docs(tutor-quality): define Stage 2B semantic evaluation

[ttbombadil]

Summary

Adds the normative Stage 2B SSOT for semantic Tutor evaluation and controlled model, prompt, and strategy benchmarking. It defines the semantic rubric, critical-failure handling, versioned Gold Corpus and human references, Judge calibration, blinded absolute/pairwise evaluation, repeated sampling, comparability, reproducibility, and bounded secret-free artifacts.

This PR is documentation and architecture only. It does not implement Stage 2B, change Stage 1 or Stage 2A contracts, alter Tutor runtime behavior, or expand Stage 2C.

Scope and boundaries

• Keeps technical execution, deterministic contract conformity, semantic Tutor quality, and actual learning effect as separate claims.
• Requires Stage 2B to consume immutable Stage 2A transcripts and results rather than duplicating provider transport, Tutor invocation, credential handling, or deterministic checks.
• Defines semantic status, per-Judge calibration status/identity, and per-comparison comparability as separate reporting axes.
• Excludes LLM-as-Judge implementation, semantic scoring code, production model selection, prompt/strategy optimization, and any change to existing hard gates.

Verification

• npm run check:docs — passed (45 Markdown files)
• git diff --check — passed
• Branch diff against main: one added SSOT file only; no implementation changes.