docs(handoff): overnight gen-quality run — Issue #1 fixed, workflow complete

Final overnight handoff + the autonomous workflow's per-provider report.

Outcome:
- Issue #1 (decks invisible to panels) FIXED & PROVEN — was 2 bugs (read-root
  unconfigured + 1-level-only discovery), not the hypothesised single 3-level one.
- Autonomous gen+judge+validate+report workflow built, launched, completed:
  Bedrock APPROVED 9/9, Ollama APPROVED 8/7, MiniMax generates cleanly (2 adapter
  bugs fixed) but quiz quality is below the Opus judge bar (model limitation).
- Playwright validated all 3 provider courses render in the panels.
- Issue #2 (6 e2e selector failures) fixed. Issue #3 (model dropdown) deferred
  with rationale.

Open product decision for the user: MiniMax rotation (flashcards-only / both /
drop-for-quizzes). No further code needed for MiniMax to function.

Author: taylaand

docs/handoffs/2026-06-01-gen-quality-workflow-report.md

@@ -0,0 +1,246 @@
+# Generation + Quality Workflow Run — 2026-06-01
+
+**Course**: Complete SQL Mastery (CodeWithMosh)
+**Sections covered**: `getting-started-0025`, `data-types-0035`
+**Providers tested**: AWS Bedrock (Claude Sonnet 4.6), MiniMax M2.7, Ollama gemma4:latest
+
+---
+
+## Summary Table
+
+| Provider | Status | Quality Rounds | Gen-Error Retries | Flashcard Score | Quiz Score | API OK | Rendered |
+|---|---|---|---|---|---|---|---|
+| AWS Bedrock — Claude Sonnet 4.6 | APPROVED | 1 | 0 | 9/10 | 9/10 | Yes | Yes |
+| MiniMax M2.7 | **FAILED** | 3 (judge) + 2 gen errors = 4 total | 3 | 8/10 (last judge) | 4/10 (last judge) | Yes | Yes |
+| Ollama — gemma4:latest | APPROVED | 4 | 0 | 8/10 | 7/10 | Yes | Yes |
+
+---
+
+## Provider Detail
+
+### AWS Bedrock — Claude Sonnet 4.6
+
+**Status**: APPROVED in 1 round (no retries required).
+
+**Final verdict** (round 1):
+> Both decks are accurate, non-trivial, and teach the real WHY behind each concept. Content is approved for use.
+
+**Strengths**:
+- Flashcards test genuine understanding (FK vs denormalisation trade-off, FLOAT IEEE 754 accumulation, `->` vs `->>` silent failure) rather than rote recall.
+- Every answer is self-contained and correct.
+- Quiz distractors are plausible — e.g. "FLOAT is slower than DECIMAL" is deliberately inverted from reality; "avoid JOINs" is the exact wrong reason engineers cite for denormalisation; "ENUM cannot be indexed" is correctly attributed to TEXT not ENUM (cross-card reinforcement).
+- Rationales are educational, distinguishing mechanism from motivation. SQL examples are syntactically valid.
+- The intentional-denormalisation exception card is a standout — it captures the nuance most intro courses skip.
+- Getting-started deck covers both tool knowledge (Workbench execution model) and foundational principles (declarative vs procedural) without near-duplicates.
+
+**Weaknesses**:
+- The ENUM rebuild card/question does not mention that MySQL 8.0 supports `ALGORITHM=INSTANT` for appending new ENUM values at the end, avoiding a full table rebuild in many cases. This is a course-level abstraction and not a correctness error, but a senior engineer may notice the gap.
+- The Workbench Execute-button flashcard is more tool-mechanics than SQL understanding — acceptable given it is a getting-started module, but the weakest card in the set.
