As for an explanation of the oscillation:
Notice the internal comparator feeding the output NPN structure. (Which is not actually a single NPN, it's more complicated than that; but calling it an NPN suffices here.) It must source base current to turn it on. With emitter basically unconnected (pulled high), that base current flows through the B-C junction instead -- to the output.
This has two unexpected effects:
- The output sources current. It's not strictly open-collector (sinking only).
- The comparator's phase is reversed. When the open-collector output is supposed to be pulling low, in fact it's pushing high instead.
This is compounded by the anomalously large collector load resistors (~100k): the base current is small, but across such large resistances, it has significant effect. Evidently, it's nearly to the supply rail(!).
Finally, oscillation frequency is, in part, determined by the input capacitance, which combined with the large resistor values, gives a time constant of, apparently fractional microseconds or so. Internal propagation delay accounts for the remainder.
(In general, comparators are not stable with negative feedback, as they have too much gain and phase shift to stabilize, as an op-amp would. They are otherwise very similar components [comparators and op-amps], indeed some differed by only one component: the omission of a compensation capacitor.)
This explanation however leaves one wrinkle, that I am not too familiar with: the SPICE model.
Models -- as the name suggests -- are only ever an approximation of the real device. Manufacturers will never (or, in many cases, contractually cannot) release models disclosing their actual as-fabricated devices; that's internal IP, trade secrets, or licensed from the fab. So, they release simplified approximations of them. They might only model the datasheet parameters (or not even all of them), and might not include parasitic effects (for example, the input phase reversal effect, that early op-amps were infamous for, might be such an effect that gets omitted in a model).
I don't know, offhand, to what accuracy the library LTSpice LM311 model is written. It could very well be constructed with a (truly) floating transistor, or even a completely virtual switch element, instead of the representative output structure.
It does seem evident, or likely at least, that they modeled these effects accurately -- accurate enough that the above explanation seems reasonable. Despite that this internal base current is, well, internal -- arguably a parasitic effect!