fix(tests): failure-writer ownership probe no longer races a real tirith download

The probe scrubs the environment down to PATH/LANG/LC_ALL/TZ, which also
dropped the harness's HERMES_DISABLE_LAZY_INSTALLS. In a fresh home with no
tirith on PATH (CI), GatewayRunner() then starts tirith's startup install on a
background thread: the PM runtime downloads and unpacks Python and uv from
GitHub while the cases run. When that unpacking overlaps the archive case, its
buffers land in the process-wide tracemalloc window and push the peak past the
8 MiB bound, so the probe reports 19/20.

The loopback fence also missed it: hermes_bootstrap's happy-eyeballs
create_connection dials with connect_ex, so the downloads reached GitHub with
blocked_external_attempts empty.

- set HERMES_DISABLE_LAZY_INSTALLS=1 in the probe environment
- fence socket.connect_ex as well as connect
- assert the receipt records no external network attempts

evals/gateway_failure_ownership/README.md

@@ -11,8 +11,10 @@ scripts/run_tests.sh -j 1 tests/gateway/test_failure_writer_ownership.py
 
 The first argument selects the production checkout, so the same fixture can A/B
 another checkout without editing it. The receipt records its temporary HERMES_HOME.
-Inherited credentials are cleared, tools are disabled, and socket connections are
-restricted to loopback. No live messaging account or paid provider is contacted.
+Inherited credentials are cleared, tools are disabled, lazy installs are off (so
+GatewayRunner does not start tirith's background download), and socket connections
+(`connect` and `connect_ex`) are restricted to loopback. No live messaging account or
+paid provider is contacted.
 
 ## Fault boundaries and controls
 

evals/gateway_failure_ownership/probe.py

@@ -1,4 +1,4 @@
-import os, sys, json, asyncio, threading, tempfile, sqlite3, socket, subprocess, tracemalloc
+import os, sys, json, asyncio, threading, tempfile, sqlite3, socket, subprocess, tracemalloc, errno
 from pathlib import Path
 from http.server import ThreadingHTTPServer, BaseHTTPRequestHandler
 from types import SimpleNamespace
@@ -14,6 +14,11 @@ os.environ.update(
     HOME=str(HOME),
     HERMES_HOME=str(HOME),
     HERMES_DISABLE_PLUGINS="1",
+    # The scrub above drops the test harness's hermetic switch. Without it, GatewayRunner()
+    # in this fresh home starts tirith's startup install on a background thread (PM runtime,
+    # Python and uv downloaded and unpacked while the cases run) and its buffers land in the
+    # process-wide tracemalloc window the archive case measures.
+    HERMES_DISABLE_LAZY_INSTALLS="1",
     NO_PROXY="127.0.0.1,localhost",
 )
 sys.path.insert(0, str(ROOT))
@@ -21,23 +26,33 @@ os.chdir(HOME)
 (HOME / "config.yaml").write_text(
     "model:\n  provider: openai-compat\n  default: fixture-model\n  context_length: 131072\nagent:\n  max_iterations: 2\ncompression:\n  enabled: false\ndatabase:\n  journal_mode: delete\n"
 )
-# Fence all network calls to loopback, including optional discovery/aux paths.
+# Fence all network calls to loopback, including optional discovery/aux paths. Both entry
+# points: hermes_bootstrap's happy-eyeballs socket.create_connection dials with connect_ex.
 orig_connect = socket.socket.connect
+orig_connect_ex = socket.socket.connect_ex
 blocked = []
 
 
+def _external(address):
+    return isinstance(address, tuple) and address[0] not in ("127.0.0.1", "::1", "localhost")
+
+
 def local_connect(self, address):
-    if isinstance(address, tuple) and address[0] not in (
-        "127.0.0.1",
-        "::1",
-        "localhost",
-    ):
+    if _external(address):
         blocked.append(str(address))
         raise OSError("fixture forbids external network")
     return orig_connect(self, address)
 
 
+def local_connect_ex(self, address):
+    if _external(address):
+        blocked.append(str(address))
+        return errno.ENETUNREACH
+    return orig_connect_ex(self, address)
+
+
 socket.socket.connect = local_connect
+socket.socket.connect_ex = local_connect_ex
 requests = []
 
 

tests/gateway/test_failure_writer_ownership.py

@@ -26,6 +26,9 @@ def test_gateway_failure_writer_preserves_accepted_turn_identity(tmp_path):
     assert receipt.exists(), result.stdout + result.stderr
     data = json.loads(receipt.read_text())
     assert all(row["reached"] for row in data["observations"]), data
+    # A startup lazy install (tirith → PM runtime) would download over the real network on a
+    # background thread and pollute the archive case's process-wide tracemalloc window.
+    assert data["blocked_external_attempts"] == [], data["blocked_external_attempts"]
     assert data["passed"] == data["total"], data["observations"]
     assert result.returncode == 0, result.stdout + result.stderr