Recover an engagement wedged on an orphan ext-mod write-ahead.

Two-step recovery, both in one transaction:

1. Flip EventAction.success from NULL → False so the _any_pending_writeahead_for_modules gate stops firing.
2. Persist a synthetic ModelRequest LLM_RESPONSE event with a ToolReturnPart for the dangling tool call. This clears the _has_dangling_tool_call / _dangling_ext_mod_tool_names gates so the next turn sees a complete tool exchange. The content explicitly marks the tool as failed and tells the model not to retry.

We deliberately do NOT republish ENGAGEMENT_RUN_REQUESTED — the engagement is already in (or destined for) the DLQ, and the existing redrive will pick it up. Republishing here would race with that and queue an extra turn.