fix: preserve unsafe integers in OpenAI trace payloads

[pragnyanramtha]

Summary

Fixes #2094 by normalizing trace input/output fields sent to the default OpenAI tracing ingest endpoint so integers outside JavaScript's safe integer range are stringified before browser-based trace viewers can round them. Runtime tool inputs and custom tracing endpoints are unchanged.

Root cause: function trace inputs can contain JSON strings or structured payloads with Python integers larger than Number.MAX_SAFE_INTEGER. Those values are valid JSON numbers, but browser JSON viewers can lose precision when parsing/displaying them.

Change:

• Add OpenAI-endpoint-only trace field sanitization for unsafe integers in JSON strings and structured input/output payloads.
• Preserve safe integers, booleans, original payload objects, and custom endpoint behavior.
• Keep existing trace field truncation behavior after the numeric normalization step.

Test plan

• uv run pytest tests/test_trace_processor.py -k 'unsafe_integers or keeps_unsafe_integers or truncates_oversized_output or keeps_small_input'
• uv run pytest tests/test_trace_processor.py -m 'not serial'
• uv run ruff format --check src/agents/tracing/processors.py tests/test_trace_processor.py
• uv run ruff check src/agents/tracing/processors.py tests/test_trace_processor.py
• git diff --check

Not run to completion:

• uv run pytest tests/test_trace_processor.py because existing serial test test_tracing_atexit_cleanup_timeout_preserves_process_exit_code_on_504 timed out; rerunning that single test also timed out.
• uv run mypy src/agents/tracing/processors.py tests/test_trace_processor.py because it produced no output for several minutes and was stopped.

Issue number

Fixes #2094

Checks

• I've added new tests (if relevant)
• I've added/updated the relevant documentation
• I've run make lint and make format
• I've made sure tests pass

Notes: no docs were needed for this internal tracing-export normalization. I ran focused ruff format/lint checks on the touched files rather than full make lint/make format.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.