test: failing repro for vocab quiz scoring bugs (#189, #191)

[davidortinau]

Stream B Step 1 of the Vocabulary Quiz bug cluster — failing-first regression tests for #189 and #191. No production code changes. Author: Jayne (Tester).

These are repro/regression tests that lock down the expected behavior so Wash (backend) and Kaylee (UI) have unambiguous targets. Wash's fix should turn the failing test green; Kaylee's UI fix is informed by the diagnostic that the service-side tests already pass.

What was added

tests/SentenceStudio.UnitTests/Integration/VocabQuizScoringRepro189And191Tests.cs — 4 tests using the existing PlanGenerationTestFixture (real EF Core + in-memory SQLite + DI), modeled on MasteryAlgorithmIntegrationTests.

Results on main (commit 2aab53d)

Total tests: 4
     Passed: 3
     Failed: 1

Test	#189/#191	Result on main	Meaning
Repro189_SingleCorrectRecognitionAttempt_ProducesExpectedPanelState	#189	✅ PASS	Service correctly records 1 attempt, 100% accuracy, no production-side bumps for a single correct MC turn
Repro189_SingleCorrectRecognition_LegacyProductionFieldsRemainZero	#189	✅ PASS	Obsolete ProductionAttempts/ProductionCorrect stay zero for a recognition turn
Repro191_NewWord_AllCorrect_DoesNotRotateOutBeforeFifthTurn	#191	❌ FAIL	Fresh word with 4 all-correct turns trips ReadyToRotateOut=True at turn 4
Repro191_CharacterizeCurrentBehavior_FreshWordRotatesAtTurnN	#191	✅ PASS (snapshot)	Documents that the first rotation turn today is 4

#189 — disambiguated

Two competing hypotheses going in:

• (a) VocabularyProgressService double-increments on a single attempt.
• (b) The Learning Details panel reads legacy/obsolete fields that don't match the new streak-based truth, or there's a duplicate UI call path.

Both #189 service-side tests PASS on main. Service math is correct:

VocabularyProgress dump (after one correct MultipleChoice attempt):
  TotalAttempts=1, CorrectAttempts=1, Accuracy=1.000
  CurrentStreak=1.00, ProductionInStreak=0, MasteryScore=0.143
  RecognitionAttempts=1, RecognitionCorrect=1
  ProductionAttempts=0, ProductionCorrect=0

→ Hypothesis (b) stands. The "2 production attempts / 50% accuracy" panel readout has to come from the UI layer, not the service. The most likely culprits are:

1. The Learning Details panel in VocabQuiz.razor reading obsolete legacy fields (e.g., ProductionAttempts directly) instead of the new streak-based fields, or
2. A double-call path through RecordPendingAttemptAsync (called from NextItem, OverrideAsCorrect, and one other site — possible duplicate-fire vector).

Both are UI/quiz-page concerns belonging to Kaylee's Stream A. The two passing service tests stay as regression guards so the service contract can't silently regress while the UI fix is in flight.

#191 — confirmed

Captured trace from the failing test (one fresh word, all answers correct, mode chosen the same way VocabQuiz.razor chooses it — MC until CurrentStreak>=3 OR MasteryScore>=0.5, then Text):

turn=1 mode=MultipleChoice streak=1.00 prodInStreak=0 mastery=0.143 sessMC=1 sessText=0 ReadyToRotateOut=False
turn=2 mode=MultipleChoice streak=2.00 prodInStreak=0 mastery=0.286 sessMC=2 sessText=0 ReadyToRotateOut=False
turn=3 mode=MultipleChoice streak=3.00 prodInStreak=0 mastery=0.429 sessMC=3 sessText=0 ReadyToRotateOut=False
turn=4 mode=Text          streak=4.50 prodInStreak=1 mastery=0.714 sessMC=3 sessText=1 ReadyToRotateOut=True   ← rotation flips here
turn=5 mode=Text          streak=6.00 prodInStreak=2 mastery=1.000 sessMC=3 sessText=2 ReadyToRotateOut=True
turn=6 mode=Text          streak=7.50 prodInStreak=3 mastery=1.000 sessMC=3 sessText=3 ReadyToRotateOut=True

Failure message:

Expected firstRotateTurn to be greater than or equal to 5 because a brand-new word with 4 all-correct turns demonstrates too little mastery to rotate out — current Tier 2 logic flips this at turn 4 (3 MC + 1 Text), which is the rapid-empty behavior #191 describes, but found 4.

Root cause is in VocabularyQuizItem.ReadyToRotateOut Tier 2 (lines 33–55 of VocabularyQuizItem.cs): once MasteryScore >= 0.50 OR CurrentStreak >= 3, the only additional gates are SessionCorrectCount>=2 AND SessionTextCorrect>=1. With the quiz's mode auto-flip kicking Text in at turn 4, those gates are met immediately. That matches Captain's report of 26 fresh words mastered in 58 turns over 8 rounds (~2.2 turns/word).

How Wash should use this

1. Read the failing trace above.
2. Tighten Tier 2 in VocabularyQuizItem.ReadyToRotateOut (and/or the mode-flip threshold). Suggested rough targets: more required Text-correct turns, higher mastery floor, or per-word session minimums independent of the global session counters. Don't pick the curve unilaterally — discuss with Captain via decisions.md first.
3. Run dotnet test --filter VocabQuizScoringRepro189And191Tests. Both Repro191_* tests will need updating after the fix:
   • Repro191_NewWord_AllCorrect_DoesNotRotateOutBeforeFifthTurn should pass.
   • Repro191_CharacterizeCurrentBehavior_* should be updated to reflect the new first-rotation turn (or removed once the curve is canonical).

How Kaylee should use this

1. The two Repro189_* tests prove the service is fine. Don't touch VocabularyProgressService.RecordAttemptAsync for Accurate and total attempt don't make sense #189.
2. Audit the Learning Details panel in VocabQuiz.razor (lines ~395–460) for any reads of legacy obsolete fields — replace with streak-based equivalents.
3. Audit the call sites of RecordPendingAttemptAsync for duplicate-fire (NextItem ~1245, OverrideAsCorrect ~1394, plus the third site near 1490).

Out of scope for this PR

• Production code (no edits to VocabularyProgressService, VocabularyQuizItem, VocabQuiz.razor, etc.).
• The actual fix — that's Wash's Stream B Step 2 + Kaylee's Stream A.
• Mode-selection changes — the test's ChooseQuizModeForTurn mirrors the current VocabQuiz.razor rule verbatim; if the rule moves, the helper moves with it.

Verification

$ dotnet build tests/SentenceStudio.UnitTests/SentenceStudio.UnitTests.csproj
Build succeeded. 0 Error(s)

$ dotnet test ... --filter "FullyQualifiedName~VocabQuizScoringRepro189And191Tests"
Total tests: 4 | Passed: 3 | Failed: 1

Branch: test/vocab-quiz-scoring-repro-189-191, off main (2aab53d). No conflicts with Kaylee's fix/vocab-quiz-ui-cluster-189-194.

[davidortinau]

Superseded by PR #198 (squash-merged to main). Jayne's repro tests landed verbatim as part of #198's atomic fix-plus-tests commit. Branch test/vocab-quiz-scoring-repro-189-191 no longer needed.