I'm currently using hypernova as my primary pool, because I like the CPPSRB method and have so far been happy with the pool performance. Now I visited this topic because I was curious about the shrinking pool reserve, but was surprised to see that the pool had silently switched to PPS. There is no information about this on the website, apart from some small changes on the front page. The help page still says CPPSRB and 1% fee though. Personally, I would like to see the pool returning to CPPSRB.

There are some concerns I have about the pool being PPS at this point; primarily with regard to the reserve. You state that the reserve is currently held entirely by the miners. That is not a proper pool reserve, as many miners will likely start withdrawing if the reserve drops outside of their comfort zone, thus accelerating the depletion of the reserve. In other words, the pool should have its own reserve, on top of what's in the miners' balances.

Given the small pool reserve, I don't think PPS is currently a viable option for this pool. You say that you have 3300 LTC, of which 10% "has never been withdrawn". If we calculate the probability of pool bankruptcy using the "never withdrawn" (300 LTC) balance as reserve, and an effective fee of 3% (since you say 1% goes to server maintenance fees) that gives a 70% risk of eventual bankruptcy. Using the full 3300 LTC as a reserve instead puts the risk at a more acceptable 2%. To limit the risk of bankruptcy to 1% a pool reserve of at least 3800 LTC is required with the fee at 3%.

tl;dr: The switch to PPS was premature and there is currently a very large (70%) risk of pool bankruptcy due to insufficient pool reserve.

PS: Calculations were based on equations found in this paper: https://bitcoil.co.il/pool_analysis.pdf
Example calculations:
http://www.wolframalpha.com/input/?i=exp%28-2*f*R%2FB%29%2C+R%3D300%2C+f%3D0.03%2C+B%3D50
http://www.wolframalpha.com/input/?i=%28B*ln%281%2Fdelta%29%29%2F%282*f%29%2C+B%3D50%2C+delta%3D0.01%2Cf%3D0.03

I had to pay transactions fees when I transferred some XPM (using primecoind). The first time it was because the coins were too new (0.01 XPM fee) and the second time because the transaction was big (200 XPM, 0.02 XPM fee).

I also notice a decline in PPS using Chemisist's code compare to the official (0.11). Getting an average of about 1800 PPS vs 2200 with official. I'm only running 12 threads as well, so it's not just affecting users with high thread counts.

I spent a long time modifying the makefile for Intel Composer XE and finally got it to compile. The result is rather disappointing however - no significant increase in PPS. I used the -fast option which is supposed to give rather good optimizations, but I'm guessing there are more optimizations to be made. I will keep looking for those.

Memory is probably the issue indeed. I think it's the -j6 part that should be removed though (-j6 means it will compile 6 files at a time if I understand it correctly). Each compiler thread uses about 300 MB, so 6 will use around 2 GB.

Supposedly, yes. However, I tried without and I'm not seeing much difference in the PPS. It is difficult to measure though, so the difference could still be there but it's probably small.

Something like that. More specifically:

This nearly tripled my PPS from the original release.

Now getting 2000-2400 PPS on an i7-980X @ 4 GHz. 7 blocks found so far (of which 1 was orphaned).

Compiled with -O3 -march=native.

The question is if that's small enough. With a maximum of 21 MBTC, the smallest unit corresponds to 1/5e16 of that. As a comparison, there's roughly 10 trillion USD in existence, where the smallest (physical) denominator is 0.01 USD, corresponding to 1e-15 of the total. Thus, bitcoin's "advantage" here is minimal.