Restore the green release matrix: Course creation hits NOT NULL on courses_course.archive_commit_sha

[alexeygrigorev]

Plan context: CI run 34189057638 (87fd3f4) is the first completed django verdict on main since Sep 5 (earlier runs were cancelled by the container/quality failures fixed in #337/#319/#341/#343): quality, container, playwright and screenshots now pass, but the django job fails and ci-gate reports failure. Dozens of tests ERROR at setup with sqlite3.IntegrityError: NOT NULL constraint failed: courses_course.archive_commit_sha. Introduced by the shared-curriculum merge (#320).

Symptom

• Failing modules include courses.tests.test_course_list_ordering, courses.tests.test_course_list_metadata, courses.tests.test_course_list_assignments, courses.tests.test_homework_optional_fields, courses.tests.test_local_question_seed, courses.tests.test_local_project_review_seed, studio_courses.tests (course/homework/project views), core.tests.test_course_platform_vendor_assets, review_import.tests.test_workflow.
• The model field (courses/models/cohort.py:192) declares default="", and the visible 0005 AddField declares default="" too, yet some INSERT into courses_course omits the column entirely.
• The visible AddField in courses/migrations/0005_shared_current_curriculum.py names model cohort, while the failing table is courses_course; a missing or default-less AddField for the course model is the prime suspect.

Direction

• Inventory every archive field AddField in migration 0005 and which model each targets; run the migration-drift check.
• Identify the creation path that bypasses the model default (missing AddField, bulk/raw insert, or fixture) and fix it at the root per the model's documented semantics: empty string for current cohorts, full 40-hex commit SHA for github_archive cohorts; the archive identity check constraint stays intact.

Scope

• The smallest change that makes Course creation safe everywhere the model default was assumed. No product behavior changes beyond what the model already specifies. No workflow or quality-contract edits.

Acceptance criteria

• The migration-drift check passes.
• All modules listed above pass locally.
• The full Django suite passes locally.
• Post-merge CI on main: the django job passes and the run reaches a green ci-gate verdict with no cancellations.

[alexeygrigorev]

Cross-reference: the same django job also fails three route-shape assertions in courses.tests.test_course_list_ordering (canonical /cohorts/ routes vs legacy expectations, helper course_card_html using reverse(course, ...)). That class is owned by #348 and is intentionally out of scope here; this lane only fixes the archive_commit_sha NOT NULL errors. Merge order may require a small rebase in courses/tests/course_list_base.py.

[alexeygrigorev]

CI trace of run 34662414649 (and the whole red streak since 2026-09-11 14:44Z): the release run is not failing on tests; the quality job fails its lint gate and then cancels the whole run (the step named "Cancel remaining release jobs after a quality failure"), so ci-gate fails on missing checks and Deploy Dev verify-ci fails downstream.

Root cause on current main (acf0093): 12 ruff errors.

• E501 line too long: content_sync/course_repository_v2.py:664, content/tests/test_person_profile.py:397, content/tests/test_podcast_catalog.py:609, content/tests/test_podcast_catalog.py:918, content/tests/test_podcast_episode_graph.py:393, events/eventbrite_content.py:396, test_support/local_course_seed_fixture.py:26, test_support/local_course_seed_fixture.py:33
• I001 import block un-sorted: courses/services/cmp_content_import.py:56, courses/services/cmp_learner_history_import.py:109, events/eventbrite_content.py:411, events/mailchimp_tag_import.py:57

Notes for whoever holds the lint fix lane:

• django, playwright and container jobs have not actually executed since the streak began; they were cancelled 20-40 seconds in each time. After the lint fix lands, expect the quality contract to continue to typecheck, migrations-check, django-check, deployment-check and test-ci, and further failures may surface there.
• This is the same blocker recorded in the plan STATUS rows for D0.1c and D1.2ca (community-base repository, docs/plan/STATUS.md).

[alexeygrigorev]

Merged to main as 950065e (f352074 on fix-ci-345-lint) after independent tester and PM gates.

