DeepSeek docs: Add real QVAC validation checklist

Approx tokens/sec: 38.08

@@ -0,0 +1,55 @@
+
+### 2.4 Failure Triage
+
+When a real‑QVAC run exits with a non‑zero code or unexpected output, use the
+following steps to investigate.
+
+1. **Validate connectivity** – Verify the endpoint URL with `curl -v $QVAC_URL`.
+   If the server is reachable, a `200 OK` response should be returned.
+
+2. **Confirm the model name** – Ensure the value passed to `--model` matches
+   a model that the endpoint actually supports. Check the provider’s
+   documentation or query the `/models` endpoint if available.
+
+3. **Verify the API key** – Double‑check that the key stored in
+   `QVAC_API_KEY` (or `OPENAI_API_KEY`) is correct and has not expired.
+   Rotate keys if necessary.
+
+4. **Check for rate limits** – Repeated runs may trigger `429 Too Many Requests`
+   responses. Wait a short time and retry, or ask your provider to increase
+   the quota.
+
+5. **Inspect sanitised output** – The CLI prints error details to stderr.
+   Sensitive tokens are redacted, but the remaining information usually
+   contains the HTTP status code and a human‑readable reason.
+   ```bash
+   node dist/cli.js --url ... --model ... 2> error.log
+   ```
+
+6. **Adjust timeouts** – If the error code is `timeout`, increase the request
+   timeout with `--timeout-ms`. This is especially helpful on high‑latency
+   connections or for slow models.
+
+7. **Interpret common error codes**
+   - `server_unavailable` – The upstream host could not be reached. Check
+     network access and firewall rules.
+   - `http_error` – A non‑2xx HTTP status was returned. Look at the sanitised
+     message for the code (e.g., 401/403 for auth problems, 5xx for
+     server‑side issues).
+   - `malformed_stream` – The server sent an unexpected SSE payload. Use
+     `curl` to inspect the raw JSON lines and confirm the server is
+     compatible with the OpenAI‑style streaming API.
+   - `timeout` – The request took longer than the configured timeout. Try
+     relaxing the limit or checking network latency.
+
+8. **Test with minimal parameters** – Use a short prompt (`"Hello"`) and a
+   small `--max-tokens` value to isolate whether the issue is specific to a
+   longer request.
+
+9. **Collect logs for further investigation** – Save both stdout and stderr
+   (e.g., `node dist/cli.js ... > out.txt 2> err.txt`). Share the sanitised
+   output when seeking help – **never** share raw API keys.
+
+10. **Compare baseline** – As a last resort, use `curl` to send a simple
+    request to the same endpoint. If the raw `curl` also fails, the problem
+    lies with the server or network, not with `qvac-bench`.

docs/validation-checklist.md

@@ -90,3 +90,4 @@ pipeline that does not have access to the real endpoint.
 7. If the run exits with a non‑zero code or produces unexpected output, refer to the [Failure Triage](#24-failure-triage) section for debugging guidance.
 
 **Example expected stdout (single iteration)**
+