Lookahead fix

[djrtwo]

address #1269

current_epoch - 1 for randomness source doesn’t give us lookahead. It is just in time and it’s the bare minimum stable source of randao for any epoch. MIN_SEED_LOOKAHEAD of 0 should equate to the current_epoch - 1 case. MIN_SEED_LOOKAHEAD > 0 should have a base of the -1

[protolambda]

Just because it is possibly the most important off-by-1 prone math in the spec, here's my take (and thanks @djrtwo for rubber duck debugging 😂 ):

• process_slots, from start state.slot == x*SLOTS_PER_EPOCH - 1, i.e. current_epoch == x - 1
• (state.slot + 1) % SLOTS_PER_EPOCH == 0 -> epoch boundary detected, process epoch.
• process_final_updates() as last epoch sub transition
• state.randao_mixes[next_epoch % EPOCHS_PER_HISTORICAL_VECTOR] = get_randao_mix(state, current_epoch), a.k.a. state.randao_mixes[x % EPOCHS_PER_HISTORICAL_VECTOR] = get_randao_mix(state, x - 1)
• epoch transition finishes
• slot transition finishes, slot is set to x*SLOTS_PER_EPOCH, now current_epoch == x

So transition looks ok at first: during an epoch randao_mixes is available up to the current epoch as index: state.randao_mixes[current_epoch % EPOCHS_PER_HISTORICAL_VECTOR] get the very latest known randao mix. But it changes with process_randao during the current_epoch.

MIN_SEED_LOOKAHEAD = 1

Without this PR:

• beacon proposer indices are looked up with get_seed(state, current_epoch). This requires state.randao_mixes[(current_epoch + EPOCHS_PER_HISTORICAL_VECTOR - MIN_SEED_LOOKAHEAD) % EPOCHS_PER_HISTORICAL_VECTOR], equivalent to state.randao_mixes[previous_epoch % EPOCHS_PER_HISTORICAL_VECTOR]. Which was initially set during the transition from previous_epoch - 1 to previous_epoch. And is then modified during previous_epoch. Unpredictable till the very end of previous_epoch. But known during all of current_epoch.
• beacon proposer indices for the next epoch looked up with get_seed(state, current_epoch + 1). This requires state.randao_mixes[((current_epoch + 1) + EPOCHS_PER_HISTORICAL_VECTOR - MIN_SEED_LOOKAHEAD) % EPOCHS_PER_HISTORICAL_VECTOR], equivalent to state.randao_mixes[current_epoch % EPOCHS_PER_HISTORICAL_VECTOR]. Which was initially set during the transition from current_epoch - 1 to current_epoch. And is then modified during current_epoch. Unpredictable till the very end of current_epoch. But known during all of current_epoch + 1.

With this PR:
offset lookups down by 1.

Next epoch proposer index lookup is now requiring: state.randao_mixes[((current_epoch + 1) + EPOCHS_PER_HISTORICAL_VECTOR - MIN_SEED_LOOKAHEAD - 1) % EPOCHS_PER_HISTORICAL_VECTOR], a.k.a. state.randao_mixes[previous_epoch % EPOCHS_PER_HISTORICAL_VECTOR]. Which was initially set during the transition from previous_epoch - 1 to previous_epoch. And is then modified during previous_epoch. Unpredictable till the very end of previous_epoch. But known during all of current_epoch.

So yes, we need this PR to make the next epoch predictable during all of the current epoch.