fix(ai): retry upstream availability failures

[minpeter]

Summary

• classify the exact upstream_unavailable provider code as transient and retryable
• keep context overflow, refusals, sensitive output, quota, billing, and generic unknown errors excluded
• prove both the provider-neutral classifier and the no-fallback same-model AgentSession path

This is an independent PR based directly on current main.

Reported failure

Error: upstream_unavailable: Codex upstream websocket send failed via proxy endpoint unknown: ConnectionClosedOK

The existing classifier covered HTTP 503, service-unavailable text, network errors, and narrow WebSocket close
wording, but not the structured upstream_unavailable code. Without a configured fallback chain the turn therefore
settled immediately instead of entering the bounded same-model retry policy.

RED -> GREEN

• RED classifier: expected retryable true, received false
• RED AgentSession: expected one auto_retry_start, received none; one provider call only
• GREEN: exact token \bupstream_unavailable\b admitted
• GREEN classifier: 7/7
• GREEN retry/fallback engine: 7/7
• full AI package suite: 1,191 passed / 761 skipped

Real-surface QA

A local OpenAI Responses server emitted the exact response.failed code/message once, then a successful response.
The real source CLI:

• exited 0
• made exactly two provider requests
• printed the recovery marker
• used only localhost and a dummy local key
• left the real auth file unchanged
• removed its sandbox and closed its server

Root build and npm run check also passed.

---

Summary by cubic

Classifies the exact upstream_unavailable provider error as transient and retryable, so Codex upstream WebSocket send failures trigger the bounded same-model retry policy instead of failing immediately. Non-retryable cases (quota/billing, context overflow, refusals, sensitive output, and generic unknowns) remain excluded; tests validate the provider-neutral classifier and the no-fallback AgentSession path.

^{Written for commit 8b9d28a. Summary will update on new commits.}

[minpeter]

Superseded by current-main Draft PR #330: #330. #330 preserves the exact retry behavior, updates the merged retry event-contract fixture, and avoids force-pushing this stale branch.