This treats any defined forkingConfig as “forking enabled” and skips adding the chain’s default genesis state. However resolveForkingConfig returns an object even when enabled is false, and other code paths (e.g. hardhatForkingConfigToEdrForkConfig) explicitly check enabled === true. Consider using forkingConfig?.enabled === true (and potentially normalizing the cache key the same way) so enabled: false behaves like “no forking”.

The collision/merge logic here is unlikely to work as intended: Map<Uint8Array, ...> uses reference equality for keys, so genesisState.get(account.address) will only hit if the exact same Uint8Array instance is used as the key. Also, even if it hits, the code mutates account but never updates the map entry, so the merged values aren’t reflected in genesisState. If you need to handle address collisions, consider doing lookups via a normalized key (e.g. hex string) and then rebuilding the final Map<Uint8Array, AccountOverride> from the merged overrides.

This is intentional, and explained above.

#resolveNetworkConfig now requires a NetworkConfigOverride, but its caller passes networkConfigOverride which is optional and can be undefined. With strictNullChecks this becomes a type error and also makes Object.keys(networkConfigOverride) unsafe if undefined is ever passed at runtime. Consider accepting NetworkConfigOverride | undefined again and normalizing to {} inside the method, or defaulting at the call site (networkConfigOverride ?? {}).

This comment doesn't make sense. It has = {}, so passing undefined defaulted to {}.

For example, this test:

interface I {
  ii?: number;
}

function test(i: {} | undefined = {}) {
  console.log(i);
}

test();
test(undefined);

Prints

{}
{}

From my reading of this - we create a WeakMap based off the network accounts. This is in effect going to be scoped to the lifetime of the hre; in a test suite this will mainly be the global hre.
We weakmap to the forking config as well, but as we use a "no forking" marker object (the most common use case) which has module scope, for most cases this level will act just as a map.
The next levels are not objects but relevant keys.

My mental model here is that this cache is scoped to the hre object with some optimizations, but in general the genesis states will be cached for the lifetime of the HRE.

I think this makes sense.

Your take is mostly correct.

The only case when it isn't is that when there are network config overrides the config gets re-resolved, so its lifetime is independent of the HRE. That means that this cache isn't hit for config overrides, because the references are different, and that those values can be GC'd after the network connection dies.

I might be tempted to collapse the chainType layer and specId into a single key + map.
I don't feel strongly on this.

We could use the same mutex trick here that you used for the ContractDecoder to ensure there is only one calculation.

That's a good point.

Added it in db3ad1d

For this PR I think this works.
That it forces an ugly test later that can't rely on reference equality for the network config - that causes me to pause.
I think this suggests that network config within networkConfigs should have the chain type set. Or am I missing a subtlety of our config resolution - and the undefined has to be preserved?

Claude suggested memoizing here - but I think that is too complicated to satisfy my desire to get the same networkConfig object.

I also found this surprising as well.

The reason chainType is not always resolved is that undefined lets you change it on connect().

Claude suggested memoizing here - but I think that is too complicated to satisfy my desire to get the same networkConfig object.

Not unreasonable to be honest, it would make the cache simpler. So the complexity would be moved somewhere else.

This is not the only place where the network config is spread, so I added a better explanation in 93d953f

There are whitespace-only lines left behind (e.g. after removing eslint-disable comments). These typically trigger formatting/lint failures (e.g. trailing whitespace/no-multiple-empty-lines) and add noise to diffs. Please remove the stray indentation-only line(s).