[Yushu-Li]

When callers reuse or mutate a metrics, prompt, artifacts or metadata dictionary after log_trace(), buffered traces currently pick up those later changes. The improvement delta and statistics were already calculated, so an exported transition can contain inconsistent metrics and deltas; immediate JSONL writes also behave differently from buffered JSONL and JSON output.

I snapshot the retained mutable payloads at log time, including nested artifacts and metadata. Code/response fields and prompt inclusion settings keep their existing behavior. Three regression methods exercise real JSON/JSONL files, reused programs across two transitions, nested payload reuse, and excluded prompts. The original source produces four failing subcases, while immediate JSONL and prompt-exclusion controls pass.

Validation (Python 3.12, no live model calls):

• Trace tests: 28 passed.
• Full unit suite: 1,006 tests run, OK with 2 skipped.
• Repository-pinned Black/isort hooks and whitespace checks passed.

HDF5 payload-type handling and export-level metadata reuse are outside this change; this is separate from #574 and #575.

I used Codex assistance for implementation and regression testing.