fix(seed): load reference data even when migrations seeded a term

[jirhiker]

What broke

The development-mode startup seed (core/app.py lifespan → seed_all) calls ensure_seed_prereqs, which loaded core/lexicon.json only when lexicon_term was completely empty:

has_lexicon = s.scalar(select(LexiconTerm.id).limit(1)) is not None
if not has_lexicon:
    init_lexicon()

Since b2c3d4e5f6a7_transducer_data_maturity the migration inserts a lexicon term of its own. On a fresh database, migrations alone now leave exactly one row in lexicon_term — measured, not guessed — so that emptiness check reports "already loaded" and the real 1144-term lexicon is skipped entirely.

The fallout, in order:

1. init_parameter violates parameter_default_unit_fkey, because ft and dimensionless were never inserted as lexicon terms. Two parameters fail and the errors are printed, not raised.
2. seed_all calls random.choice(organization_terms) on an empty organization category → IndexError: Cannot choose from an empty sequence.
3. Application startup failed. Exiting.

The frontend felt this as a CI outage: OcotilloUI's Cypress job polls http://localhost:8000/docs for the API, which was never going to answer, so the job burned its full 720s and failed with exit 124. See DataIntegrationGroup/OcotilloUI#350.

The fix

• ensure_seed_prereqs now runs init_lexicon() and init_parameter() unconditionally. Both already leave existing rows alone — init_lexicon upserts categories, terms, and associations with on_conflict_do_nothing — so the emptiness shortcut was buying nothing and cost correctness.
• init_parameter skips (parameter_name, matrix) pairs that already exist, so a re-run no longer trips uq_parameter_name_matrix and prints a wall of caught DatabaseErrors.
• New assert_lexicon_ready fails with the empty categories named, rather than leaving the next contributor to decode an IndexError thrown deep inside the seed.

Verification

Against a fresh PostGIS database, recreating the CI sequence (drop schema → alembic upgrade head → seed_all):

• With this branch: lexicon_term rows after migrations only = 1, then the seed loads all 1144 and completes — contacts=2, SEED OK.
• With staging's versions of the two files restored: the exact CI failure reproduces, same two FK violations on ft and dimensionless, same IndexError: Cannot choose from an empty sequence.

Tests: 4 new cases in tests/test_seed_prereqs.py covering that both initializers run even when the tables hold rows, that assert_lexicon_ready names the empty categories, that every required category has terms once reference data loads, and that init_parameter is idempotent.

uv run pytest --ignore=tests/transfers → 664 passed, 46 skipped, 6 xpassed. One unrelated failure, tests/test_ogc.py::test_ogc_project_areas_items_expose_groups_with_project_areas, reproduces identically on unmodified staging on this machine and is green on CI runners — local environment (arm64 host, amd64 PostGIS image), not this change. black and the pre-commit flake8 selection are clean.

🤖 Generated with Claude Code

[github-actions]

Coverage

✅ 79.05% total — gate is 75%.

Coverage for the Python files changed in this PR

Name	Stmts	Miss	Cover	Missing
core/initializers.py	137	20	85%	67-69, 73-82, 184, 188, 206, 271-284, 288-289
TOTAL	137	20	85%