hermes - 💡(How to fix) Fix Classify image_generate empty_response and provide bounded fallback [1 pull requests]

Problem: Gateway sessions can receive repeated image_generate tool failures where the tool returns success=false, error_type=empty_response, and error="Codex response contained no image_generation_call result". In observed logs this happened repeatedly for gpt-image-2-medium, with calls taking roughly 6-136 seconds before failing.

This is distinct from safety rejection, network timeout, and invalid prompt input. Today the failure is surfaced as a generic tool error, which can lead to repeated attempts and weak user-facing recovery guidance.

Expected behavior:

• Classify empty_response as a first-class image generation failure mode.
• Bound retries for this class, ideally at most one retry unless the user explicitly asks again.
• Preserve structured evidence in the tool result: model, error_type, duration, and whether retry was attempted.
• Give the agent a concise recovery hint, e.g. retry later, reduce prompt complexity, or switch generation route.

Acceptance:

• Unit coverage for a response with no image_generation_call result.
• The agent does not enter repeated same-turn image generation attempts for the same empty_response.
• Safety rejections, network failures, and empty provider responses remain distinguishable in logs and user-facing guidance.
• Existing successful image generation behavior is unchanged.