fix(logger): preserve traceback in JSON sink

[mikasenghaas]

Summary

• Pre-format the loguru traceback into record["extra"]["traceback"] via a traceback_patcher that runs at log time, before enqueue=True pickles the record. The JSON sink reads it back into the "exception" field, restoring full stack frames in JSON logs.
• Keeps enqueue=True — required for multiprocess safety of trainer torchrun workers sharing stdout.

Why

When --log.json_logging is enabled, exception logs lose their stack frames. Reproduction with the orchestrator entrypoint produced:

{"level": "ERROR", "message": "Fatal error in orchestrate", "exception": "RuntimeError: ...\n"}

No Traceback (most recent call last): header, no frames — just the exception type + message.

Root cause

loguru/_recattrs.py:68 — RecordException.__reduce__ unconditionally sets the traceback to None before pickling, because traceback objects aren't picklable:

def __reduce__(self):
    try:
        pickled_value = pickle.dumps(self.value)
    except Exception:
        return (RecordException, (self.type, None, None))
    else:
        return (RecordException._from_pickled_value, (self.type, pickled_value, None))
                                                                                  # ^^^^ traceback dropped

With enqueue=True the record is pickled to cross the queue, so by the time json_sink receives it both exc.traceback and exc.value.__traceback__ are None, and traceback.format_exception(...) renders only the head line.

Verification

Injected raise RuntimeError("...") at the start of orchestrate(), ran orchestrator @ configs/debug/orch.toml --log.json_logging:

Before:

{"level": "ERROR", "exception": "RuntimeError: ...\n"}

After:

{"level": "ERROR", "exception": "Traceback (most recent call last):\n  File \".../utils.py\", line 59, in async_wrapper\n    ret = await func(*args, **kwargs)\n  File \".../orchestrator.py\", line 96, in orchestrate\n    raise RuntimeError(...)\nRuntimeError: ...\n"}

Why not drop enqueue=True?

Per loguru docs:

"Whether the messages to be logged should first pass through a multiprocessing-safe queue before reaching the sink. This is useful while logging to a file through multiple processes."

The trainer's torchrun workers all write JSON lines to a shared inherited stdout. Without the queue, concurrent writes from different processes can interleave at byte boundaries and corrupt JSON lines. Loguru's per-sink lock handles intra-process threading but is not multiprocess-safe.

---

Note

Medium Risk
Changes how exceptions are captured and serialized in JSON logging mode (via Loguru patchers and extra mutation), which could affect error visibility and structured log fields across async/multiprocess logging.

Overview
Fixes JSON logging so exception stack traces are preserved when the sink runs with enqueue=True.

This pre-formats the traceback at log time via a new traceback_patcher (stored in record["extra"]["traceback"]) and updates build_log_entry() to prefer that preformatted traceback when populating the JSON exception field, falling back to formatting record["exception"] only when needed.

^{Reviewed by Cursor Bugbot for commit 7d6b55b. Bugbot is set up for automated code reviews on this repo. Configure here.}