Read-modify-save races on conversation metadata can clobber concurrently updated fields (title processor)

SetTitleCallbackProcessor fetches the conversation, polls the agent-server for a generated title (up to ~12s), then saves the ENTIRE stale snapshot via save_app_conversation_info (openhands/app_server/event_callback/set_title_callback_processor.py, ~110-148). save_app_conversation_info rewrites metrics, llm_model, and other fields from the snapshot, so any stats update that lands during the poll window is silently erased (classic lost-update).

Other read-modify-save callers of save_app_conversation_info (e.g. PR-metadata updates) have the same shape.

OpenHands/OpenHands#15354 reduces the metrics blast radius (the next stats event/run-end pull restores totals, and the cost-event ledger is append-only so per-model attribution survives), but the underlying pattern remains. Fix direction: column-specific updates (e.g. a dedicated update_title) or optimistic locking/version stamping on the metadata row.

Found while root-causing cost mis-reporting on LLM switches (#15354/#15355); confirmed by two independent code reviews.

Read openhands/app_server/event_callback/set_title_callback_processor.py around lines 110-148, then trace save_app_conversation_info and the other callers mentioned in the issue. Reproduce or add coverage for a concurrent metadata update during title polling; done means title changes no longer overwrite fields updated by another operation.