docs: add evidence-driven RTDS troubleshooting guide

README.md

@@ -233,3 +233,6 @@ When the connection is stablished, if a `filter` is used, the server will dump a
 | symbol | string | Symbol of the asset                                              |
 | data   | array  | Array of price data objects, each containing timestamp and value |
 
+## Troubleshooting and bug reports
+
+For production diagnostics and an actionable issue-report format, see [RTDS troubleshooting](./TROUBLESHOOTING.md). Reports are most useful when organized as **symptom → evidence → workaround → requested platform fix**.

TROUBLESHOOTING.md

@@ -0,0 +1,100 @@
+# RTDS troubleshooting and actionable bug reports
+
+The most useful production reports let another builder reproduce the failure and let maintainers identify the responsible layer. Use this sequence:
+
+> **Symptom → evidence → workaround → requested platform fix**
+
+Do not include API keys, wallet private keys, authentication headers, session tokens, or other credentials in an issue or log attachment.
+
+## Report template
+
+### Symptom
+
+Describe the externally visible failure in one or two sentences.
+
+Examples:
+
+- "The socket remains `OPEN` and answers pings, but all subscribed topics stop delivering payloads."
+- "A subscription is accepted, but the requested topic never produces an initial snapshot."
+- "Messages continue to arrive, but their source timestamps are delayed by more than 30 minutes."
+
+Include whether the failure is intermittent or permanent and whether a new connection recovers it.
+
+### Evidence
+
+Provide enough detail to distinguish transport, subscription, and data-source failures:
+
+- UTC start time, failure time, and recovery time
+- RTDS host and exact subscription frame (`topic`, `type`, and `filters`)
+- Client package version, Node/browser version, operating system, and WebSocket implementation
+- Last valid payload time per topic—not only the last ping, pong, or metadata frame
+- Socket `readyState`, close code/reason, and error callbacks
+- Sanitized raw frames immediately before and after the failure
+- Message counts and reconnect attempts
+- Whether multiple topics on the same socket stopped together
+- Whether a hard reconnect restored data
+
+When possible, run a raw-frame capture beside the SDK client. This distinguishes an SDK parsing problem from a server-side publishing problem.
+
+### Workaround
+
+Document the smallest mitigation that restored service, including thresholds and possible false positives. A useful workaround should be safe for other builders to copy while the root cause is investigated.
+
+### Requested platform fix
+
+Name the invariant the platform should provide rather than only asking maintainers to "fix the socket."
+
+Examples:
+
+- A live gateway connection must retain valid backend subscription state.
+- If backend subscription state is removed, the gateway should close the socket with an explicit reason.
+- Subscription failures should return stable application-level error codes rather than internal database errors.
+
+## Distinguish transport liveness from data liveness
+
+A WebSocket that is `OPEN` and answering pings is not necessarily delivering subscribed data. Ping/pong may only prove that the gateway and network path are alive.
+
+For production consumers:
+
+1. Track the last **valid data payload** independently from ping/pong traffic.
+2. For naturally quiet subscriptions, add a normally active topic to the same socket as a data-plane heartbeat.
+3. If all expected payloads become stale, hard-close the socket and create a new connection.
+4. Restore subscriptions from `onConnect` after reconnecting.
+5. Add jitter to reconnect delays to avoid synchronized reconnect storms.
+
+Do not use repeated duplicate subscriptions or intentionally triggered server/database errors as a production health check. They add load and couple the client to unstable implementation details.
+
+## Known silent-stall failure mode
+
+Issue [#26](https://github.com/Polymarket/real-time-data-client/issues/26) documents a failure where all topics on a connection can stop while the socket remains open and ping/pong continues. In the reported tests, frozen connections did not recover through re-subscription; a new socket restored delivery.
+
+The current client reconnects after socket close/error events. Applications that require continuous delivery should also implement data-payload staleness detection until the client or server provides an explicit data-plane liveness contract.
+
+## Minimal watchdog pattern
+
+```ts
+const STALE_AFTER_MS = 15_000;
+let lastPayloadAt = Date.now();
+
+const client = new RealTimeDataClient({
+  onConnect: current => {
+    current.subscribe({ subscriptions: activeSubscriptions });
+  },
+  onMessage: (_current, message) => {
+    if (message?.payload) lastPayloadAt = Date.now();
+    handleMessage(message);
+  },
+});
+
+setInterval(() => {
+  if (Date.now() - lastPayloadAt <= STALE_AFTER_MS) return;
+
+  // Replace the socket. Re-subscribing on a stale connection may not restore
+  // backend subscription state.
+  client.disconnect();
+  client.connect();
+  lastPayloadAt = Date.now();
+}, 1_000);
+```
+
+Choose the threshold based on the expected cadence of the heartbeat topic, not on a naturally quiet market topic. Production implementations should also guard against overlapping reconnects and reset/cancel watchdog timers during intentional shutdown.