Voltage has a far bigger impact on the overall lifespan of a CPU rather than heat. With increased voltages come increased heat, but heat will be the last of those two to degrade a chip. CPU's these days are fairly robust and can in fact handle temperatures out with the recommended operating environment, although and its best practice to keep them within that recommended range they can still operate for a long period under such conditions.
Voltage is a completely different matter as it is a lot more sensitive. Exposing the chip to voltages higher than that which the chip maker strongly recommends for 24/7 use can have a crucial effect on the lifespan of that chip over time. Running high amounts of voltage (Vcore) through a CPU for a long period of time can end up leading to electromigration, ultimately resulting in the chip being unstable at it's set voltage and requiring more for the exact same frequency.
The Intel Whitepaper states that the absolute maximum amount of Vcore for 32nm Sandy Bridge based CPU's is 1.52v, but exposing the chip to this amount in a daily usage scenario could possibly cause malfunction and therefore ins't recommended that you do so. Having LLC enabled whilst running at such levels is also not recommended, as this breaks Intel's chip design specification, not to mention that it can lead to problems too.
Personally, I would try to stick to <1.45v for 24/7 use, the lower the better. You'll also obviously want to make sure that temperatures are within a decent range whilst stability testing.
For those saying that LinX finds errors quicker than Prime95, please understand that these two applications use completely different CPU testing methods. Just because LinX will stess the CPU far more and find stability errors quicker doesn't necessarily mean that it's better. Both applications have their advantages and should be used whilst stability testing an overclock, including being used properly. That doesn't mean a mere 20 passes of LinX or a 2 hour run with Prime95.