Thanks for the support Nick. I think your proposed idea would still result in rare but confusing behavior, which is the type of surprise Python should avoid.

The hash-based pyc files doesn't seem like a solution to me, because it won't be enabled by default. And I think it's obvious the performance loss of doing so is unacceptable.

If changing the bytecode format is unacceptable, though it seems like the cleanest answer, the import machinery could just avoid caching bytecode when the int(mtime) of the source file is potentially in the future.