It'd be helpful if we thoroughly documented some caveats with the legacy driver's approach of converting BSON documents to arrays (which is still possible in the new driver, albeit not the default).

At the very least, we should discuss the edge case where the driver cannot reliably infer if ['0' => 'foo'] should be serialized back to a BSON document or array. Additionally, we can discuss some best practices for using classes for BSON documents and arrays in the driver's type maps, which can allow users the convenience of ArrayAccess (for objects and/or arrays) without sacrificing any type preservation.