LGTM! I asked Claude if any classes were accidentally left out, but we're good to go.

I'm interested what you got cooking with mypyc and what the results are!

I added Claude's summary below on what which classes get the decorator and which classes are omitted (and why) as documentation.

What got @final

491 classes, all leaf @slotted_freezable @dataclass (or plain @dataclass) types, across all 24 fork directories plus 5 shared modules (crypto/blake2.py, genesis.py, merkle_patricia_trie.py, state.py, trace.py). The decorator is placed outermost (above @slotted_freezable / @dataclass), which is correct: slotted_freezable returns a freshly built class object, and since decorators apply bottom-up, the outermost @final marks that final object.

These data types are the ones that benefit: @dataclass generates __init__, __eq__, and __hash__, which are real calls in the codebase (equality, dict/set keys), and @final lets mypyc bypass the vtable and emit direct native calls for them.

What was omitted, and why it is correct

mypyc's @final optimization only bypasses the vtable for method calls and property accessors. A class with no methods or properties has nothing to devirtualize, so the following omissions cost nothing:

• Leaf exception classes (OutOfGasError, StackUnderflowError, TransactionTypeError, etc.): pass bodies, no methods or properties, and never method-called in hot paths.
• KZG scalar types in crypto/kzg.py (KZGCommitment, KZGProof, BLSFieldElement, VersionedHash): also pass bodies, nothing to optimize.
• ForkCriteria subclasses in fork_criteria.py (ByBlockNumber, ByTimestamp, Unscheduled): these do have methods (check, __repr__), but they are invoked through the ForkCriteria base type (so devirtualization would not fire at the call sites) and only at fork-scheduling time, not a hot path.

These are also correct on the merits regardless of mypyc:

• GasCosts: a plain namespace class carrying the comment "may be patched at runtime by a future gas repricing utility", so it is intentionally left mutable.
• Ops and other Enum subclasses: enums with members cannot be subclassed anyway, so @final would be redundant.
• PreState and EvmTracer: typing.Protocol subclasses, where @final would be semantically wrong since protocols are meant to be implemented.

Sources

• mypyc devirtualizes method and property calls on final classes: python/mypy#17886.
• mypyc native classes overview: https://mypyc.readthedocs.io/en/latest/native_classes.html.
• typing.final semantics: https://mypy.readthedocs.io/en/stable/final_attrs.html.