Publish authoritative homework questions for every current non-legacy course

Outcome

Every homework exposed by a current, non-legacy, live schema-v2 cohort must contain its real, source-authoritative questions rather than the generated placeholder “Describe one practical idea you took from this module.” The structured YAML in each course repository remains the publication authority; the website must not patch content in templates, Python, migrations, or fallback files.

Normative references:

• zoomcamp-ops/docs/shared-curriculum-v2.md
• zoomcamp-ops/docs/curriculum-contract.md
• _docs/design/specs/unit-content-pipeline.md
• _docs/runbooks/course-content-push-and-pull.md
• _docs/architecture/database-only-content.md

Boundary and authority decisions

• In scope are mapped homework manifests of the one course.yaml.current_cohort whose cohort.yaml says delivery: live and curriculum: current in schema-v2 repositories. Today that means AI Dev Tools 2026, Data Engineering 2027, Machine Learning 2026, and LLM 2026.
• MLOps self-paced is intentionally excluded: its v2 contract requires homework: []; this issue must not weaken that rule or invent scored self-paced assignments.
• Stock Markets Analytics is schema v1 with no current/non-legacy homework-manifest graph, so CMP’s 2026 questions are not published by this issue. Migrating that course is a separate product/content decision.
• Authority precedence is: (1) sanitized current CMP public-content prompts/options/points for a matching current cohort/homework (currently AI homework 1–2 and LLM homework 1–5); (2) the matching current repository homework.md and current teaching material for gaps (AI homework 3–4, LLM homework 6–7, ML 2026, and DE 2027 drafts); (3) a named course owner. Older cohorts are corroboration only, never silently copied forward.
• CMP answers are sensitive and were not queried or logged. Never inspect, export, infer, or commit them. A native choice question without an approved encrypted answer may remain a choice question only with points: 0; do not degrade it into generic free text or pretend it is gradable.

Scope

1. Build a checked inventory of every in-scope {repository, cohort, module, homework.yaml, homework.md}, its current placeholder/real questions, and the authoritative source used for each replacement.
2. Replace placeholder-only question arrays with the actual current assignment questions, preserving the authoritative wording, order, question type, options, and points that can be supported safely.
3. Give every structured question a stable unique content_id and stable meaningful id; preserve existing published homework identities/slugs/routes.
4. Add encrypted answer envelopes only through the approved authoring workflow and key id. If no approved envelope exists, set the question to unscored without inventing an answer.
5. Commit and push reviewed source changes to each course repository’s main in focused per-repository commits, then run the normal website pull/import and verify the persisted/rendered result. Record source and website sync SHAs.

Non-goals

• Writing new substitute questions when no authority can be found.
• Reading or exporting CMP learner answers, correct-answer tables, submissions, or other protected data.
• Migrating schema-v1 Stock Markets Analytics, adding MLOps self-paced homework, or editing legacy/archive cohorts.
• Changing homework due dates, open/closed state, form telemetry fields, routes, grading policy, or lesson content.
• Adding a website-side content fallback or hardcoded question copy.
• Bundling the community-notes removal in #424 or lesson/video/navigation work in #425; each must remain an independently reversible change.

Dependencies and safety

• Requires read access only to the sanitized CMP public-content projection plus the public source repositories. Any answer encryption requires the existing approved key/authoring procedure; lack of it is not permission to access sensitive answers.
• Source repositories may have unrelated work in progress. In particular, AI Dev Tools and Data Engineering were dirty during grooming. Implement from a clean, pinned worktree or coordinate ownership; never stash, overwrite, or include unrelated changes.
• Run each source repository’s own AGENTS.md/checks plus zoomcamp-ops validation before push. Pushing focused source commits to main is authorized by the raw request, but only after review/checks.

Acceptance criteria

• The inventory accounts for every mapped homework in all four in-scope live current cohorts and explicitly records the MLOps and Stock Markets exclusions above.
• No in-scope homework.yaml contains the exact generic placeholder or a paraphrased filler replacement; every question has a cited current authority.
• CMP-backed entries preserve sanitized prompt, option labels/order, type, and points unless a documented safety normalization is required. No protected answer value appears in a diff, log, issue, screenshot, or report.
• Repository-backed entries are traceable to the exact current homework.md/teaching source path and commit; wording is not inferred from an older cohort.
• Choice/check-box questions remain native structured choices with at least two stable option ids. Free-form questions declare the correct answer type. Stable question ids/content ids are unique and survive replay.
• A question with positive points has a valid approved encrypted answer envelope. Without one it is explicitly points: 0; imports and UI do not claim it can be auto-scored.
• Homework manifest, source-repository, parser/import, duplicate-id, answer-envelope, and mapping checks pass in each affected repository and the website.
• Each source repository change is a focused reviewed commit on main; the report records repository SHA, authority paths, exclusions, and rollback commit.
• Website sync/import pins those accepted SHAs, completes without unresolved homework/question mappings, and a replay is a no-op with unchanged question counts/identities.
• On a real current homework page per affected family, the rendered prompt/order/options/points match source YAML; submission controls and scored/unscored messaging match the manifest. No placeholder remains in the database or rendered current routes.
• Legacy/archive course pages and their stored submissions/questions are byte/row-count stable.

Verification scenarios

Source/repository

1. Generate the inventory and provenance table without learner data.
2. Run each repository’s curriculum checker and tests against the exact commit to push; run zoomcamp-ops contract/fixture tests for every supported question shape.
3. Search current live manifest paths for the placeholder and assert zero matches; separately prove excluded archive paths were untouched.
4. Import the pinned commits into an isolated website database twice and compare homework/question identities, counts, prompts, options, points, and answer-envelope metadata.

Browser

1. Open at least one current homework for AI Dev Tools, Data Engineering, Machine Learning, and LLM, including one CMP-backed and one repository-backed gap.
2. Verify questions/options/order and scored versus unscored treatment against YAML; submit a synthetic response through each supported control type.
3. Check anonymous preview plus signed-in open/closed states at desktop and 390px mobile in both themes; no prompt/options overflow and keyboard focus/order remain correct.
4. Open a representative legacy/archive page and confirm its content and route are unchanged.

Blocker rule

If an in-scope homework has no defensible current authority, stop on that homework, record the exact missing authority, and request a course-owner decision. Do not close the issue by inventing content.

Evidence to attach

• Inventory/provenance/exclusion table, without sensitive answers.
• Source SHAs and focused check results per repository.
• Website import/replay report and rendered desktop/mobile evidence.

Start with the mapped current-cohort homework.yaml and homework.md files in the AI Dev Tools, Data Engineering, Machine Learning, and LLM repositories, then read the linked curriculum and content-pipeline contracts. Build the inventory and provenance table first, run each repository’s checks plus the website pull/import, and consider the work complete only when replay is a no-op, current rendered homework matches source, and legacy/excluded content is unchanged.