docs: standardize model-judge calibration reporting

[aryeko]

Summary

• add shared manual pointwise model-judge calibration and reporting guidance
• clarify covered/partial/missing/contradicted/unknown interpretation and expected-bad fixture handling
• update eval-kit skills to report deterministic evidence first and keep raw provider bundles local unless curated
• bump package/version references for v0.1.5

Verification

• pnpm install --frozen-lockfile
• pnpm check
• git diff --check
• rg sweeps for stale v0.1.4 current examples, default-config toggling guidance, and calibration wording

Notes

Docs/skills-only release. No runtime behavior, schemas, prompts, fixtures, or consumer semantics changed.