Complete the human-grounded promotion evidence loop - #3

Outcome

Turns a repeated promotion run into an auditable evidence package that a human decision owner can review.

What changed

• adds prepare-review to generate a blinded, hash-bound A/B review packet from a valid promotion run
• adds finalize-review to validate an independent holdout attestation and exactly two complete human reviews
• reports human agreement, automated-to-human agreement, disagreements with rationales, per-task outcomes, trial variance, regressions, usage, and operator-recorded cost
• retains and hash-lists the manifest, holdout attestation, and both reviewer label files
• documents the holdout custodian, evaluator operator, reviewers, and decision owner handoff
• keeps the automated OpenAI judge explicitly provisional and non-independent
• leaves the promotion decision with an accountable human owner

Scientific boundary

This PR completes the framework workflow. It does not invent a client holdout, fabricate human labels, or claim that any target skill passed promotion. A client-specific promotion result still requires real independently controlled tasks and two real reviewers.

Verification

• 50 unit and end-to-end tests pass
• Ruff passes
• Python compilation passes
• packaged healthcheck passes
• standalone copied-skill healthcheck passes
• diff check passes