fix(arxiv): remove a rejected failed-conversion page before falling back

src/gpd/mcp/servers/arxiv_bridge.py

@@ -689,7 +689,15 @@ async def _intercept_download(
             else:
                 if _arxiv_ar5iv.is_conversion_failure(content):
                     # An older bridge cached ar5iv's failed-conversion page.
-                    logger.info("ignoring cached conversion failure for %s", paper_id)
+                    # Remove it so no fallback, including the upstream
+                    # server's own cache check, can serve it again.
+                    try:
+                        cache_path.unlink()
+                    except OSError as exc:
+                        return _tool_error(
+                            f"Cached failed-conversion page for {paper_id} could not be removed: {exc}"
+                        )
+                    logger.info("removed cached conversion failure for %s", paper_id)
                 else:
                     return _content_envelope(
                         "cache",

tests/mcp/test_arxiv_bridge.py

@@ -1691,3 +1691,38 @@ def test_topic_match_rules(query, paper, expected) -> None:
     from gpd.mcp.servers.arxiv_bridge import _topic_match
 
     assert _topic_match(query, paper) == expected
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("tool", ["download_paper", "read_paper"])
+async def test_rejected_cached_conversion_failure_is_removed_before_upstream_fallback(
+    monkeypatch: pytest.MonkeyPatch, tmp_path, tool
+) -> None:
+    """When ar5iv and the PDF mirror both miss, the call falls through to the
+    upstream server, which serves any cached Markdown file as it is; the
+    rejected failure page must be gone by then."""
+    from gpd.mcp.servers import _arxiv_ar5iv, _arxiv_gcs, _arxiv_token_bucket
+    from gpd.mcp.servers.arxiv_bridge import ArxivBridge, ArxivBridgeConfig
+
+    _arxiv_token_bucket._reset_for_tests()
+
+    async def no_sleep(_seconds: float) -> None:
+        return None
+
+    monkeypatch.setattr(_arxiv_token_bucket.asyncio, "sleep", no_sleep)
+    monkeypatch.setattr(_arxiv_ar5iv, "fetch_html_content", lambda pid: None)
+    monkeypatch.setattr(_arxiv_gcs, "fetch_pdf_from_gcs", lambda pid: None)
+    stub = tmp_path / "hep-th_9711200.md"
+    stub.write_text("No content available\nConversion to HTML had a Fatal error and exited abruptly.", encoding="utf-8")
+
+    fake, log = _make_fake_session(call_outputs=[("upstream answer", False)])
+    bridge = ArxivBridge(ArxivBridgeConfig(storage_path=tmp_path, backend="hybrid"))
+    bridge._session = fake  # type: ignore[assignment]
+    try:
+        result = await bridge.call_tool(tool, {"paper_id": "hep-th/9711200"})
+    finally:
+        bridge._session = None
+
+    assert not stub.exists()
+    assert [call[0] for call in log] == [tool]
+    assert "Fatal error" not in result.content[0].text