I'd agree simply because Windows presents a much larger attack surface. The amount of code running on a minimal Unix installation plus Tor is a lot less than a Windows system, especially network facing code.
... Running code, or network accessible code? Either way I don't see how you came to that calculation. 'Minimal' Unix + Tor + SSH restricted by SSH Key vs 'Minimal' Windows + Tor + RDP restricted by Client Certificate. I also don't know what you mean by 'minimal' as very few ... I think a Windows Server, properly configured, is roughly as secure as a properly configured Linux Server. ... I think there have been more bugs that result in RCE on production Linux servers running SSH and a webserver in the past 4 years than there have been in production Windows servers running RDP and IIS. ... I think if you're pointing fingers at China and the NSA, you should assume they have RCE in both Windows and Linux. ... I think running relays on Windows Servers is no worse than running relays on Linux Servers, and therefore it is good to do, because it adds diversity to the network.
Attack surface on a well adminned relay comes down to three things: - Network stack itself (kernel) - Daemon software itself (tor + remote admin) - Their respective use of other kernel/library/shell provided resources.
I might suggest the current proportion of Windows to Linux is roughly ideal. This is primarily because, all other things set equal at 'minimal' (= tor + remote admin), good adminning, and good control of corporate secrets (or moles)... Windows still has one huge strategic weakness at that point... the magic packet. It's the whole binary vs. opensource argument. So essentially, the correct ratio of the two might be the odds you place that a binary OS has a magic packet. Today's node count shows 73% to opensource platforms. I'd suspect 73% is about where a lot of analysts might bet on Windows being magical, whether by/for the NSA, or any other reason or source.
(Remember this... https://en.wikipedia.org/wiki/NSAKEY That was just from running 'strings'. Good luck trolling all of Windows with a disassembler... a nice fat payoff if you do. And the number of disassembling vs. opensource auditors is probably even more heavily skewed. And Windows is 'trusted' by buyers, nor can you replicate their binaries from any 'source code sharing agreements'. Then it's Patch Tuesday again... so it could be no one has or ever will disassemble audit it. So odds end up being pitched instead. And for many applications, that's good enough.)
The real problem below is the 96% allocation of opensource to Linux and 4% to Other opensource. That's something that should be fixed. For these purposes, it doesn't matter which BSD/Other you pick... once you get the security odds there back towards say 50:50 Linux:Other, then you can debate userland and relative security amongst them all you want.
Here's some links to get you started, including two other main branches of the Unix Kernel family tree at the bottom...
5939 Linux 1591 Windows 173 FreeBSD http://www.freebsd.org/ 56 Darwin 44 OpenBSD http://www.openbsd.org/ 7 NetBSD http://netbsd.org/ 6 SunOS 4 Bitrig https://www.bitrig.org/ 2 GNU/kFreeBSD https://www.debian.org/ports/kfreebsd-gnu/ 2 DragonFly http://www.dragonflybsd.org/ 0 Illumos (OpenSolaris) http://wiki.illumos.org/display/illumos/Distributions 0 Minix http://www.minix3.org/
Official metrics... https://metrics.torproject.org/network.html
Someone should really do an analysis of platform vs. exit bandwidth as well. Anyone?
Also, isn't there some project out there that is counting the historical number of kernel bugs+severity per OS over time?
[To cpunks to cover all the other volunteer node based networks out there that could benefit from tuning their platform ratios.]