Implement divergent EC chain in simulations

[masih]

So far simulations used the same EC chain across all participants after the first instance. This approach was simple but not representative of the real-world where the EC chain may diverge across multiple subsets of participants.

The changes here implement a comprehensive set of APIs that allow a set of EC generation strategies to mix-and-match and replicate complex EC chain evolution through iterations of a simulation. The chain evolution is controllable by instance by actor ID. This allows us to simulate fluctuating EC evolution patterns throughout a simulation. The changes include an initial set of EC generators that cover:

• Fixed EC for all instances and participants.
• Random EC per instance per participant.
• Random EC per instance but uniform across all participants.
• Common prefix EC per instance with probabilistic length
• Majority common prefix across a set of participants with probabilistic length

The EC generators can be aggregated and set per number of participants, which enables simulation of a flexible set of scenarios including evolving majority common prefix.

Fixes #115

[anorth]

I skimmed, looking mainly at the tests and how they changed. The main code still needs a proper review.

Looks good except for some confusion around which value is being expected, or how it's fetched, in the multi-instance tests.

[Kubuxu]

SGTM

[anorth]

Thanks for pushing forward to land something.

I think what the suggested code snippet meant to be is requireConsensusAtInstance(t, sm, instanceCount-1, <basechain-of-instance-after-last>.Head())

Yes. basechain-of-instance-after-last is what I meant by expected.

which is an unreasonable expectation in scenarios with EC divergence among participants

But what about scenarios without EC divergence? I'm just trying to preserve the behaviour in those.

The code that landed in multi_instance_test looks fine. If the same thing can be achieved without running the sim into instance one-past-the-end, that will be fine too. But if what I'm saying sounds like nonsense to you, probably best we leave it as-is.

[masih]

If the same thing can be achieved without running the sim into instance one-past-the-end, that will be fine too.

Thanks Alex, I think at the depth current tests assert things it makes no difference.

We might want to revisit this if we decide to simulate failure scenarios, where we expect the simulation to not converge and eventually converge back. The current simulation logic simply errors out in such cases. For now what's there seems reasonable. 🙏