Retry transport failures and name their cause

[rommeld]

Fixes #39

• Add reqwest's wording to RETRYABLE_PATTERNS. The list was ported from pi and uses Node's fetch wording, so reqwest transport failures never matched it, and the turn ended instead of auto-retrying. Two phrases are added:
  • error sending request: nothing came back. This is pi's fetch failed.
  • error decoding response body: the SSE stream broke mid-response. This is pi's terminated.
• Print the reqwest error's source() chain in InferenceError::Http, so the retry notice and the session file say why the request failed. Example: …error sending request for url (…): client error (SendRequest): connection error: Connection reset by peer (os error 54). This deviates from pi, whose message is only fetch failed.
• Tests: both messages are now in transient_errors_are_retryable. The new transport_errors_name_their_cause checks that a refused connection mentions Connection refused.

End-to-end check (a throwaway test, not committed): the real Codex provider and agent loop ran against a local server that resets the first connection.

• Before: 0 retries, and the turn fails.
• After: 1 retry, and the turn completes.

A reset in the middle of the SSE stream gave the same result.

Validation: cargo fmt --check; cargo clippy --workspace --all-targets -- -D warnings; cargo test --workspace (436 passed).

🤖 Generated with Claude Code