Allow long high-reasoning reviews

[dymoo]

Outcome

• raises the fixed non-streaming model-call deadline from 3 minutes to 10 minutes
• applies one shared deadline across every retry attempt and response decoding
• treats timeout/abort as terminal so inference is never duplicated after a timeout
• retries only fast connection failures, 429s, and 5xx responses while backoff fits the remaining budget
• keeps the limit product-owned and documents the consumer workflow/proxy timeout contract

Production evidence

NeptuneQuant PR #8 exercised the dedicated internal CLIProxy key from its ARC runner. Five requests returned HTTP 200; three consecutive high-reasoning synthesis attempts were client-cancelled at exactly 180 seconds and surfaced as proxy HTTP 500 responses. Authentication, DNS, routing, and model selection are healthy; the action deadline is the demonstrated failure boundary.

Validation

• npm run format
• npm run check-all
• 100/100 tests passed, including deterministic timeout/non-duplication, response-decoding deadline, and shrinking shared-budget retry regressions