feat(bft): keepalive + typed payloads + pending-map race fix

[satyakwok]

Summary

WebSocket robustness pass on `SubscriptionManager`. Three concrete bugs + one ergonomic add + a small naming drift fix.

Bug 1: pending-map race on socket error / close

Pre-fix: only the first pending subscribe saw the rejection; every other in-flight subscribe hung until its 10 s timeout fired one-by-one. Now error + close handlers iterate the pending map and reject every entry with the same surfaced error so callers see failure immediately.

Bug 2: subscribe-response error path swallowed

Pre-fix: only `{ id, result }` was handled; a server-side error reply `{ id, error: { message } }` left the pending caller hanging until timeout. Now the message handler checks for `.error` and rejects with the server's message.

Bug 3: middlebox idle-kill

Caddy `reverse_proxy idle_timeout`, NAT, AWS ALB all drop quiet connections at 60–120 s. Added:

• `KEEPALIVE_INTERVAL_MS = 30 s` — WebSocket ping every 30 s.
• `STALE_TIMEOUT_MS = 90 s` — if no frame (event, pong, subscribe-response) lands within 90 s, terminate the socket; close handler reconnects via exponential backoff.

Feature: `subscribeTyped()`

`ChannelPayloadMap` discriminated union maps each Channel to its payload type — `subscribeTyped("newHeads", ...)` gives `payload: NewHeadsPayload` instead of `unknown`. Backwards-compatible: untyped `subscribe()` still works.

Plus per-sub `onError` stored on `InternalSub` so a reconnect-time re-subscribe routes failures back to the original caller, not just the manager-level handler.

Status method

`mgr.status()` returns `{ socketState, subs, secondsSinceLastFrame }` for ops dashboards / debug pages.

Naming drift fix

• `sentrixMainnet.name` was `"Sentrix Mainnet"` — canonical brand is `"Sentrix Chain"` (matches chainlist registry `Add Sentrix Chain mainnet (chainId 7119) and testnet (chainId 7120) ethereum-lists/chains#8266` + every frontend chain config).
• Testnet `explorerUrl` was the shared `scan.sentrixchain.com` (testnet tx links opened the mainnet view) — now `scan-testnet.sentrixchain.com` for proper EIP-3091 deeplink routing.

Test plan

• `pnpm build` passes
• Open a long-running subscription against testnet; verify the connection stays alive across a 5-min idle window (ping keeps Caddy from killing it)
• Trigger a subscribe-response error (eg invalid filter) — caller's promise rejects with the server's message instead of timing out
• Trigger a sudden socket close (ngrok kill, Caddy reload) — every in-flight `subscribe()` rejects immediately with `"websocket closed before subscribe response"`
• Use `subscribeTyped("newHeads", ...)` in a TS playground — payload is typed as `NewHeadsPayload`, `.number` autocompletes