feat(api): FastAPI skeleton with two-role RLS auth foundation

Phase 0's API half: a working FastAPI app with register/login/me/logout,
backed by a Postgres schema where row-level security is real and
independently proven, not just declared.

The core design decision, and the reason this lands as one PR instead of
several: request-scoped queries run as `velodrome_app` (NOBYPASSRLS), but
looking up identity in the first place — login by email, a session by its
token hash — has to happen *before* app.user_id can be set, so those specific
lookups run as a second role, `velodrome_auth` (BYPASSRLS), used nowhere else
in the codebase. See velodrome/db.py's module docstring and apps/api/README.md
for the full rationale. This is genuinely one reviewable unit: the migration,
the models, and the auth service only make sense evaluated together, since
they're three views of the same invariant.

tests/test_auth.py::test_rls_blocks_cross_user_session_reads is the test
worth reading first — it doesn't trust the RLS policy SQL because it reads
correctly, it proves isolation by registering two users and confirming a
scoped read of `sessions` for user A returns exactly one row, never two.

Bugs found and fixed while actually running this against real Postgres
(everything below was verified against a live postgis/postgis:16-3.4
container and a built Docker image, not just read for correctness):

- CREATE ROLE's PASSWORD clause is DDL, not DML — it doesn't accept bind
  parameters (`PASSWORD $1` is a syntax error). Fixed with dollar-quoting.
- Postgres roles are cluster-wide, not per-database — a second database in
  the same cluster hit "role already exists" on a plain CREATE ROLE. Fixed
  with a DO block catching duplicate_object.
- A bare `Mapped[datetime]` on the ORM models infers a naive timestamp,
  silently disagreeing with the migration's correct `DateTime(timezone=True)`
  — asyncpg rejected the mismatch at insert time. Fixed once, at the
  declarative Base level via type_annotation_map, rather than per-column.
- The session cookie's `secure` flag was gated on `!= "development"`, so
  anything else — including local testing and a real deploy running
  temporarily without TLS in front — got a Secure cookie no HTTP client
  will ever send back, breaking every authenticated request after login
  with no visible error. Gated on `== "production"` instead.
- `alembic check` initially flagged every PostGIS/TIGER-installed table
  (dozens of them) as drift, because they're not in our metadata. A
  schema-based denylist doesn't work — reflected foreign tables come back
  with schema=None regardless of their real schema. Fixed with an
  allowlist keyed on target_metadata.tables instead, which is also more
  robust against future PostGIS versions adding more tables.
- The migration itself was missing `nullable=False` on three timestamp
  columns that the ORM model assumed were never null — a genuine
  model/migration drift that alembic check caught once the PostGIS noise
  above was filtered out. Fixed in 0001 directly, since it's never shipped.
- Two indexes the migration creates explicitly weren't declared on the
  ORM models, causing the same kind of drift. Added index=True to match.

Deliberately deferred, not forgotten: per-IP/per-account login rate
limiting (docs/PLAN.md mentions it; Phase 0's bar is a working skeleton,
and this needs its own design pass) and the procrastinate job runner /
worker container (nothing to run yet — arrives with the ingestion
pipeline).

ci.yml updated to match: the api and migrations jobs now provision the
same two runtime roles this code actually needs, replacing the single
placeholder DATABASE_URL from before any code existed.

Verified: ruff check, ruff format --check, and mypy --strict all clean.
12/12 pytest passing against a real Postgres. Full alembic upgrade ->
downgrade -1 -> upgrade cycle run twice (once standalone, once inside a
two-database cluster to specifically catch the role-collision bug).
alembic check clean. Docker image builds and serves real traffic —
register and an authenticated GET /me both exercised against the actual
built container, not just the test suite.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

apps/api/Dockerfile

@@ -0,0 +1,31 @@
+# syntax=docker/dockerfile:1
+FROM python:3.12-slim AS builder
+
+RUN pip install --no-cache-dir uv
+
+WORKDIR /app
+COPY pyproject.toml uv.lock ./
+# Split into two syncs so dependency installation caches independently of source changes: this
+# first one installs only dependencies (--no-install-project), so touching velodrome/ doesn't
+# invalidate this layer.
+RUN uv sync --frozen --no-install-project --no-dev
+
+COPY velodrome ./velodrome
+COPY alembic ./alembic
+COPY alembic.ini ./
+# Now install the project itself into the same venv.
+RUN uv sync --frozen --no-dev
+
+FROM python:3.12-slim AS runtime
+
+RUN groupadd --system velodrome && useradd --system --gid velodrome --create-home velodrome
+WORKDIR /app
+COPY --from=builder /app /app
+ENV PATH="/app/.venv/bin:$PATH"
+USER velodrome
+
+EXPOSE 8000
+
+# Migrations run as an explicit step before this in deploy.yml (see docs/PLAN.md) — never from
+# the entrypoint, so a failed migration fails the deploy visibly instead of crash-looping here.
+CMD ["uvicorn", "velodrome.app:app", "--host", "0.0.0.0", "--port", "8000"]