+
+**Round-by-round history**:
+
+| Round | Kind | Approved | Flashcard Score | Quiz Score |
+|---|---|---|---|---|
+| 1 | judge | Yes | 9 | 9 |
+
+**Output files**:
+- `/Users/taylaand/Obsidian/Personal/Study/CodeWithMosh/Complete_SQL_Mastery__bedrock/flashcards/getting-started-0025-flashcards.json` (5 cards)
+- `/Users/taylaand/Obsidian/Personal/Study/CodeWithMosh/Complete_SQL_Mastery__bedrock/quizzes/getting-started-0025-quiz.json` (5 questions)
+- `/Users/taylaand/Obsidian/Personal/Study/CodeWithMosh/Complete_SQL_Mastery__bedrock/flashcards/data-types-0035-flashcards.json` (5 cards)
+- `/Users/taylaand/Obsidian/Personal/Study/CodeWithMosh/Complete_SQL_Mastery__bedrock/quizzes/data-types-0035-quiz.json` (5 questions)
+
+---
+
+### MiniMax M2.7
+
+> **ATTENTION: This provider hit the retry ceiling without achieving approval. Human decision required — see section below.**
+
+**Status**: FAILED. Exhausted 3 gen-error retries and 3 judge rounds without approval. The final generation attempt (retry 3) produced an incomplete output — `data-types-0035-quiz.json` was never written because MiniMax returned XML-style tool-call markup instead of the expected `tool_use` block format.
+
+**Final judge reasons** (round 3, last successful evaluation):
+> Fix these specific defects before resubmitting:
+> 1. Getting Started quiz Q1 (Elka/relational model): The correct answer option contains garbled non-English text ("看到她新地址") embedded in an otherwise English sentence.
+> 2. Data Types quiz Q4 (JSON_SET path prefix): Marks "JSON_SET automatically normalizes bare paths by prepending $ internally" as CORRECT — this is factually wrong. MySQL's JSON_SET does NOT auto-prefix bare paths.
+> 3. Data Types quiz Q1, distractor D rationale: States VARCHAR(2) and CHAR(2) store the same number of bytes — incorrect; CHAR(2) stores 2 bytes, VARCHAR(2) stores 3 bytes.
+
+**Round-by-round history**:
+
+| Round | Kind | Approved | Flashcard Score | Quiz Score | Detail |
+|---|---|---|---|---|---|
+| 1 | judge | No | 8 | 4 | Two critical defects: Q1 option text contained a question fragment not an answer; Q5 `isCorrect` flag on wrong option |
+| 2 | gen_error | — | — | — | Generation failed: "1 task(s) failed" (retry 1) |
+| 2 | judge | No | 8 | 6 | Three distractor errors: wrong topic in distractor C; TINYINT UNSIGNED as a 5M-row distractor; copy-paste rationale error in Q5 distractor D |
+| 3 | judge | No | 8 | 4 | Garbled non-English characters in Q1 correct option; factually wrong correct answer in Q4; incorrect byte-count rationale in Q1 distractor D |
+| 4 | gen_error | — | — | — | Generation failed: "1 task(s) failed" (retry 2) |
+| 4 | gen_error | — | — | — | Generation failed: "1 task(s) failed" (retry 3 — ceiling reached) |
+
+**Last generation partial output** (3 of 4 files written; 1 failed):
+- `getting-started-0025-flashcards.json`: OK (12.1s)
+- `getting-started-0025-quiz.json`: OK (60.1s)
+- `data-types-0035-flashcards.json`: OK (36.6s)
+- `data-types-0035-quiz.json`: **FAILED** — MiniMax returned XML-style `<minimax:tool_call>` markup instead of a `tool_use` block; the parser could not extract `emit_quiz_deck`.
+
+**Root cause pattern**: MiniMax alternated between content quality failures (garbled text, incorrect `isCorrect` flags, factual errors) and tool-call format failures. The format failure (`<minimax:tool_call>` instead of JSON `tool_use`) is a protocol-level bug, not a content quality issue — it recurred on retries 2 and 3 and appears to be a reliability regression in the model's tool-use compliance.
