Reliability is a product feature. For every meaningful release, we follow a repeatable loop:

1. Validate local behavior in a clean worktree.
2. Run the fast verification suite.
3. Run required e2e checks for critical flows.
4. Publish only after deterministic checks pass.

When something fails, we prefer actionable failures over silent fallback behavior. That makes issues visible quickly and prevents partial rollouts from masking real problems.