🧪 test(models): pin the OpenRouter catalog fetch tests to repo-controlled ids

[cwest]

Why

cwest/integration was RED on its own: tests/hermes_cli/test_models.py
yielded 2 failed, 85 passed (CI slice 7/8), so every PR targeting the branch
inherited a failing required check regardless of its content.

Both failures were AssertionError: assert 'qwen/qwen3.7-max' in ['anthropic/claude-opus-4.8'].

Root cause

fetch_openrouter_models builds its preferred_ids list from the curated
catalog (get_curated_openrouter_models() remote manifest, falling back to the
in-repo OPENROUTER_MODELS snapshot), then keeps only the ids the live payload
also carries. The two failing tests mocked the live payload but did not pin
the curated source — so preferred_ids came from whatever the ambient remote
manifest currently returns. When the curated list rotated, the mocked live
entries were no longer in preferred_ids and were silently skipped, failing an
assertion with no underlying defect.

This is exactly the snapshot-not-contract class AGENTS.md calls out ("Behavior
contracts over snapshots. Tests should assert how two pieces of data must
relate, not freeze a current value — model lists…"). The sibling test
test_filters_out_models_without_tool_support already got this right by pinning
the curated list; the two failing tests were never updated to match.

The fetch/filter path itself is correct, so no production code changed — the
fix is entirely in the test file.

What

Rewrite both tests as behavior contracts:

• Pin the curated list to repo-controlled placeholder ids
  (test-vendor/...) via monkeypatch.setattr(OPENROUTER_MODELS, …) and patch
  get_curated_openrouter_models to empty, so preferred_ids is deterministic
  and independent of the live catalog / remote manifest.
• Assert the invariant under test rather than a frozen list:
  • test_live_fetch_recomputes_free_tags — tags are recomputed from live
    pricing (zero-priced → free, non-zero → paid), and every covered curated
    id is retained.
  • test_permissive_when_supported_parameters_missing — an entry with no
    supported_parameters is retained, while an explicit list omitting tools
    is dropped (negative control preserves the fix(gateway): hide models without tool support from list Kilo-Org/kilocode#9068 behavior).

The rewritten tests contain no third-party model id whose presence is outside
this repo's control; adding or removing a model from the curated list can no
longer red them.

Verification

• RED-first: the two tests reproduced 2 failed, 85 passed on this branch's
  base (bb87eebfc), matching CI slice 7/8, before the fix.
• scripts/run_tests.sh tests/hermes_cli/test_models.py → 87 passed, 0
  failed after the fix.
• hermes_cli/models.py is untouched (git diff --stat shows only the test
  file), confirming this was the behavior half, not a lookup defect.
• The 7 other tests/hermes_cli failures (WSL gateway, service manager, signal
  handler, qwen-oauth resolution) are pre-existing on the pristine base —
  verified by reverting this fix and re-running those four files (same 7
  failures) — and are unrelated to the OpenRouter catalog / out of this card's
  scope.

[cwest]

No changes needed.

The diagnosis holds up against the code. fetch_openrouter_models builds preferred_ids from get_curated_openrouter_models() (imported lazily at hermes_cli/models.py:1359) and falls back to the in-repo OPENROUTER_MODELS only when that returns empty. The two old tests mocked the live payload but left preferred_ids sourced from the ambient manifest, so once the curated list rotated away from the mocked ids the entries were skipped at models.py:1392 and the assertion compared against whatever the manifest happened to hold. That is the snapshot-freeze the repo guidance rules out.

The rewrite pins both the curated source (get_curated_openrouter_models -> [] and OPENROUTER_MODELS -> a fixed test-vendor/* list) and the live payload, then asserts the relationship rather than a frozen result: zero-priced entries tag free, non-zero stay paid, curated[0] carries recommended, a missing supported_parameters field is retained, and an explicit list omitting tools is dropped. Every assertion maps to a real branch in models.py:1390-1406, and the patch targets are the right ones given the lazy import.

I ran the negative control and the pricing invariant through mutation checks on a throwaway copy of models.py: making the tools filter unconditionally permissive turns test_permissive_when_supported_parameters_missing red, and inverting the free/paid tag turns test_live_fetch_recomputes_free_tags red. Both catch the regression they claim to. tests/hermes_cli/test_models.py is 87 passed locally, and the id rotation this test used to freeze on can no longer red it since the two rewritten tests carry no third-party id whose presence is outside this repo's control. CI is green and the branch is clean.