Add a committed regression test for the model-fallback path

Context

The model-fallback loop (outer model list × inner retry) in .github/workflows/gemini-reviewer.yml was validated only with an ad-hoc mock server during PR #7 (a model returning 404 fell through to the next, which returned 200). There is no committed test, so future changes (e.g. the boilerplate-refactor in #8, or a new provider) can silently break fallback without anyone noticing.

Goal

Add a regression test that spins up a mock OpenAI-compatible server and asserts:

1. Fallback works — a removed model (HTTP 400/404/422) is skipped and the next model is used; the comment heading names the model that actually reviewed.
2. All-fail path — when every model fails, an All models failed comment is posted (not a hang/crash).
3. Fail-fast on auth — a 401/403 (or permanent 429) aborts instead of burning the list.

Approach

A small script (bash or node) that starts the mock server, exports PROVIDER=openrouter MODELS="removed-model good-model" plus a dummy key/endpoint, sources the relevant part of the workflow's run step, and checks review_report.md. Run it in CI on changes to the workflow, or as a manual/local check.

Acceptance

• Test reproduces the 3 scenarios above.
• It fails if fallback regresses.
• Documented in AGENTS.md/README under testing.

Start with the run step in .github/workflows/gemini-reviewer.yml and inspect how it produces review_report.md. Build the mock-server check for fallback, all-fail, and auth fail-fast cases, then run it locally or in CI. Done means all three scenarios fail on regression and the test is documented in AGENTS.md or README.