fix: prevent mutation of original dict when record.msg is a dict

[gaoflow]

Summary

When record.msg is a dict (used as a structured log message), the format() method assigns it to message_dict by reference. Subsequent mutations to message_dict -- such as injecting exc_info or stack_info -- leak into the caller's original dict, causing unexpected side effects.

Reproduction

import logging
from pythonjsonlogger import jsonlogger

logger = logging.getLogger("test")
handler = logging.StreamHandler()
handler.setFormatter(jsonlogger.JsonFormatter())
logger.addHandler(handler)

data = {"key": "value"}
try:
    1 / 0
except:
    logger.exception(data)

# data now contains "exc_info" -- the original dict was mutated!

Fix

Use record.msg.copy() instead of a direct reference, so the original dict remains untouched.

Tests

All 135 existing tests pass.

(This contribution was made under my direction and I take full responsibility for its correctness.)

[nhairs]

Hi @gaoflow, thanks for putting this together.

Overall it makes sense what it's trying to achieve.

I would however like to take some time to think about if this is a change I want to make. In particular, if my understanding of dict.copy() is correct, this will only create a shallow copy of the data. Meaning that if there is any other mutable object within the msg dict, it's possible that it would also be modified. However I'm not sure if something like deepcopy is robust enough here - stability is the number one priority of this library.

Assuming I decide to go ahead with this change, it would also need a unit test to ensure there are no regressions (the sample provided in the description could form the basis of it). We'd also need consider what documentation / doc-strings need to be updated to make this API contract clear.

Finally, it looks like the failing CI is unrelated to your change.