fix(tests): runtime_provider no longer permanently captures a mocked load_config

Test-pollution class: runtime_provider is usually imported lazily (inside
switch_model's resolution path), so its first import in a pytest worker can
happen while a test has hermes_cli.config.load_config patched. The
module-level from-import then bound the MagicMock permanently — after the
patch exited, every later caller in the process silently read the dead
test's config. Live victim: MoA aggregator context-length resolution
(resolve_runtime_provider -> AuthError 'Unknown provider custom:example'),
making TestMoAContextLength::test_moa_custom_context_configures_compressor_threshold
fail whenever it shared a process with
TestLocalOllamaModelDiscovery::test_switch_model_on_current_ollama_custom_endpoint_keeps_base_url.

Fix: load_config / get_compatible_custom_providers / normalize_extra_headers
become late-bound delegates resolving hermes_cli.config attributes at call
time. Both patch targets (config.load_config and
runtime_provider.load_config) keep working. Regression tests pin the
late-binding property and fail if the delegates revert to from-imports
(sabotage-verified).

hermes_cli/runtime_provider.py

@@ -40,14 +40,36 @@ from hermes_cli.auth import (
     is_actual_local_base_url,
     normalize_actual_base_url,
 )
-from hermes_cli.config import (
-    get_compatible_custom_providers,
-    load_config,
-    normalize_extra_headers,
-)
+from hermes_cli import config as _config_mod
 from hermes_cli.providers import custom_provider_aliases, custom_provider_slug
 from hermes_constants import OPENROUTER_BASE_URL
 from hermes_cli.providers import is_official_openai_host
+
+
+def load_config():
+    """Late-bound delegate to :func:`hermes_cli.config.load_config`.
+
+    Deliberately NOT a module-level ``from hermes_cli.config import
+    load_config``: this module is often imported lazily (inside functions),
+    so its first import can happen while a test has
+    ``hermes_cli.config.load_config`` patched — a from-import would then
+    bind the MagicMock *permanently*, poisoning every later caller in the
+    process (the mock's fixed config shadows the real one long after the
+    patch exits). Delegating at call time keeps both patch targets working:
+    patching ``hermes_cli.config.load_config`` OR
+    ``hermes_cli.runtime_provider.load_config`` behaves as expected.
+    """
+    return _config_mod.load_config()
+
+
+def get_compatible_custom_providers(config=None):
+    """Late-bound delegate — see :func:`load_config` for why."""
+    return _config_mod.get_compatible_custom_providers(config)
+
+
+def normalize_extra_headers(value):
+    """Late-bound delegate — see :func:`load_config` for why."""
+    return _config_mod.normalize_extra_headers(value)
 from utils import base_url_host_matches, base_url_hostname, env_int
 
 

tests/hermes_cli/test_runtime_provider_late_binding.py

@@ -0,0 +1,63 @@
+"""Regression: runtime_provider's config helpers must be late-bound.
+
+Test-pollution class found Aug 2026: ``hermes_cli.runtime_provider`` is
+frequently imported lazily (inside functions like ``switch_model``'s
+resolution path), so its first import in a process can happen while a test
+has ``hermes_cli.config.load_config`` patched. With a module-level
+from-import, that bound the MagicMock into ``runtime_provider.load_config``
+for the life of the process: after the patch exited, every later caller
+(e.g. MoA aggregator context-length resolution reading ``providers:``)
+silently got the long-dead test's config. Live pair that exposed it:
+
+  tests/hermes_cli/test_models.py::TestLocalOllamaModelDiscovery::
+      test_switch_model_on_current_ollama_custom_endpoint_keeps_base_url
+  tests/agent/test_model_metadata.py::TestMoAContextLength::
+      test_moa_custom_context_configures_compressor_threshold
+
+The fix: runtime_provider exposes late-bound delegates that resolve
+``hermes_cli.config`` attributes at call time. These tests pin exactly
+that property — a patch on ``hermes_cli.config.<fn>`` must be visible
+through ``runtime_provider.<fn>`` while active, and must fully release
+when it exits (no permanent capture). They fail if anyone reverts the
+delegates back to module-level from-imports.
+
+NOTE: deliberately no ``sys.modules`` pop/re-import here — a fresh
+re-import under an active patch is itself a polluter (duplicate module
+objects with diverging state).
+"""
+
+from unittest.mock import patch
+
+
+def test_load_config_patch_is_visible_and_releases():
+    import hermes_cli.runtime_provider as rp
+
+    sentinel = {"providers": {"only-the-mock-has-this": {"api": "http://x/v1"}}}
+    with patch("hermes_cli.config.load_config", return_value=sentinel):
+        # While the patch is active, the delegate must see the mock…
+        assert rp.load_config() is sentinel
+
+    # …and once it exits, the same module must resolve the real function
+    # again instead of replaying the mock's config (the permanent-capture
+    # failure mode of a from-import binding).
+    result = rp.load_config()
+    assert result is not sentinel
+    assert "only-the-mock-has-this" not in (result.get("providers") or {})
+
+
+def test_sibling_delegates_are_late_bound_too():
+    """get_compatible_custom_providers / normalize_extra_headers share the
+    same import shape and must share the same late-binding behavior."""
+    import hermes_cli.runtime_provider as rp
+
+    marker = [{"name": "marker", "base_url": "http://m/v1"}]
+    with patch(
+        "hermes_cli.config.get_compatible_custom_providers", return_value=marker
+    ):
+        assert rp.get_compatible_custom_providers({}) is marker
+    assert rp.get_compatible_custom_providers({}) is not marker
+
+    hdrs = {"X-Marker": "1"}
+    with patch("hermes_cli.config.normalize_extra_headers", return_value=hdrs):
+        assert rp.normalize_extra_headers(None) is hdrs
+    assert rp.normalize_extra_headers(None) is not hdrs