We have added retries for "pipeline unreachable" one layer at a time, usually after a CI flake:

The result is retries that stack. With #7868, one unreachable pipeline can cost about 154 s with default client settings instead of about 35 s (see #7868 (comment)). The websocket proxy and the metrics scrape still make a single attempt. is_connect_failure in the manager and HostUnreachable in the coordinator classify the same failures separately.

Proposed rule: handle each kind of failure in exactly one layer, the lowest one that can tell (1) that a repeat is safe and (2) when to stop.

Work:

• Manager retries connect failures on all proxy paths, including websocket, with one budget.
• Manager reports the attempt count in the error details. Clients do not retry when the manager already retried.
• Document the policy (what is retried, the budget, which layer owns it) and the manager budget in the client timeout docs.
• Decide whether the manager and the coordinator should share one connect-failure classification.