Language intricacies aside, is there a reason to use Clojure over Elixir, Erlang? Genuinely curious what JVM has to offer vs BEAM / OTP if you're going to use dynamic languages.
Sadly no, but I don't think there's much of a ping difference on a good network. With WebRTC (datachannels in UDP mode) you just get the benefits of UDP (or lack thereof :)).
Only played it with my friends, don't think a map ever exceeded 10 players. So I can't say much about scaling - although it was a client-server model over UDP and not peer to peer networking.
Quake's source code is very hackable. I forked ioquake3 for a personal project and it was surprisingly easy to add WebRTC support to the C codebase, so you could play it in your browser over WebRTC :)
You can also prerender the tiles on a powerful machine and host them on a cheap instance (assuming you don't need very large regions). A while ago I created an OSM server (https://github.com/seemk/TileSweep) to do this for me, you can select regions you want to render via a polygon and you get a SQLite database of the tiles. It can also run as an alternative for mod_tile + renderd.
I rendered all of Estonia from zoom levels 0 to 18 (~16.5M tiles) which is around 20GB and host the tiles on a cheap DigitalOcean instance.
Totally agree with your point on HOG + SVM, I think it is obsoleted by convolutional neural networks.
I wrote a realtime human detection library [1] for a robotics project that used HOG + a simple neural net for classification. While it worked okay, I wasn't happy with the precision (around 90%) and decided to try out a simple convnet from Torch (doing the classication on depth images instead of HOG descriptors). The Torch version was slightly slower on a CPU, but both the precision and recall jumped up drastically.