The diff covers the 12 ruff errors from the trace above plus everything the job cancellations masked behind them, layer by layer:

• 8 E501 wraps (string contents preserved) and 4 I001 import sorts.
• 9 files that would have failed the format gate once lint stopped dying first.
• 12 mypy errors in 4 files: typing annotations in events/tests/test_eventbrite_content.py, an isinstance chain replacing an equivalent all() guard in scripts/prod/registration_sources/eventbrite_registrants.py, a TYPE_CHECKING Iterable[PendingEventRegistrants] annotation in scripts/prod/import_event_registrants.py, and a defensive ValueError guard in scripts/prod/legacy_zoomcamp/certificate_import.py whose only call site already checks for None.
• A latent quality-contract-v3 bug: scripts/ci.py's quality dispatcher had no test-ci branch, so target 9 of 9 could never run even with lint green. It delegates to the existing test task now.
• 15 files added under excluded trees since the BE-16 baseline without gate coverage: 11 opted into scripts/ci.py's ADOPTION_INTEGRATION_PYTHON plus pyproject.toml's strict mypy block, 4 scripts/prod files recorded in ci/adoption_baseline.txt (they were already linted and strictly type-checked via the scripts/prod directory opt-in).
• tests_ci guard: the retired temporary/ staging directory's reviewed exception removed (2aaafb4), and .agents/ added to surface.agent_process in ci/ownership.json (the website-illustrations skill from cee4bb0 had no graph entry; the assertion was unreachable until now).

Local verification on the same tree (base acf0093): full quality contract exit 0 with all nine targets green, ending 671 contract tests passed; focused Django tests over the five touched test modules 117 OK; the touched strict pytest files 56 passed. No behavior change beyond the one defensive guard; no user-facing surface touched.

Heads-up for this CI run: the django, playwright, container and deploy jobs execute for the first time since the streak began, so anything failing there is a new trace under this issue, not a regression of this merge. On-call is watching run 34676767683.

[alexeygrigorev]

On-call trace of CI run 34676767710 (push 950065e, 05:54Z): the lint-stack fix unblocked the first real django/playwright execution since 2026-09-11 14:44Z, and the complete-suite run failed as predicted in the 05:56Z heads-up. Deploy Dev 34676767683 verify-ci failed downstream on the same sha.

Django job failure taxonomy (39 failing tests, grouped by terminal error):

1. Corpus-root reads (FileNotFoundError under /home/runner/prod/dtc-data/content-staging/): scripts.tests.test_verify_local_dataset (4), scripts.tests.test_public_projection_media_digest (6), scripts.tests.test_reviewed_release_import CompetingActivationTests (2, faq_projection.json), content.tests.test_public_media_tooling (2). These read the external reviewed corpus through DEFAULT_PROJECTION_ROOT; CI is corpus-less by design (PUBLIC_MEDIA_STORE_BACKEND: memory, checkout carries no projection), so they cannot pass as written.
2. Bridge/route-registry reads (EventDescriptionBridgeError / BridgeBuildError "public route registry is unavailable"): scripts.tests.test_event_description_bridge (7), scripts.tests.test_luma_event_descriptions (7). Same root: corpus-derived registry absent.
3. Corpus-pinned fixtures: core.tests.test_design_system_shell DesignFiveAShellTests setUpClass KeyError 'alexeygrigorev' (people catalogue empty without the projection), core.tests.test_homepage MainHomepageRoutingTests (2, ValueError substring not found while rendering course_list), events.tests.test_models EventIdentityRouteTests.test_long_event_slug_uses_short_canonical_and_stale_slug_redirects, core.tests.test_web_runtime_coherence.test_deployed_browser_smoke_receives_complete_runtime_identity (TestSafetyError: remote tests require the reviewed registry target class).
4. Genuine regression from 4c9b523 (09-11 22:50, retired EventAlias): review_import.tests.test_workflow.test_fresh_scrub_removes_event_dependencies_with_foreign_keys_enabled queries the dropped events_eventalias table (sqlite3.OperationalError: no such table).

