If Substrate has a block 500 with a key gen, and a block 600 with a key gen,
and the session starting on 500 never cosigns everything, everyone up-to-date
will want the cosigns for the session starting on block 500. Everyone
up-to-date will also be rebroadcasting the non-existent cosigns for the session
which has yet to start. This wouldn't cause a stall as eventually, each
individual set would cosign the latest notable block, and then that would be
explicitly synced, but it's still not the intended behavior.
We also won't even intake the cosigns for the latest intended session if it
exceeds the session we're currently evaluating. This does mean those behind on
the cosigning protocol wouldn't have rebroadcasted their historical cosigns,
and now will, but that's valuable as we don't actually know if we're behind or
up-to-date (per above posited issue).
We don't have consensus on the session's last block, so we shouldn't check if
the cosign is before the session ends. What matters is that network, within its
set, claims it's still active at that block (on its view of the blockchain).
An archive of all GlobalSessions is necessary to check for faults. The storage
cost is also minimal. While it should be avoided if it can be, it can't be
here.
Cosigns was an archive of every single cosign ever received. By scoping
NetworksLatestCosignedBlock to be by the global session, we have the latest
cosign for each network in a session (valid to replace all prior cosigns by
that network within that session, even for the purposes of fault) and
automatically have the notable cosigns indexed (as they are the latest ones
within their session). This not only saves space yet also allows optimizing
evaluation a bit.
Not only cleans the existing cosign code but enables non-Serai-coordinators to
evaluate cosigns if they gain access to a feed of them (such as over an RPC).
This would let centralized services not only track the finalized chain yet the
cosigned chain without directly running a coordinator.
Still being wrapped up.
