docs(tutor-quality): plan stage 2b implementation

[ttbombadil]

Summary

Adds the non-normative implementation plan for Tutor Quality Stage 2B, derived from the merged Stage-2B SSOT.

The plan splits implementation into six independently reviewable PRs and prioritizes an early absolute semantic-evaluation vertical slice before multi-model benchmarking, pairwise evaluation, or ranking.

Architecture

• Keeps the Semantic Corpus Stage-2B-owned.
• Uses a narrow Stage-2A Case Adapter as the primary Tutor execution path.
• Keeps runTutorQualityEvaluation and the Stage-2A execution contract unchanged.
• Supports compatible existing Stage-2A transcripts as a secondary ingestion path.
• Keeps candidate-specific CaseExposureRecords separate from Semantic Corpus versioning.
• Keeps the existing Tutor LLMProvider contract unchanged and plans a separate structured-output capability for Judge calls.
• Computes Stage-2B transcript digests without modifying Stage-2A artifacts.

Delivery sequence

1. Semantic contracts, Corpus, and Stage-2A adapter
2. Absolute Semantic Judge foundation
3. Human Reference and calibration evidence
4. Repeated multi-model benchmark and raw reporting
5. Blinded pairwise evaluation
6. Conditional reporting/ranking

FIRST_USEFUL_STAGE_2B requires PR 2 plus DG-1 and DG-10 for the bounded real-Judge smoke.

MODEL_BENCHMARK_MVP is targeted after PR 4 once sufficient complete reviewed cases exist.

Decision gates

The plan deliberately leaves unresolved normative decisions explicit rather than encoding them as implementation defaults, including rubric scale, calibration thresholds, statistics, held-out governance, ranking, cost normalization, pairwise protocol, and operational limits.

Scope

Documentation/planning only.

• No Stage-2B implementation.
• No Tutor runtime change.
• No Stage-1 or Stage-2A contract change.
• No provider calls.
• No production model selection.
• No Stage 2C work.

Verification

• npm run check:docs (Node 24.20.0)
• git diff main...HEAD --check
• Only docs/plan-tutor-quality-stage-2b.md changed.