fix(llm): surface llm-connect's error body instead of a bare 502

llm-connect maps every provider API error onto HTTP 502 and puts the real
cause in the body (llm_connect/server.py::_error_response: error,
provider_status). LLMConnectClient.complete called raise_for_status() and threw
that body away, so a revoked OpenRouter key was indistinguishable from a downed
gateway — four days of production evidence read as "llm-connect is down".

Live check confirms one fault, not two: the llm-connect pod is Running 1/1 with
healthy endpoints, and today's FI/Binky/triage runs still 502 after yesterday's
rollout, matching the sanitized upstream 401 railiance-platform reported.

The client now raises with error, provider_status, provider, model and a
bounded copy of llm-connect's already-sanitized message, under a field
allowlist so no provider blob or key material reaches the run artefact.

Refs ACTIVITY-WP-0031-T01, T03.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

src/activity_core/llm_client.py

@@ -55,7 +55,8 @@ class LLMConnectClient:
             json=payload,
             timeout=self.timeout_seconds,
         )
-        resp.raise_for_status()
+        if resp.status_code >= 400:
+            raise RuntimeError(_llm_connect_error_text(resp))
         data = resp.json()
         self.last_response_metadata = _extract_response_metadata(data)
         content = data.get("content")
@@ -64,6 +65,43 @@ class LLMConnectClient:
         return content
 
 
+_SAFE_ERROR_KEYS = ("error", "provider_status", "provider", "model")
+_MAX_ERROR_MESSAGE_CHARS = 400
+
+
+def _llm_connect_error_text(resp: httpx.Response) -> str:
+    """Describe an llm-connect error using its own diagnostic body.
+
+    llm-connect maps every provider API error onto HTTP 502 and puts the real
+    cause in the body (``error``, ``provider_status``) — see
+    ``llm_connect/server.py::_error_response``. Reporting only the transport
+    status makes a rejected provider key look identical to a dead gateway,
+    which is exactly the ambiguity ACTIVITY-WP-0031-T03 evidence must not have.
+
+    llm-connect sanitizes its own ``message`` before returning it; this keeps a
+    bounded copy of that text plus an allowlist of non-secret fields.
+    """
+    base = f"llm-connect returned HTTP {resp.status_code}"
+    try:
+        body = resp.json()
+    except ValueError:
+        return base
+    if not isinstance(body, dict):
+        return base
+
+    parts = [
+        f"{key}={body[key]}"
+        for key in _SAFE_ERROR_KEYS
+        if body.get(key) not in (None, "")
+    ]
+    message = body.get("message")
+    if isinstance(message, str) and message.strip():
+        parts.append(f"message={message.strip()[:_MAX_ERROR_MESSAGE_CHARS]}")
+    if not parts:
+        return base
+    return f"{base}: " + "; ".join(parts)
+
+
 _SAFE_RESPONSE_METADATA_KEYS = {
     "finish_reason",
     "usage",