Most of the issues pointed out by int_19h would be handled at compile time during static analysis and wouldn't change much to the generated bytecode. I'm talking about strong typing, immutability by default, less error-prone syntax, tail calls, evaluation order etc...
Even replacing 256bit ints with arbitrary precision "bigints" wouldn't add too much of a cost if it's a native type of the underlying VM (as it should be for such an application IMO). It might even reduce code size by removing overflow tests.
But I would still expect arithmetic to be overflow-checked by default, as in e.g. VB.NET. This would mean that careless arithmetic on unvalidated inputs could still cause the contract to fail - but at least it would be a controllable failure, with all state changes being reversed, and hence not exploitable.
Even replacing 256bit ints with arbitrary precision "bigints" wouldn't add too much of a cost if it's a native type of the underlying VM (as it should be for such an application IMO). It might even reduce code size by removing overflow tests.