Runner: sidecar oversized Codex tool results

[gaofeng21cn]

Problem

Large Codex runner.tool_result events can bloat quest event logs and consume follow-up context. The runner also had no durable telemetry for prompt/tool-result byte budgets or compacted result counts.

Solution

• Add a generic evidence_packets helper that writes oversized tool-result payloads to quest-local .ds/evidence_packets/<run_id>/ sidecars.
• Keep runner.tool_result events parseable by replacing oversized output with a compact JSON packet containing path, hash, payload bytes, summary, and key blockers.
• Add Codex runner telemetry for prompt bytes, stdout bytes, tool-result bytes, compacted result count, full-detail tool-call count, token usage, and the telemetry sidecar path.

MDS provenance

Adapted the generic context-budget parts of MedDeepScientist commits 793ff56 and b911a53. MedDeepScientist, MAS, medical-fork, manuscript-readiness, and publication-authority semantics were intentionally excluded.

Tests

• python -m py_compile src/deepscientist/evidence_packets.py src/deepscientist/runners/codex.py
• git diff --check
• uv run pytest -q tests/test_codex_runner.py::test_codex_tool_event_preserves_parseable_bash_exec_payload_and_metadata tests/test_codex_runner.py::test_codex_tool_event_truncates_oversized_bash_exec_log_but_keeps_json_parseable tests/test_codex_runner.py::test_codex_compacts_extreme_tool_result_payload_before_event_write tests/test_codex_runner.py::test_codex_runner_sidecars_oversized_tool_result_event tests/test_codex_runner.py::test_codex_runner_leaves_normal_tool_result_event_parseable tests/test_codex_runner.py::test_codex_runner_writes_tool_result_telemetry_and_sidecar