Is there a write-up somewhere describing in detail the 'accidental selfish mining' problem that this mitigates? I think a link in the BIP to a fuller description of the problem and how validation-skipping makes it go away would be helpful.

RE: which bit to use:  the draft versionbits BIP and BIP101 use bit 30; to avoid confusion, I think it would be better to use bit 0.

I agree with Jannes Faber, behavior with respect to SPV clients should be to only tell them about fully validated headers. And I also agree that immediately relaying full-proof-of-work blocks before validation (with an indication that they haven't been fully validated) is a good idea, but that discussion didn't reach consensus when I brought it up two years ago (https://github.com/bitcoin/bitcoin/pull/3580).