CCing Cherie since it's tangentially related (through a recent discussion) to the work she's doing with making HTTP/2 push more cache aware.  

Cherie: This isn't directly relevant to what you're doing, but it's part of the whole preload/push/no-store caching question we discussed last week, and I don't know of a narrower place where no-store caching issues are being discussed, so I'm casting a wide net.  Feel free to ignore.

We should probably take the design discussion to net-dev@, but yes, the description of the current behaviour sounds correct, and doing #2 (drop the request) also sounds correct.

The broader question is whether we should be using flags like LOAD_PREFETCH to signal the "cache but don't use" behaviour. It's seems quite weird to have the //net layer indirectly, tangentially aware of the "pull into cache" semantics of prefetch, while still requiring the full consumer interface (of a URLRequestContext).

+1 to that....the LOAD_PREFETCH stuff in net/ seems pretty weird to me, too.  And is there some sort of arbitrary freshness timer in net/ as well, to prevent really stale resources via that path?

Yes, #2 seems strictly better than the current state.

However a way to do #1 might also be interesting for NoStatePrefetch, although I don't really have clear metrics about how important this is for now.
I've also seen it mentioned in the context of improving back navigations.