fix(websocket): tighten upstream WS dead-peer window under ~30s head silence

[snowkide]

Summary

• Shorten upstream WS liveness from 30s ping / 75s pongWait to 10s / 25s so a black-holed upstream is torn down and re-dialed before typical downstream head-silence thresholds (~30s).
• Keeps pongWait > 2× pingInterval (two ping chances before declare-dead).

Why (single Chainlink RPC = eRPC)

Chainlink pins one sticky client WS to one eRPC pod. eRPC’s internal HTTP/upstream failover on other pods does not move that socket.

If the head-feeding upstream on that pod blackholes, eRPC previously waited up to 75s before re-dial. Consumers (e.g. Polygon NoNewHeadsThreshold = 30s) already treat silence as unhealthy by then — so the node looks “eRPC down” while the rest of the eRPC fleet can still tip.

This does not refuse client eth_subscribe or push failure into Chainlink. It only makes upstream self-heal start sooner so the sticky client keeps getting heads.

Out of scope

• Process shutdown 1001 GoingAway (DaemonSet roll) — still closes the client WS; needs roll/drain hygiene
• Selection/breaker probe wedge — see fix(selection): make recovery probes breaker-ineligible (probe/selection wedge defence-in-depth) #4
• Initializer auto-retry — see fix(initializer): keep auto-retry alive for transient tasks despite fatal or hung siblings #5

Test plan

• go test ./clients/ -run 'WsJsonRpc|Pong|Reconnect|Liveness|Dead'
• go test ./indexer/adapters/wsupstream/ -run 'Reconnect|Healthy|Resub'
• go test ./erpc/ -run UpstreamDiesUngracefully_SelfHeals
• Canary: under upstream blackhole, erpc_upstream_websocket_connected drops and re-subscribes within ~25s; Chainlink head sub on that pod does not sit silent past ~30s

Made with Cursor