test(tutor-quality): run Tutor Quality scenarios with the real planner only (TutorQuality §5)

[ttbombadil]

Purpose

Conformance finding from the Tutor SSOT audit. TutorQuality SSOT §5: "The runner MUST use TutorService and, where Course Content is supplied, CurriculumTutorAdapter. It MUST fake only the external provider." The Stage-1 plan says the same ("with only the provider faked").

Finding

Since the runner was created (33b87cc9), it accepted a planning override (scenario.planning ?? new CurriculumTutorAdapter()). Five scenarios used it to script planning outcomes:

Scenario	Faked
retains a repaired transition/blocked × exact/near duplicate question (4)	planFollowup returning hand-written transition/blocked metadata
lets an application-owned TutorPlan replace provider question and metadata	planInitial returning a hand-written plan
TQ-REG-001 › repairs repetition … coherent transition metadata	the whole planner while Course Content was supplied, mutating state to an invented pwm-output Topic

So §3 items 6/7 and the TQ-REG-001 clause "result/state metadata must remain coherent through transition or blocked planning outcomes" were checked against invented outcomes, not the real planner.

Change (tests only)

• The runner has no planning override. Course Content is always planned by the real adapter; without Course Content no planner exists.
• The same outcomes now come from Course Content states:
  • transition: a strong answer completes LEARN mastery of a Topic that has no further LEARN question;
  • blocked: a DEEPEN answer with no further post-mastery question ends in content-exhausted;
  • plan authority: the real initial plan replaces the provider's question and forged metadata;
  • TQ-REG-001: a minimal PWM test Topic (activates on type-used: int, the only fact the extractor yields for the PWM sketch) is mastered by one strong answer. The real transition serves the turn under LEARN (R-PH-1), and the state commits DEEPEN on pwm-output; the near-repeat is still repaired. The old fake claimed learningPhase: DEEPEN on the served turn, which the real planner does not produce.
• No product code changes. No Course Content outside test fixtures.

Evidence

• RED: removing the override makes the three faking call sites fail to type-check. They are the only uses.
• Mutation: when the repaired question is dropped for transition/blocked outcomes in TutorService.applyDialogFollowUp, all five rewritten scenarios fail; reverted.
• npm run check 0 errors, npm run test:tutor-quality 407 passed, ESLint clean, pre-push (unit + Sonar) passed.

🤖 Generated with Claude Code