+
+**Output files** (incomplete — missing data-types quiz):
+- `/Users/taylaand/Obsidian/Personal/Study/CodeWithMosh/Complete_SQL_Mastery__minimax/flashcards/getting-started-0025-flashcards.json` (5 cards)
+- `/Users/taylaand/Obsidian/Personal/Study/CodeWithMosh/Complete_SQL_Mastery__minimax/quizzes/getting-started-0025-quiz.json` (5 questions)
+- `/Users/taylaand/Obsidian/Personal/Study/CodeWithMosh/Complete_SQL_Mastery__minimax/flashcards/data-types-0035-flashcards.json` (5 cards)
+- `data-types-0035-quiz.json` — **NOT WRITTEN**
+
+**Note on API/render validation**: Despite the failed generation, the validation report shows `flashcard_count=10` and `quiz_count=10` for the minimax course. This is inconsistent with the generation record (only 3 of 4 files written). The validation agent may have encountered stale data from a prior run or the API is counting from a partially-consistent state. This discrepancy requires investigation.
+
+---
+
+### Ollama — gemma4:latest
+
+**Status**: APPROVED in 4 rounds (no gen-error retries). This provider required the maximum number of judge iterations but ultimately met the approval bar.
+
+**Final verdict** (round 4):
+> Content is genuinely useful and accurate. Approve with these improvements for the next generation pass.
+
+**Strengths**:
+- Data Types 0035 flashcards are outstanding — all five test design reasoning (ENUM rebuild cost, IEEE 754 vs DECIMAL, Y2K38 TIMESTAMP limit, JSON sparse-table avoidance, BIGINT range) with correct SQL examples.
+- Getting Started 0025 flashcards for FK/SSOT, junction tables, and Views are well-framed with concrete scenarios.
+- Quiz distractors throughout are plausible real-world anti-patterns (comma-separated columns, nullable column explosion, SET type, VARCHAR PKs) rather than throwaway wrong answers.
+- Rationales are educational and explain the why, not just the what.
+- Data Types 0035 quiz Q3 (ENUM-to-lookup-table migration) is particularly strong — all four options are patterns engineers actually propose.
+
+**Weaknesses (final round — deferred improvements, not blockers)**:
+- Getting Started 0025 has one trivia flashcard (loopback IP / default port) that adds no SQL design value. Replacement suggested: View vs CTE trade-offs.
+- Quiz Q5 tests MySQL Workbench UI mechanics rather than transferable SQL knowledge.
+- Quiz Q4 distractor D has a logically unsound rationale — dismisses a trigger as a design flaw when a trigger is actually a valid implementation mechanism.
+- Data Types 0035 quiz Q2 distractor B risks misleading learners: DECIMAL(19,4) is marked wrong for being "arbitrary" but it is the dominant industry convention for financial data; the rationale needs nuance.
+
+**Round-by-round history**:
+
+| Round | Kind | Approved | Flashcard Score | Quiz Score | Key issues |
+|---|---|---|---|---|---|
+| 1 | judge | No | 8 | 6 | Duplicate distractors in data-types Q5; imprecise correct answer in getting-started Q3; throwaway distractors; UI-trivia flashcard |
+| 2 | judge | No | 7 | 6 | Two quiz questions test GUI/CLI knowledge not SQL; Q3 "correct" answer contestable; getting-started flashcard 4 is Workbench trivia |
+| 3 | judge | No | 9 | 6 | Data Types Q5 JSON_EXTRACT rationale factually wrong in MySQL 8; Getting Started Q4 tests Workbench UI mechanics |
+| 4 | judge | Yes | 8 | 7 | Approved with deferred improvements noted above |
+
+**Output files**:
+- `/Users/taylaand/Obsidian/Personal/Study/CodeWithMosh/Complete_SQL_Mastery__ollama/flashcards/getting-started-0025-flashcards.json` (5 cards)
+- `/Users/taylaand/Obsidian/Personal/Study/CodeWithMosh/Complete_SQL_Mastery__ollama/quizzes/getting-started-0025-quiz.json` (5 questions)
+- `/Users/taylaand/Obsidian/Personal/Study/CodeWithMosh/Complete_SQL_Mastery__ollama/flashcards/data-types-0035-flashcards.json` (5 cards)
+- `/Users/taylaand/Obsidian/Personal/Study/CodeWithMosh/Complete_SQL_Mastery__ollama/quizzes/data-types-0035-quiz.json` (5 questions)
+
+---
+
+## Playwright / API Validation
+
+### Complete_SQL_Mastery__bedrock
+
+- **API**: `/api/courses` confirms course present with `flashcard_count=10`, `quiz_count=10`. Both `/api/cards` calls returned 10 items each.
