Gate Replicate endpoint promotion on a PR-triggered regression bench

Turns every merged model-update PR into an automatic go/no-go decision. It runs your held-out eval set through the newly pushed Replicate version, compares the scores against the version currently serving production, and only repoints the production endpoint alias when the candidate is at least as good.

Use it when an ML team ships model changes through GitHub and you want a hard quality gate before any new weights reach live inference. It removes the manual "did we check the bench?" step and prevents silent accuracy or latency regressions from reaching users.

1. 1A merged PR labeled `model-release` fires the GitHub trigger with the new version SHA.
2. 2The flow reads the candidate Replicate version reference from the PR body.
3. 3It runs the golden eval set against the candidate version on Replicate and captures per-case scores plus p95 latency.
4. 4A logic step compares candidate accuracy and latency to the live version's recorded baseline against your thresholds.
5. 5If it passes, it updates the production endpoint alias to the candidate version on Replicate.
6. 6It posts the verdict, score delta, and decision to the release Slack channel.