test: cover the documented command surface and the relation invariant

[imakecodes]

Closes #12.

Bottom of stack #61. Merge this first; #60 sits on top.

Why

Coverage was 89%, but the gap was not evenly spread. Three of the seven options the README documents had no test at all.

Untested before	What it does
--table_names	switches every table identifier to the physical DB name
--group_by_app	emits TableGroup blocks
--color_by_app	emits headercolor per table

Also unverified: selecting a whole app by label, the error on an unknown label, the default update timestamp, explicit through models, the database-engine mapping, router-aware table notes, and choices_to_markdown_table.

The part that is not about a percentage

Issue #38 was a schema dbdiagram rejected outright, because a ref named a column its table never declared. test_every_relation_endpoint_references_a_declared_column encodes that rule rather than any single column name, so it catches the whole class of failure instead of the one instance that was reported. It runs over both naming modes, since --table_names rewrites every table identifier and every relation endpoint with it.

How the assertions work

Through tests/dbml_parser.py, which reads generated DBML back into tables, columns, relations and groups. Substring matching cannot express "this ref points at a column that exists", and it passes silently when output moves. The parser is an assertion aid scoped to what the renderer emits, not a DBML implementation.

Fixture models gain the shape reported in #38 (explicit AutoField pk plus a custom db_table), a one-to-one alongside a foreign key to the same model, and an explicit through model, which is the branch where no join table should be synthesized.

Result

	Before	After
Tests	4	47
Coverage	89%	99%
selection.py	87%	100%
renderer.py	84%	100%
utils.py	71%	100%

The four remaining statements are the defensive dedup return in _build_many_to_many_table and the base_field.choices branch, which needs django.contrib.postgres. Both are documented in CLAUDE.md as deliberate so nobody chases them for the number.

pytest-cov and a make coverage target keep this measurable. No threshold is enforced in CI; that is a maintainer call, and a floor that fails unrelated PRs is worse than none.

Verification

make ci clean. 47 tests pass on both ends of the support matrix, Python 3.11 with Django 4.2 and Python 3.14 with Django 6.1.

No source changes in this PR.

🤖 Generated with Claude Code