[amar-python]

Fixes #84 and #81. Part of ART Milestone 1 (BR-23, BR-28).

#84: Alembic migrations

• Adds alembic.ini and alembic/. api/db.py bootstrap() already ran alembic upgrade head whenever alembic.ini existed. It never did, so every start took the fallback CREATE TABLE IF NOT EXISTS path. It now migrates on startup.
• 0001 baseline mirrors the fallback schema exactly, with IF NOT EXISTS. Databases created by the old fallback adopt it with no change.
• The version table lives in the uploads schema, so each schema carries its own migration state. That's ready for schema-per-tenant (ART decision 6).

#81: registry uniqueness

• 0002 adds UNIQUE (file_hash) and UNIQUE (file_name) on csv_files. dynamic_loader.py relied on these to settle concurrent identical uploads, but they never existed. The fallback schema gets the same constraints.
• If a database already holds duplicates (possible because of the old race), the upgrade stops and lists them rather than deleting anything. The API then won't start until they're cleaned up. This is deliberate: it's loud, and no data is lost.
• Dynamic upload losing a race: now returns duplicate_file. Before, it surfaced as a generic DB error from the concurrent CREATE TABLE.
• T&E load losing a race: now returns a structured error and rolls back its rows. With the new constraint it would otherwise have been an uncaught 500.
• UNIQUE (file_name) matches today's semantics. ART decision 2 (versioned datasets) revisits it in M5.

Evidence

Local PostgreSQL 16, replicating the CI integration job step for step:

Check	Result
scripts/test_report.py --strict	247 collected, 247 passed, 0 failed, 0 skipped (239 before + 8 new)
evals/runner.py --tiers p,i,s,e	26/26
pytest tests/test_migrations.py	8 passed; with the fix removed: 7 failed, 1 passed
scripts/lint.sh (now also covering api/db.py, the migrations and the new test)	flake8 clean, bandit clean

The lost-race tests are deterministic, not timing-based. A competing transaction holds an uncommitted registry row, the loader blocks on the UNIQUE index, then the competitor commits.

Deployment note: after this merges, the next API start creates csv_uploads.alembic_version and adds the two constraints. On your local te_mgmt_dev, if startup fails with "Cannot add UNIQUE", the error lists the duplicate rows to remove.

---

Generated by Claude Code