Job_chat: Replace correction loop with tool outputs

If code replacements fail in job_chat, we run a code correction LLM call once. This made more sense when the service wasn't agentic. We should instead allow the agent to call the code correction tool more than once, and surface errors to the editing model instead of a separate instance.

Removing the code correction and allowing code correction loops will simplify the code, but possibly introduce latency and streaming issues. If we allow the service to emit multiple text and code edit events, we need to make sure that we capture them appropriately when the service is called as the main agent and as a subagent.

Start by tracing job_chat and the code correction tool, then inspect how text and code edit events are captured for the main agent and subagents. The work is done when correction can be called repeatedly, errors reach the editing model, and multiple emitted events are handled without breaking streaming.