AIKIDO-2026-768557

TL;DR

When Python integers larger than the u64/i64 range are passed into the Starlark sandbox, pyobject_to_value routes them through JSON serialization where the BigInt fallback branch was commented out, so they silently widen to 64-bit floats instead of becoming Starlark integers. Distinct large integers then compare equal and lose precision, so checks such as amount <= limit can pass when they should fail, with no error raised. The widening is invisible from Starlark because reads return the corrupted values. The fix converts Python integers directly to arbitrary-precision Starlark integers.