Transport: httpx network errors (ConnectError/NetworkError/etc.) never retry, even for GET/HEAD/OPTIONS #240

Summary

Transient ECONNRESET, DNS hiccups, HTTP/2 RST_STREAM, server half-closes, and TCP connect refusals on a safe idempotent GET (markets list, portfolio snapshot, etc.) surface as a one-shot failure to the caller. The retry loop only handles TimeoutException and PoolTimeout; every other httpx.HTTPError subclass lands in a final except and re-raises with no retry.

Industry-standard SDKs (openai-python, anthropic-python, stripe-python, urllib3) all retry idempotent verbs on transport-level errors. ConnectError on POST is in principle also safe to retry (request never reached the wire, mirroring the PoolTimeout reasoning from #204).

Location

• kalshi/_base_client.py:284-286 — sync transport final except httpx.HTTPError
• kalshi/_base_client.py:466-467 — async transport mirror

Evidence

except httpx.HTTPError as e:
    raise KalshiError(f"HTTP error: {method.upper()} {path}") from e

# httpx hierarchy: ConnectError/NetworkError/RemoteProtocolError/ReadError/WriteError
# all inherit from TransportError -> RequestError -> HTTPError.

Recommended fix

Insert an except httpx.TransportError (or split out ConnectError/NetworkError/RemoteProtocolError/ReadError/WriteError) between TimeoutException and HTTPError that participates in the same RETRYABLE_METHODS + backoff + total_timeout budget loop. Mirror the PoolTimeout "never reached the wire" branch for ConnectError so POST/DELETE can retry only on pre-wire failures.

Severity & category

high / reliability