Fourth way

I don't think the OpenNTPD algorithm is of sufficient quality to driver stratum 1 or even stratum 2 servers. OpenNTPD isn't really suitable for scientific and similar workloads. It also doesn't support NTPv4, but only SNTPv4 and NTPv3. Their priorities may have evolved, but when OpenNTPD first came out they were fairly explicit about the use cases they were targeting--principally networking equipment--and that their primary focus was on simplicity of implementation.

Also there's the issue with support for reference clocks. See my comment elsethread regarding chrony.

Other people know far better than me, but years ago I asked myself the same questions as you and these are the answers I came up with. If the slate could be wiped clean and the entire global infrastructure magically transported into the year 2017, with 2017 hardware and 2017 software, ntpd likely wouldn't figure into the equation at all. AFAIU, the core algorithms in ntpd have been hashed and rehashed in countless scientific papers and other discourse, so there's nothing special per se about the code that doesn't exist elsewhere or can't be easily rewritten from scratch. And the bulk of the remainder is effectively legacy support. But if you want those algorithms (because they're tried-and-true, despite the downsides) and need the legacy support, it's not crazy to want to keep ntpd around.