Write conversation logs from the in-loop benchmark runner

[t-timms]

Why

experiments/swebench_lite/run_in_loop.py never created a ConversationLogger, so log_conversations: true (which the CLI honours) did nothing for benchmark runs. After a full night of agent-in-loop evaluation (dozens of agent sessions) there was no training/ directory at all. Benchmark runs generate exactly the trajectories any later tuning needs (tool calls, thinking, outcomes, all scored by the official harness), and they were being discarded.

Changes

• The runner creates a ConversationLogger when settings.log_conversations is on, passes it to Conversation, writes the terminal session_end record (exit reason and code from EXIT_REASON_TO_CODE, iterations, tool calls, duration, cost) and closes the file on every exit path (normal, timeout, exception).
• CHANGELOG entry.

Verification

• tests/test_run_in_loop_wiring.py gains two tests: with log_conversations: true one <session>.conversation.jsonl exists, starts with the system record and ends with a session_end record (exit_reason: stopped, exit_code: 0); with it off no file is written. The enabled-case test fails without this change and passes with it (checked).
• ruff check . / ruff format --check . clean.

Not verified

• Not run against a live model; the stub agent loop produces only the system and session_end records, so the per-turn records rely on the existing Conversation logging that the CLI path already exercises.
• The full suite was not re-run on this branch (CI runs it).

[github-actions]

🤖 Godspeed Review

LLM review skipped (no API key configured).

Security

No secrets detected in changed files.