[kvlonge]

Why

#978 made three parts of a SQL test file optional. The optional TEST() header is being reverted: every test file should again start with a TEST() header so test files are unambiguous to read, lint and format. Mode inference and the optional trailing SELECT 1 from #978 stay.

Changes

• Discovery: a test file without a TEST() header fails with D003 again ("must start with a TEST() header as the first non-whitespace content"). The header-less parse path is removed; mode inference from __macro_actual__, __udf_actual__ and __table_fn_actual__ is kept for headers without mode.
• Rules and formatter: files under tests/unit are no longer treated as tests without a header (LintFileRole.in_sql_test_directory removed). Omitted SELECT 1 handling for TEST and SCENARIO bodies is unchanged.
• Tests: header-less cases now expect D003; mode-inference, formatter and E2E cases use TEST(); and still omit SELECT 1; a new E2E case checks a header-less test file is rejected with D003.
• Docs, README, skill and waffle-shop example tests: TEST(); restored. Docs state that every test file starts with a TEST() header (fields optional; an unnamed test is named after its file), the mode is inferred, and the trailing SELECT 1 is optional. Skill references regenerated.

Verification

• uv run pytest on discovery test_sql_tests.py, lint test_run_lint.py, compile helper test_tests.py and test_scenarios.py, and E2E test_optional_test_ceremony.py (182 passed)
• ruff check, ruff format --check, uv run fensu check
• make check-ci