Fix silent stalls in native_subscribe WS subscriptions (Solana, eth pending tx)

[a10zn8]

Summary

• Surface idle WS subscription stalls as TimeoutException (instead of silent completion) so the wrapping DurableFlux re-issues subscribe with a fresh subId.
• Stop swallowing subscription errors with onErrorResume { Mono.empty() } — let them propagate to DurableFlux, which already handles reconnect with exponential backoff.
• Applied to both subscription paths: GenericSubscriptionConnect (Solana slotSubscribe, plus all chains that go through GenericIngressSubscription: Polkadot, Ripple, etc.) and WebsocketPendingTxes (Ethereum newPendingTransactions).

Root cause

Solana RPC nodes commonly stop delivering events on a specific subscription while keeping the WS TCP/frame layer fully alive (no ping/pong on our side either — WsConnectionImpl runs with handlePing(false) and no idle handler). The previous code in GenericSubscriptionConnect.createConnection() had:

.timeout(Duration.ofSeconds(85), Mono.empty<ByteArray>().doOnEach { log.warn(...) })
.onErrorResume { log.error(...); Mono.empty() }

When the upstream stalled for 85 s:

1. .timeout(d, Mono.empty()) switched to an empty fallback → Flux completed (no error).
2. DurableFlux.connect() only retries on onError, not onComplete → no resubscribe.
3. SharedFluxHolder propagated onComplete to all downstream subscribers and cleared current.
4. The 30 s heartbeat in NativeSubscribe.nativeSubscribe kept firing → the client connection looked healthy but never received another data event.

WebsocketPendingTxes.createConnection() had the identical anti-pattern.

The healthy reference is GenericWsHead.listenNewHeads() which already uses .timeout(d, Mono.error(...)) and resubscribes on the error.

Test plan

• GenericSubscriptionConnectTest.emits TimeoutException when upstream goes silent past idle timeout — virtual-time test pinning the new error signal.
• GenericSubscriptionConnectTest.idle timeout triggers resubscribe via DurableFlux retry — end-to-end: first subscription stalls (Flux.never), virtual time advances 90 s, second subscribe() delivers a "recovered" event; verifies subscribe(...) was called twice.
• WebsocketPendingTxesSpec.Emits TimeoutException when upstream goes silent past idle timeout — same pattern for the Ethereum pending-tx path.
• Existing Produces values test still passes (happy path unchanged).
• ./gradlew compileKotlin compileTestKotlin clean.

Follow-up (out of scope)

WsConnectionImpl has no WS-level ping/idle handler, so a half-open TCP connection still relies on OS TCP timeouts (~minutes). With this PR, the subscription layer at least heals on its own; the connection layer is a separate hardening task.

🤖 Generated with Claude Code