Attribution: clusters 1-3 are the un-decoupled remainder of the corpus decoupling campaign (commits 6313cd9..acf0093 of 09-12 01:04-02:40 decoupled podcast, person, media/routes and events fixtures; these 10 files were not covered). They predate the streak, but CI had not run the complete suite since 2026-09-10 02:39Z (run 34430301722), so they were last exercised before the corpus left the checkout. Cluster 4 is a test-side miss in 4c9b523, verified only under the cancelled-jobs regime.

Playwright job: ~22 failures concentrated in design-parity, podcast, person and foundation smoke tests; trace pending from whoever takes that lane (a font/header fix sequence is already in flight on the local main checkout, unpushed).

Routing: clusters 1-4 route back through engineering on this issue, following the established decoupling pattern (synthetic fixtures, no real-corpus reads). CI stays red and D0.1c/D1.2ca stay blocked until the complete suite passes.

[alexeygrigorev]

Implementation claim (worktree-holds-it): the django clusters 1-4 lane of the CI-run-34676767710 taxonomy is owned by worktree .tmp/worktrees/issue-345-django on branch issue-345-django, cut from origin/main.

Scope: clusters 1-3 (corpus-root reads, bridge/route-registry reads, corpus-pinned fixtures) decoupled onto the synthetic reference fixture following the established 6313cd9..acf0093 pattern, and cluster 4 (review_import querying the retired events_eventalias table) fixed test-side. The Playwright lane stays with the font/header fix sequence already in flight on the main checkout; this lane touches no templates and no product render surface, so screenshots are not_applicable and the applicable browser tier is the backend-only smoke tier.

Non-goals: no production import path changes, no real-corpus re-coupling, no quality-contract edits.

[alexeygrigorev]

Engineer trace for the run that followed the clusters 1-4 taxonomy, plus a corpus-less reproduction recipe for the django lane owner.

CI run 34681609718 (push 8a6f6e5, 07:46Z, the docs-refresh-380 merge that also carried the design-token commits 695a671..37b57e9): resolve-release, classification, container and screenshots succeeded, but quality, django and playwright were cancelled together 4 minutes in (07:49:45-07:49:58Z), before any of them produced a failure. ci-gate therefore failed on three cancelled checks and publish/deploy were cancelled behind it. Deploy Dev 34681609984 then failed at verify-ci (the same source has no green CI verdict). There are no lint errors and no failed steps in the cancelled jobs, so this is not the quality-cancels-run path from the earlier streak; the run needs a fresh verdict. Main still has no completed release verdict on its current HEAD.

Corpus-less reproduction for the django clusters (verified on a .tmp/worktrees/fix-main-red snapshot of the uncommitted lane work): the corpus-dependent tests resolve the projection from Path.home() / prod / dtc-data / content-staging / public_projection, so this development machine hides the CI behavior whenever the corpus is present. Run the affected files with HOME pointed at an empty directory (for example HOME=$(mktemp -d) uv run --frozen pytest ...). With the corpus hidden, content.tests.test_public_media_tooling, scripts.tests.test_verify_local_dataset and review_import.tests.test_workflow pass on the current uncommitted work (72 passed), and the bridge, luma and media-digest files now carry corpus-conditional skip guards, so they skip as designed.

Worktree bookkeeping: the 08:02Z claim names .tmp/worktrees/issue-345-django, but the lane edits are landing in .tmp/worktrees/fix-main-red (last modified 08:24Z, still uncommitted, file set growing between checks). One owner should consolidate before the tester gate, and the branch used for the merge should be the one that actually holds the diff.

Lane claim (worktree-holds-it): the post-landing verdict on the playwright side of run 34676767710 is owned by worktree .tmp/worktrees/issue-345-playwright-verify, cut detached from origin/main. Scope: run the applicable playwright tiers corpus-less against current main to check whether the design-token commits retired the ~22 design-parity, podcast, person and foundation-smoke failures. No template or product edits; the django clusters lane stays with its owner.

[alexeygrigorev]

Playwright verdict for the unowned lane, from worktree .tmp/worktrees/issue-345-playwright-verify (detached 8a6f6e5, corpus-less HOME): the smoke tier finishes 11 failed, 19 passed, 2 teardown errors. The design-token commits did not retire these failures; they split into two roots, both test-side, and neither needs a template or product change to verify.

