Create evaluation and calibration harness for decision packs

[Sunny-commit]

Goal

Make decision-pack quality measurable and claims reproducible.

Scope

• Labeled fixtures and deterministic evaluation runner.
• Accuracy, confusion counts, per-class metrics, Brier score, and calibration/ECE where probabilities are available.
• Separate mock-provider tests from optional live-Laya evaluation.
• Machine-readable and human-readable reports.

Acceptance criteria

• Dev and held-out fixtures are clearly separated.
• Reports include model/checkpoint, configuration, sample size, hardware, and methodology.
• Loading, routing, network, and inference latency are distinguished.
• Negative results and known limitations are retained.
• CI prevents unsupported benchmark claims from being presented as verified.

Dependency

Decision-pack pull requests should add fixtures compatible with this harness.