub-directory so bitcoin-qt can't find them.
If any of those files were damaged, bitcoin-qt will now launch and start downloading the whole block chain from scratch.
If bitcoin-qt still fails, put the block chain files where they belong and deploy your backed-up wallet.

Notes:
(1) Actually, I should have said "At least one" but an event resulting in damage both to the block chain data and wallet is very unlikely.

Neither VECTORS nor BFI_INT worked with your nvidia card? Bleh...
Did you per chance try changing the kernel to phatk or phatk2? (-k phatk / -k phatk2)?
As to phoenix being admittedly flawed, your only chance would be going with another miner.
If phatk does work for you I suggest you go with cgminer.

Nvidia? Huh...

One of these should work:
phoenix -k poclbm VECTORS BFI_INT PLATFORM=0 WORKSIZE=128  AGGRESSION=9 -u http://user:password@bitcoinpool.com:8334
phoenix -k poclbm VECTORS PLATFORM=0 WORKSIZE=128 AGGRESSION=9 -u http://user:password@bitcoinpool.com:8334
phoenix -k poclbm PLATFORM=0 WORKSIZE=128 AGGRESSION=9 -u http://user:password@bitcoinpool.com:8334

Bullshit.

Yes, you are most efficient at 50% but by no more than 2 or 3% it your PSU is 80+certified.
Any decent 1200-watter will pull 950W at the wall indefinitely.

As to leaving the memory at 1375, God help you. Even with the hardware-limited "core_clk -125MHz" cards you'd easily drop ±20 W per card. Less power, less heat, less PSU load.

Right on, P4.  I'm glad you raised the often overlooked point of diminishing returns at high fan speeds.
Just keep in mind that "below 70°C" is an extreme oversimplification:
   The 69xx series can never be expected to run that cold. They are all flawed products, having been meant to be manufactured at 32nm instead of 40, and will run more like high 70s - low 80s.

@BCMan, now you're talking dude Welcome to the land of safe and efficient mining.

Find max stable overclock at stock voltage and stick to it.
Overvolting will hurt your power efficiency as the power usage rises by core voltage squared.

Beg your pardon but I was under an impression that Phoenix stable version was 1.7.3.

If your miner actually is 1.7.4 I suggest you switch to 1.7.3.
If you're pulling some crazy overclocking speeds clock your cards down for a while and see if that helps.

You are quite mistaken. That switch there be fan control (Silent = higher temps). Switch it to Performance and rejoice much lower temps.
Just make sure you power off the rig. It's all too easy to make a costly boo-boo when messing around with a live machine.

The MSI R6970 Lightning card is clocked at 940 MHz, it's already factory overclocked.

Linux 2.6.32-5-686 #1 SMP Mon Oct 3 04:15:24 UTC 2011 i686 GNU/Linux
AMD driver ver. 11.11 && AMD APP ver. 2.5

just an example:

EDIT::Oh, since it isn't obvious from the uname output: it's Debian.

Way to go BCMan!
Give that card a day. If it's still perfectly stable tomorrow, undervolt again.
If you manage to repeat this 3 or 4 times the card should be about 5 °C colder taking care of the thermal issue you raised earlier.

Sorry to hear this :<
It's a total crapshoot, some GPUs have just higher quality chips than others.

If you do have a secure backup of the wallet file why don't you research that issue?

Try re-installing 5.2 again. You have nothing to lose but a bit of time.
I strongly believe that your "bug" amounts to a typo, caps lock being turned on or some other trivial matter.

Doing backups of an unencrypted wallet is an insanely wrong idea.
The only worse thing that comes to mind is never bothering with any kind of backups at all.

It's critically important that you only download the client from a trusted source, like the Bitcoin homepage. It will redirect you to sourceforge for the actual download.
Never download from random sites, rapidshare, or bit-torrent as it well could be booby-trapped.

How can anyone tell you whether your particular card has a good enough chip to do that?
You're going to need to test that.

Try dropping the voltage by 0.020 V and running the rig for a day. Repeat until you've found maximal stable undervolt.
Now, there are two choices: you can drop the clock slightly and keep undervolting or you can keep the stable voltage.
If you're actually paying for your energy, dropping the core clock and a serious underclock job are usually the way to go.
Also, undervolting significantly drops core temps.

Pruning is only done on end client's side!
If you downloaded a pruned chain, that would amount to trusting other nodes that the pruned-out transactions were all legit.
There is no place for this kind of trust on the protocol level.
You still download the whole block chain, verify each and every block, and then - trusting only yourself - prune out some of the transactions.

Just do mind the core temperature k?
You don't want to fry those cards.

Obviously, once you got mining going now is the time for overclocking and fine-tuning.

Read those to get yourself up to speed:
GUI Miner thread
Hardware subforum

When in doubt, use the Search function && RTFM.
Report specific issues here.

Once you get familiarized with mining I suggest you switch to cgminer which is the most advanced miner currently available.
CGminer thread

If your machine is going to be used as a dedicated mining rig:
(1) Never use crossfire. Make sure the crossfire connector is removed.
(2) Don't overclock that CPU. You aren't going to need it. You might also opt to disable half of the cores - that will increase your power efficiency.
(3) There's no good explanation to stick with Windows.

For me it's 0/7 GPUs. All in fine working order, the oldest of them running since summer. Undervolting is crucial IMO.

They've not been idling, either. The ASUS hd6950 DCII 1GB (which I detest for its non-standard VRM, BTW) has been running since early September at 942 MHz. This is the max stable setting for this particular card using any recent AMD driver and 2.5 SDK.  At 945 it will lock up in minutes.

Scoring a perfect 10/10 failured cards, surely you must have done something wrong?  
May I ask what temps and fan speeds you've been running those cards at?