Improve AI provider reliability: timeouts, retries, streaming errors, and fallback handling

[michaelmanly]

I noticed a few AI provider/runtime paths where production reliability behavior may currently depend on broad defaults or single-route execution:

• packages/fetch/src/getAgentOptions.ts

  • TIMEOUT appears to default to 7200 seconds
  • this may allow hung provider calls to tie up CLI/worker resources for up to 2 hours unless overridden

• core/util/withExponentialBackoff.ts

  • retries appear to run up to 5 times with deterministic exponential delay
  • I did not see jitter, max-delay caps, or an overall retry budget
  • this may cause synchronized retry storms during 429 or overload events

• core/util/withExponentialBackoff.ts

  • retry matching appears to include "malformed json"
  • this may retry deterministic parse/schema failures instead of surfacing structured diagnostics

• core/llm/llms/OpenAI.ts

  • _streamResponses appears to catch malformed streaming chunk parse errors and ignore them
  • this may hide partial response corruption or dropped stream events

• core/llm/streamChat.ts

  • streaming appears to call model.streamChat(...) and rethrow on failure
  • I did not see an optional runtime fallback chain if the selected provider/model fails

Would maintainers be open to a small PR to improve one of these reliability paths?

Possible small PR options:

1. reduce or make safer the default timeout behavior
2. add jitter/caps/retry budgets to exponential backoff
3. avoid retrying deterministic malformed JSON errors
4. add structured logging for dropped streaming chunks
5. document or support OpenAI-compatible gateway base URLs for teams that want external retries, fallback routing, timeout handling, and request tracking

The goal would be optional production reliability improvements without changing default behavior aggressively or adding a required dependency.