Summary

When context compression fires, the model sometimes returns empty on the next turn. Hermes sends (empty) to the user and the conversation dies. Two fixes:

1. Post-compaction empty recovery — sets a _just_compacted flag after compression. If the very next model response is empty, injects a continuation prompt and retries instead of sending (empty).

2. Retry via provider routing — if the primary summary model fails (bad config, API error, timeout), retries the same call_llm(task="compression", ...) call without the explicit model override. call_llm already has its own provider fallback chain, so this goes through the normal routing pipeline instead of hardcoding a specific provider/model in core code.

Also upgrades summary-failure logging from warning to ERROR with model name and error details, so misconfigured summary models are visible in logs immediately.

What happens today without this

• User has summary_model pointing at a broken or unreachable endpoint
• Compression fires, summary generation fails silently, middle turns get dropped without any summary
• Model sees the truncated context, returns empty
• User gets (empty) sent to them, conversation is effectively dead

Test plan

• py_compile passes on both files