high level response:

including a small number of block headers (10?) directly as op_return
metadata (or something) doesn't have that high overhead necessarily, but
could be super effective at helping miners participate with lower hashrate.
the reason to include this as on-chain data is so that the mining pool
doesn't require any external network software.

this would balance out the issues if the data is somewhat bounded (e.g., 10
headers). what's nice is this data has no consensus meaning as it's client
side validated by the DCFMP block filter.

interestingly, the participating pools could 'vote' on how difficult shares
should be as a metaparameter to the pool over blocks... but analysis gets
more complex with that.

cheers,

jeremy