What executing our own guide broke

Applies to The Fabrica v0.1.2. Documentation source: 1ee1bc9cb661, including post-release documentation corrections.

What this run fixed

Executing this procedure end to end surfaced six defects. All six are fixed in the same change that added this guide, because a guide that documents its way around a broken product is not a guide.

The generated tests could not run. They are written against authed_client / other_user_client / anon_client, and no such fixtures existed — every one errored on collection. The mandatory IDOR regression the scaffold advertises had never executed. The three fixtures now ship in conftest.py.
PATCH 500’d on every successful update (MissingGreenlet), and the test suite could not see it: the IDOR test only PATCHes a foreign row, which 404s before the handler body runs. The router template refreshes the row; the test template now has a happy-path PATCH.
Buyer dashboard pages were unroutable. The documented location, app/[locale]/_dashboard/_domain/, is a Next.js private folder. A probe page there produced no route while a control page under (dashboard)/ did:
```
├ ƒ /[locale]/probe-b     ← (dashboard)/probe-b
(no probe-a route)        ← _dashboard/_domain/probe-a
```
It is now the route group (dashboard)/(_domain)/, which routes and inherits the dashboard chrome.
alembic revision always failed with Could not find entrypoint console_scripts.ruff — after writing the file. The post-write hook is now type = exec.
The factory’s own tests rejected the first buyer resource. test_db_naming_convention.py and test_db_roundtrip.py pinned the table list with ==, so any _domain/ model turned the suite red in three assertions, none of which mentioned the scaffold. They are subset checks now; the round-trip additionally asserts that a re-upgrade reproduces the first, which covers buyer migrations too.
A find-and-replace mangled every generated docstring. The templates documented their own placeholders using the literal {{...}} markers, so substituting multi-line field blocks injected them into prose — and with long enough field lists it failed ruff on line length. The instruction docstrings now name block placeholders without braces.

A seventh came from the fresh-clone rehearsal rather than the scaffold: test_job_dispatch.py::test_celery_app_uses_project_slug asserted "thefabrica" in app.main, so the repository’s own checks failed for every buyer the moment factory-cli init renamed the project — the supported first step. It now asserts what celery_app.py actually builds, f"{project_config.project_slug}-workers", with no literal at all.

One thing was found and deliberately not fixed at the time: alembic revision --autogenerate reported large pre-existing drift across eight existing tables (column comments and index definitions the factory’s migrations never wrote). It was unrelated to the scaffold and reconciling it was a schema change, not a documentation one — so this guide originally routed around it.

That drift is now closed. Migration 007_schema_metadata_alignment emitted the 79 missing column comments, created the ix_api_keys_created_at index that TimestampMixin declares, and replaced two redundant constraint/index pairs with the single unique index the models describe. alembic check against a migrated database now reports nothing to generate, and test_migrated_schema_matches_model_metadata fails the build if that ever stops being true.

Provenance. Documentation source: 1ee1bc9cb6619c19f57766731e7884b37f515dc9 (main).

Extracted from docs/first-resource.md.

That commit is later than the release these pages describe: it carries documentation corrections made after thefabrica-v0.1.2 went out. Nothing that changes how the product behaves landed between the two, which is what makes naming that version honest.