+- **Rendered**: Playwright navigated to `http://127.0.0.1:8567/#flashcards`; course card visible showing "10 flashcards 10 quiz questions" with Flashcards/Quiz buttons.
+- **Console errors**: 5 non-blocking pre-existing errors; no new regressions.
+- **Sample flashcard front**: "You need to store product prices in MySQL. A colleague suggests using FLOAT for efficiency. Why is this wrong, and what should you use instead?"
+- **Sample quiz question**: "A financial application stores product prices. A developer defines the column as `price FLOAT`. What is the most serious consequence of this choice in production?"
+
+### Complete_SQL_Mastery__minimax
+
+- **API**: `/api/courses` reports `flashcard_count=10`, `quiz_count=10`. Both `/api/cards` calls returned 10 items each.
+- **Rendered**: Playwright confirmed course card present with "10 flashcards 10 quiz questions".
+- **Discrepancy**: The generation run recorded only 3 of 4 files successfully written (data-types quiz failed). The API and render counts of 10/10 are inconsistent with this. Either the API is serving stale data from a prior successful run, or there is a partial-write scenario that needs investigation.
+- **Sample flashcard front**: "A colleague argues that MySQL's `JSON_SET` is smart enough to normalize bare paths like `'weight'` automatically. They write `JSON_SET(properties, 'weight', 20)` and expect it to work. Are they right? If not, what is the correct syntax?"
+- **Sample quiz question**: "You need to store two-character US state abbreviations (always exactly 2 characters). Which storage comparison is correct?"
+
+### Complete_SQL_Mastery__ollama
+
+- **API**: `/api/courses` confirms course present with `flashcard_count=10`, `quiz_count=10`. Both `/api/cards` calls returned 10 items each.
+- **Rendered**: Playwright confirmed course card renders at `#flashcards` with "10 flashcards 10 quiz questions" and Flashcards/Quiz buttons visible.
+- **Sample flashcard front**: "When designing a schema, why is using a lookup table pattern (e.g., for 'size') superior to using an ENUM type?"
+- **Sample quiz question**: "When designing a product catalog where products have highly variable attributes (e.g., a TV has a resolution, a shirt has a size), which pattern is best suited to store these optional, schema-per-row attributes while maintaining relational integrity?"
+
+---
+
+## What Needs Human Decision
+
+### 1. MiniMax — FAILED provider: decide whether to retry or deprecate
+
+MiniMax hit the 3-gen-error-retry ceiling and produced unapproved content across 3 judge rounds. The failures fall into two distinct categories:
+
+- **Content quality**: Garbled non-English characters in output, incorrect `isCorrect` flags, factual SQL errors. These suggest the model does not reliably follow the judge's remediation instructions across multiple rounds.
+- **Protocol failure**: MiniMax returned `<minimax:tool_call>` XML-style markup instead of the expected JSON `tool_use` block. This is a tool-call compliance bug that prevented the `data-types-0035-quiz.json` from being written at all on the final two retries.
+
+**Decision needed**: Should MiniMax be retried with a different prompt strategy (e.g., stricter tool-call formatting instructions), or should it be removed from the provider rotation until the XML markup regression is investigated and fixed?
+
+### 2. MiniMax API/render count discrepancy
+
+The validation report shows `flashcard_count=10` and `quiz_count=10` for the minimax course, but the generation run only wrote 3 of 4 files. Someone needs to confirm whether the API is serving data from a previous (stale) run or whether there is genuinely a complete set of cards for this course that pre-dates this run. If the quiz counts are stale, the minimax course data should be treated as invalid for use.
+
+### 3. Ollama — deferred improvements (approved but non-trivial feedback)
+
+The Ollama content was approved but with four documented improvement items that the judge explicitly flagged for the "next generation pass". These are not blockers, but if content quality is a priority they should be addressed before the course is distributed to learners:
+
+- Replace connectivity-trivia flashcard (127.0.0.1 / port 3306) with a View vs CTE trade-offs card.
