0.1.8 Beta7: Bound request-local retries without circuit breaking #256

This was generated by AI during triage.

Parent

#319 (part of #248)

Outcome

Freeze the request-local retry contract before the Beta7 candidate is published and RC qualification begins, without adding circuit breaking. Main generation may retry only a proven-safe pre-transmission transport failure within fixed attempt and delay budgets; qualification cases never hide retries.

Acceptance criteria

• Third-party main generation uses at most two total attempts.
• Retry eligibility is limited to a proven-safe DNS/connect/TLS failure before request-body transmission.
• Cumulative retry delay is at most 15 seconds.
• No retry occurs after response headers, body bytes, SSE events, or downstream-visible output.
• Idempotent catalog/model GET retry remains a separate bounded policy.
• Qualification cases perform no hidden retry and report degraded/recovered attempts explicitly.
• Responses and Chat regressions prove no duplicate tool execution or terminal event.

Blocked by

• 0.1.8 Beta6 gate: Publish usable Ollama Chat with complete Collaboration V2 #259