Block re-orgs need some thought.
If I have chain A-B-C, and get IBLT's for an alternative chain A-B'-C'-D'...
... then the current memory pool won't work to try to reconstruct B' C' D'.
Using B and C to reconstruct B' and C' should work pretty well. Then the remaining memory pool transactions can be used to reconstruct D.
If any of the reconstructions fail, just fall back to fetching all of B' C' D'.
Then again, re-orgs are rare enough that always falling back to fetching full blocks would be OK.