+- Fix quiz Q4 distractor D rationale (trigger dismissal is logically unsound).
+- Replace quiz Q5 (Workbench highlight-to-execute) with transferable SQL knowledge question.
+- Nuance the DECIMAL(19,4) distractor rationale in data-types quiz Q2.
+
+**Decision needed**: Accept as-is for learner use, or queue a targeted regeneration pass for these four items?
+
+### 4. Bedrock ENUM/ALGORITHM=INSTANT gap
+
+The judge noted that the ENUM rebuild card does not mention `ALGORITHM=INSTANT` for MySQL 8.0 (appending ENUM values without a full table rebuild). This is a course-level abstraction and not a correctness error, but it is a meaningful omission for a senior engineer audience.
+
+**Decision needed**: Accept the omission as appropriate course-level simplification, or add a follow-up card covering the MySQL 8.0 exception?
+
+---
+
+## STEP-3 ADDENDUM (post-workflow targeted fixes) — 2026-06-01
+
+After the workflow completed, I root-caused MiniMax's FAILED status and fixed the
+**code** defects it exposed, then re-ran MiniMax. Summary of what changed:
+
+### MiniMax code bugs found & fixed (proven live)
+
+1. **error 2013 → tool_result correction** (commit `2c26c2a`, fixed earlier in the
+   night before the workflow): the schema-correction retry appended a plain-text
+   user turn after an assistant `tool_use`; MiniMax's strict `/anthropic` shim
+   rejects that. Now emits a protocol-valid `tool_result`.
+2. **inline-XML tool calls** (commit `c4b2030`): MiniMax M2.7 intermittently emits
+   the forced tool call as `<minimax:tool_call><invoke …><parameter …>` text
+   markup instead of a native `tool_use` block (~half the time). Added
+   `_parse_inline_tool_call` fallback **plus** a transient-retry in
+   `call_with_correction` (a single bad emission no longer hard-fails). This is the
+   exact protocol failure the workflow flagged as needing investigation.
+3. **`--max-retries` knob** (commit `209d0e4`): flaky providers can be given a
+   larger retry budget so consecutive bad emissions don't exhaust it.
+
+**Proof**: MiniMax data-types quiz (previously failed reliably) now generates
+4/4 clean consecutive runs (0 errors); a full clean two-section run wrote 4/4
+files (5+5+5+5 = 20 cards).
+
+### MiniMax content quality — STILL below bar (model limitation, NOT code)
+
+After clean generation, an Opus-4.8 judge re-assessed the fresh MiniMax decks:
+
+> **VERDICT: REJECTED — flashcard 9/10, quiz 6/10.** Getting-Started quiz Q4 has
+> `isCorrect` on the wrong option for the scenario described (Workbench
+> execution-scope confusion); the Data-Types quiz has two rationale-to-option
+> copy-paste mismatches. Flashcards are excellent; quizzes carry subtle
+> correctness defects.
+
+This matches the workflow's 3 in-loop rejections. **Conclusion: MiniMax M2.7's
+generation pipeline is now fully working (zero errors), but its quiz content
+quality is a model limitation the strict judge keeps catching.** Its flashcards
+(score 9) are genuinely strong.
+
+### Final provider scorecard
+
+| Provider | Generates cleanly? | Judge verdict | Notes |
+|---|---|---|---|
+| **Bedrock Sonnet 4.6** | ✅ | **APPROVED 9/9** (round 1) | Best — accurate, deep, plausible distractors |
+| **Ollama gemma4** | ✅ | **APPROVED 8/7** (round 4) | Needed 4 judge rounds; approved with deferred nits |
+| **MiniMax M2.7** | ✅ (after code fixes) | **REJECTED 9/6** | Flashcards strong; quiz quality is a model limit, not a bug |
+
+### Decision for the user (MiniMax)
+
+The code is fixed and committed — MiniMax no longer errors. The open question is
+purely product: **keep MiniMax (it produces excellent flashcards) and accept
+weaker quizzes, restrict MiniMax to flashcard-only generation, or drop it from
+the rotation for quizzes.** No further code work is required for MiniMax to
+*function*; the gap is content quality intrinsic to the model.