I like the behavior because, as you mentioned, there's some interesting sonic territory, and there's a perfectly serviceable way to get a "traditional" sync.
With oscillator sync, the only thing that matters (in terms of wave reset) is the period of the controlling oscillator. You'll notice that the selected wavetable doesn't matter, and the current wave shape also doesn't matter. The only thing that matters is how often the wave starts a new cycle.
And the cycles on Oscillator 3 are weird, due to how they're stored in memory. At low frequencies, it's a single 1024-sample waveform, then there's a single 512-sample waveform, then two copies of a (downsampled) 256-sample waveform, and finally eight copies of a (downsampled) 128-sample waveform. My best guess is that Osc 2/3 sync behaves so differently because of these variations as Oscillator 3's pitch changes. A regular frequency "sweep" isn't really very regular. The eight-copy range comes in sooner than you'd expect (around F5), and so your Oscillator 2 reset starts happening a lot, causing the timbre to kind of "break up" like it does.
I'm not totally sure, but I suspect that there's some math going on behind the scenes that takes advantage of these downsampled copies by increasing the software's actual played wavelength. Higher frequencies are maybe an illusion and this gets reflected in sync.
The workaround is to keep Oscillator 3's pitch low. You're in the highest-quality 1024-sample section of the waveform up to around F2, so stay below that. Or, even better... embrace the instability.