test(gateway): make the fallback-chain tests independent of host config

[OmarB97]

What does this PR do?

Removes a latent host-config dependency in TestResolveRuntimeWithFallback (5 tests in tests/test_tui_gateway_server.py).

The tests patch server._load_fallback_model, but _resolve_runtime_with_fallback() consults _load_effective_fallback_model() — the policy-filtered chain added by #269. That wrapper only reaches the raw loader when the ambient policy is not off:

def _load_effective_fallback_model():
    cfg = _load_cfg()
    policy = get_fallback_policy(cfg)
    if policy == "off":
        return []
    chain = _load_fallback_model() or []

This is not a CI failure, and I want to be precise about that. CI has no hermes config, so get_fallback_policy({}) returns the "any" default, the wrapper falls through to the patched raw loader, and all five tests pass. They are green on main today.

On a host whose config.yaml sets fallback_policy: off, the wrapper short-circuits to [], the stub is never consulted, no fallback is ever attempted, and all five fail — asserting nothing about the code they exist to cover. The result is a suite that behaves differently for a contributor than it does in CI, and five tests that are silently inert for everyone else (they pass through a stub that is only reachable by accident of the default policy).

The fix is to patch what the code actually calls. test_auth_error_all_fallbacks_fail_raises additionally pins _load_cfg, because it asserts the policy-specific exhaustion wording — "no usable configured backup route remained" is the any branch — and would otherwise assert against whatever policy the host happens to have.

Scope note: this file has two other failures on main (the golden turn-isolation transcript and the session-activate double). Those are the CI-blocking ones and they are owned by #296, which is already green. This change is disjoint from that PR — different tests, no overlapping lines — so the two compose. Until #296 lands, Python tests on this PR will still show those two failures; they are pre-existing on main and untouched here.

Related Issue

No filed issue — found while driving main's CI back to green after the stale-lockfile fix (#286) unmasked the Python suite.

Type of Change

• ✅ Tests (adding or improving test coverage)

Changes Made

All in tests/test_tui_gateway_server.py, class TestResolveRuntimeWithFallback:

• Five monkeypatch.setattr targets changed from _load_fallback_model to _load_effective_fallback_model — the function _resolve_runtime_with_fallback() actually calls.
• test_auth_error_all_fallbacks_fail_raises pins _load_cfg to {"fallback_policy": "any"} so its assertion on the exhaustion wording is deterministic.
• A comment at the first patch site explaining the wrapper/policy interaction, so the next person does not re-point it at the raw loader.

How to Test

uv sync --locked --python 3.11 --extra all --extra dev

Green either way after this change; before it, the first command produced 5 failures and the second 0:

HERMES_HOME=/path/to/a/home/whose/config/sets/fallback_policy_off \
  .venv/bin/python -m pytest "tests/test_tui_gateway_server.py::TestResolveRuntimeWithFallback" -q

env -u HERMES_HOME .venv/bin/python -m pytest "tests/test_tui_gateway_server.py::TestResolveRuntimeWithFallback" -q

Both now report 8 passed. Whole-file run in a CI-like environment: 382 passed, 2 failed — the two failures being the pre-existing golden-transcript and session-activate cases owned by #296.

Checklist

Code

• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate — fix(deps): uv.lock check is red on every PR — main's lockfile has been stale since #262 #286–feat(desktop): local spawn API — start app-owned chats from the CLI, streamed live #298 checked; test(gateway): update turn-isolation parity and activate doubles for #270 #296 covers this file's other two tests and does not touch these five
• My PR contains only changes related to this fix (no unrelated commits)
• I've run the affected tests and they pass in both environments described above
• I've added tests for my changes — N/A, this repairs the isolation of existing tests; no behavior change
• I've tested on my platform: macOS 15 (Darwin 25.6.0), Apple Silicon, Python 3.11.15

Documentation & Housekeeping

• I've updated relevant documentation — N/A, test-only change
• I've updated cli-config.yaml.example — N/A
• I've updated CONTRIBUTING.md / AGENTS.md — N/A
• I've considered cross-platform impact — N/A, no platform-specific code; this removes an environment dependency
• I've updated tool descriptions/schemas — N/A

Risks / gaps

• ruff check tests/test_tui_gateway_server.py is clean.
• Patching _load_effective_fallback_model means these tests no longer exercise the policy gate itself. That is the correct split — they are about walking the chain on AuthError, and the gate has its own coverage — but it does mean a regression in the off/local-only filtering would not be caught here.
• I did not touch this file's other two failing tests, deliberately, to avoid conflicting with test(gateway): update turn-isolation parity and activate doubles for #270 #296. If test(gateway): update turn-isolation parity and activate doubles for #270 #296 is abandoned rather than merged, those two still need fixing.