Recalculated CA ratings

I think the opposite is true. 4 kills is worth just as much as 100dmg, so waiting until the end of the round and then making kills won't get you much damage and when you play 1 vs 3, it's likely that you get killed before you even make any significant damage at all.

You're not complaining about the elo system, you're complaining about typical clan arena and the same ol same ol that happens every game EVERY time there is an elo whore on the server. Yes, they gain some elo hiding but they don't exactly gain a ton either soley by hiding. This is their way of saying ''My team can't hold a position so I am going to make the best of it by camping and railing''. You gain a ton of elo playing good, that's it. You're good, you frag and gain elo.

Seriously, if you hate +back so much just play a different gamemode or a different game altogether.

First players who don't meet minimum match time/round requirements are removed from the player list.
For the remaining players a performance value is calculated, then the values of all players are compared to sort the list. You basically "beat" everyone below you in that list and lost to everyone above you in the list. A multiplayer match is treated like n*(n-1)/2 duels. For this calculation teams are completely irrelevant.

I calculate a "perf(ormance)" value for each player in the match. "beat" means a lower perf value than the player you are currently being compared to. If there are 10 players on the server, you are compared 9 times. It only cares about higher / lower / draw (very small margin), not about the absolute "perf" value.
The amount of points gained or lost depends on the outcome (win/loss/draw), your rating before the match, everyone else's rating before the match, everyone elses RD before the match and your own RD.
Low RD values (like 30) mean that the system is confident about the current rating and will only adjust it slightly, even for unexpected outcomes. This aspect of glicko is actually something that worries me. Once your RD is low, you get stuck around your rating and won't gain much by beating severely better players.
If you get beat by players with a higher rating, it also tends to keep your loss small.
But there is no simple rule-of-thumb. The Glicko update formula is quite complicated and with so many input factors it's hard to generalize things. Too many dependencies between the various factors.

When you only play on one server where most people are better than you, you will have a low rating.
If someone else plays only on another server, where he beats most players, his rating will be higher. As long as the 2 communities stay on their servers, those ratings are fine for shuffling on those servers. But the ratings are not comparable to eachother. Only when everyone plays everyone else, the ratings are comparable.
That's why it is possible for mediocre players to have a high rating or good players to have a low rating. They don't provide enough input to the system to be properly rating. The more players you play, the more accurate the rating is. Playing the same 10 player 100000 times won't do any good.

KDR doesn't matter at all. What matters is the Perf(ormance) value. Yours is compared to everyone else's and if it's smaller, you lose against that particular player and if it's larger than you win against that particular player - and there is also a small margin for a draw. Only win/loss/draw matters for Glicko, not the actual Perf value.

You don't get completely stuck. I keep a minimum RD of 30 atm, which still leaves room for reasonable changes.
The problem is that the initial RD of 350 allows changes that are way too large. When I recalculate the data the next time, I'll most likely lower this or add some other measure to prevent severe overshooting (e.g. a 1500 player with RD 350 beating 5 players with 2200 could gain 1200 in a single match)

you need >= 50% of rounds played in a match to get rated.
the "perf" value is scaled up from your played rounds to the total rounds, so on paper the number of rounds you played should not matter.

people will change behavior as soon as they know they are being observed - and that change is quite often based on wrong assumptions than on actual facts. even if i spent the rest of my lifetime tweaking this system, the amount of criticism would stay the same.
i wanted to create a system for shuffling teams better than QL's 50/50 random generator. for that the ratings are good enough. beyond that, i don't care much.