feat: add Tutor Quality Stage 2A real-provider evaluation foundation

[ttbombadil]

Scope

• Real-Provider Evaluation Foundation using the normal TutorService/LLMProvider path
• Versioned Anchor Corpus with 10 scenarios, including TQ-REG-001
• Separate executionStatus and invariantViolations axes
• Raw capture of the parsed provider result before TutorService validation/repair
• Secret-free allow-list transcript and report artifacts
• evaluationIdentity covering Git, Course, Corpus, Provider, Model, Prompt revision and digest metadata
• Fixed model identifier; auto is rejected
• Provider-call budget counting listModels and generation calls
• Prompt revision and digest metadata
• Manual workflow_dispatch only; no required CI dependency

Stage 1 safety

• Existing deterministic Tutor Quality hard gates remain unchanged
• Stage 2A is not a pull-request hard-gate dependency
• No direct provider HTTP calls
• Existing Tutor and provider trust boundaries remain authoritative
• Shared diagnostic validation is used instead of duplicated rules

Explicitly not included

• LLM-as-Judge
• Semantic learning-support scoring
• Learning-effect measurement
• Adaptive Tutor strategies
• Fact Extractor expansion
• Learner profiles

Verification

All local verification was run under Node 24.20.0:

• npm run test:tutor-quality: 29 files, 216 tests passed
• npm run test:unit: 2477 tests passed, 1 existing skip
• npm run check passed
• npm run check:docs passed
• git diff --check passed
• Fake-Provider Anchor Run over all 10 scenarios: 0 technical failures, 0 invariant violations
• No-credential CLI: not-run / missing-credential / 0 provider calls / report.json only
• Sonar Quality Gate passed: 0 new violations, 100% reviewed security hotspots, 82.5% new-code coverage, no open issues

Real Provider Smoke Test

• Provider: KI:connect
• Model: mistralai-mistral-small-4-119b-2603
• Scenario: TQ-REG-001
• Status: completed
• Returned model matched the requested model
• No technical failures
• invariantViolations = []
• Secret scan passed
• Credential removed after the run
• No secrets or evaluation artifacts committed

No credential value or token fragment is included in this description.