Root 1, corpus-pinned fixture seeding (7 failures plus both teardown errors), the same cluster-3 gap as the django tier, not covered by 6313cd9..acf0093: test_internal_event_to_person_flow (both viewports) and test_public_pages_remain_meaningful_without_javascript fail because FEATURED_SPEAKER_PATH (/people/alexeygrigorev.html) 404s -- the people registry the browser tests stand on is the corpus-derived one, matching the KeyError 'alexeygrigorev' in core.tests.test_design_system_shell. test_podcast_latest_middle_and_oldest_seasons (both viewports), test_all_public_hub_aliases_redirect_once_with_query and test_oldest_latest_details_and_media_fallback die at StopIteration in next(... catalogue.podcasts() ...) -- no row with slug practical-llm-engineering-and-rag exists corpus-less. Fix shape: seed these modules from the synthetic reference fixture the decoupled podcast/person suites already use.

Root 2, stale pins against shipped copy (2 failures): test_public_home_and_hubs (both viewports) dies at CLIMB_HEADING ("From \u201cWhat does that mean?\u201d to \u201cLet me show you.\u201d"), a string that no longer exists anywhere in the product after 78ae0c9 (2026-09-10 22:46, homepage learner-journey rewrite). The pin needs the shipped copy, chosen the way 44459519 did it.

One failure needs template inspection before routing: test_docs_and_faq_root_trailing_slash_browser_contract (both viewports) gets 200 on /faq/ with the right canonical/og tags but finds zero a[href="/faq/"] where the contract expects exactly one self-link -- either the docs-cards redesign dropped the self-link or the contract predates the whole-surface card pattern.

Full failure log: /tmp/aisl-pw-smoke.log on the dev host; the worktree holds the exact reproduction command in its lane note.

[alexeygrigorev]

Follow-up on the cancelled 07:46Z main run: a rerun cannot produce a test verdict, so a fresh run was dispatched instead.

Rerun trace: rerunning 34681609718 fails structurally in attempt 2. Attempt 1 was cancelled before the release-source-identity-34681609718 artifact was sealed, and attempt 2 hard-fails in resolve-release at the step Restore the sealed source identity on a workflow rerun (Artifact not found for name: release-source-identity-34681609718), which skips every downstream job (django, playwright, quality, screenshots all skipped). Conclusion failure here is harness mechanics, not a new or changed test result.

Fresh verdict: CI workflow_dispatch on release_sha=8a6f6e51 with deploy_development=false, probe_development=false is running as 34684578554 (https://github.com/DataTalksClub/website/actions/runs/34684578554). Expect the same cluster set as 34676767710 plus the design-token effect on the two /faq/ self-link and CLIMB_HEADING nodes already traced from the detached worktree. Verdict will be recorded here when it completes; note a cancelled-before-seal run should not be rerun again — dispatch a fresh run on the SHA instead.

[alexeygrigorev]

Verdict on the fresh dispatch 34684578554: it completed but produced no test verdict, and its green django/playwright conclusions must not be read as the failures being gone.

What actually happened: resolve-release and classification succeeded, and the classifier profiled the run as documentation_or_contract (a workflow_dispatch carries no push diff, so the classifier sees the docs-only delta of 8a6f6e51). On that profile the django and playwright jobs only validated the selection and provenance — zero tests executed, conclusion success with no pytest output. Container and screenshots did real work. The gate then correctly failed: QUALITY_RESULT: cancelled — the quality job was cancelled downstream of the gate decision, which also cancelled the tail (auto-capture-prior, publish, deploy). ci-gate requiring every non-probe check is the pipeline working as designed on an incomplete run.

Net position: main's standing code-profile failure record is still run 34676767710 at 950065e0 (the cluster taxonomy from the earlier traces). A docs-only SHA cannot get a code-profile verdict through dispatch because the classifier will always skip the code suites; the fresh full-code verdict will come naturally from this lane's fix PR CI. One operational note: a run cancelled before sealing release-source-identity-<run id> cannot be rerun to a verdict (attempt 2 fails in Restore the sealed source identity on a workflow rerun) — dispatch a fresh run instead of rerunning.