Model IDs were hardcoded across the e2e suite, so every provider
deprecation meant a scatter-shot edit and stale models (gemini-2.5-flash,
gpt-4o-mini) kept running in CI. Tests now import role-named pins from
tests/e2e/model_matrix.py, and a bump is a one-file change plus the
compose gateway config.
check_e2e_model_freshness.py (wired into test-code-quality) fails when a
pin is missing from model_prices_and_context_window.json, is within 30
days of its deprecation_date, drifts from the docker-compose gateway
config, or when a test hardcodes a versioned model literal again.
Bumps: gemini-2.5-flash -> gemini-3.5-flash, gpt-4o-mini -> gpt-5.4-mini
