It is always possible I'm being dense, but I still don't understand how
this proposal makes a chain-forking situation better for anybody.

If there are SPV clients that don't pay attention to versions in block
headers, then setting the block version negative doesn't directly help
them, they will ignore it in any case.

If the worry is full nodes that are not upgraded, then a block with a
negative version number will, indeed, fork them off the the chain, in
exactly the same way a block with new hard-forking consensus rules would.
And with the same consequences (if there is any hashpower not paying
attention, then a worthless minority chain might continue on with the old
rules).

If the worry is not-upgraded SPV clients connecting to the old,
not-upgraded full nodes, I don't see how this proposed BIP helps.

I think a much better idea than this proposed BIP would be a BIP that
recommends that SPV clients to pay attention to block version numbers in
the headers that they download, and warn if there is a soft OR hard fork
that they don't know about.

It is also a very good idea for SPV clients to pay attention to timestamps
in the block headers that the receive, and to warn if blocks were generated
either much slower or faster than statistically likely. Doing that (as
Bitcoin Core already does) will mitigate Sybil attacks in general.

-- 
--
Gavin Andresen