Proposed prompt; not executed or evaluated.

Review this redacted AI-service retry design using only the supplied SDK/provider docs
and configuration. Do not send requests, run code, change settings or ask for API keys.
Identify each retry layer and count maximum total attempts, including the original.
Flag missing evidence about Retry-After, deadlines, timeouts, cancellation, concurrency,
spend/quota errors, partial streams and repeat safety. Do not infer idempotency from
a function name or an error status. Separate documented behavior from assumptions.
For every proposed change, cite the supplied source and state what test would verify it.
Leave unknowns unresolved. Return a short review checklist, not an unbounded retry loop.

Paste approved redacted configuration and relevant current primary docs here.
