arch(acl): ADR-0006 — provision the ACL integration test against the compose stack (refs #46)

Records why the integration test targets the running compose stack rather than a
Testcontainers graph (no .NET compose support; not hermetic anyway due to the
Selectielijst dependency), the opt-in publish seed, and the chunked-body bug the
test caught. Proposed in #53.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

docs/architecture/adr-0006-integration-test-provisioning.md

@@ -0,0 +1,84 @@
+# ADR-0006: Provision the ACL integration test against the compose stack
+
+- **Status:** Accepted
+- **Date:** 2026-06-29
+- **Deciders:** Respellion engineering
+- **Relates to:** S-04a (#46); proposed in #53; builds on ADR-0001 (loose coupling), ADR-0002 (catalogus design), ADR-0003 (default-fill); supports CLAUDE.md §11 (integration tests via real containers)
+
+## Context
+
+S-04 delivered the ACL's one operation — `OpenZaakGateway.OpenZaakAsync` — with unit
+tests against a stubbed `HttpMessageHandler` and a Reqnroll scenario over an in-memory
+stand-in. The deferred S-04 acceptance criterion (S-04a) is the one a stub cannot meet:
+
+> Integration test using Testcontainers against real OpenZaak passes.
+
+The test must drive the gateway against a **real** OpenZaak — real ZGW JWT auth, the real
+`POST /zaken/api/v1/zaken` contract, real CRS handling — and assert a zaak comes back.
+
+Two ways to stand OpenZaak up were considered (the issue's open question): (a) a full
+**Testcontainers** graph started by the test, or (b) target the **running compose stack**
+the repo already defines (`infra/openzaak/docker-compose.yml`, `make openzaak-up`).
+
+Investigation reversed the initially-favoured Testcontainers option:
+
+1. **Testcontainers .NET has no docker-compose support.** OpenZaak needs PostGIS + Redis +
+   a `setup_configuration` one-shot (the JWT client) + the API. Honouring "full graph" would
+   mean re-implementing that five-service stack — init ordering, the config volume, health
+   gating — by hand in C#, duplicating the maintained compose file and rotting with it. That
+   rubs against CLAUDE.md §13 ("if a test is hard to write, the design is wrong").
+2. **The test cannot be hermetic anyway.** OpenZaak's Zaken API rejects a zaak against a
+   *concept* zaaktype (`not-published`), and a *published* zaaktype requires ≥1 resultaattype,
+   which OpenZaak validates by fetching the external **Selectielijst** reference API
+   (`selectielijst.openzaak.nl`). So a real zaak POST already depends on outbound internet
+   from the OpenZaak container — the self-containment that motivated Testcontainers is lost
+   regardless of how the containers are started.
+
+## Decision
+
+**The ACL integration test targets the running compose stack; it does not start containers
+itself. No new test dependency is added.**
+
+- A gated test project `Acl.IntegrationTests` (`[Trait("Category","Integration")]`) talks to
+  OpenZaak with a plain `HttpClient`, reusing the same endpoint + JWT-client config the seed
+  uses (`OZ_BASE` / `OZ_CLIENT_ID` / `OZ_SECRET`, defaulting to the local stack). It locates
+  the published `BIG-REGISTRATIE` zaaktype via the Catalogi API and exercises the real
+  `OpenZaakGateway` against it.
+- **The lane is kept out of the fast checks.** `make unit` runs with
+  `--filter "Category!=Integration"`; Stryker is pinned to `Acl.Tests` (`test-projects`), so
+  neither the unit nor the mutation lane needs a live stack. A new `make integration` target
+  brings the stack up, seeds, runs the lane, and always tears down — mirrored by a Gitea
+  Actions `integration` job. This matches `make` being the single source of truth (ADR-0005).
+- **Publishing is opt-in in the seed.** `infra/openzaak/seed_catalogus.py` gains an
+  `OZ_PUBLISH=1` path that adds the relations OpenZaak's publish requires — two statustypen
+  (begin/eind), a roltype, and a resultaattype whose Selectielijst procestype is matched onto
+  the zaaktype — then publishes. The default seed (S-01 / ADR-0002) still leaves the zaaktype
+  a concept; only `make integration` flips the switch.
+
+## Consequences
+
+- **Positive:** a small, honest test over the real ZGW contract with no bespoke orchestration
+  to maintain; the compose stack is exercised exactly as operators run it; no new dependency.
+- **It caught a real bug.** The gateway sent the zaak body via `JsonContent` without a
+  `Content-Length`, so .NET framed it as `Transfer-Encoding: chunked`, which OpenZaak's uwsgi
+  rejects with 400. A stubbed handler accepts either framing, so only a real OpenZaak surfaced
+  it. Fixed by buffering the body (`LoadIntoBufferAsync`); guarded in the fast lane by a unit
+  test asserting a `Content-Length` is set. This is the concrete justification for §11's
+  integration tier.
+- **External dependency:** the integration job needs the OpenZaak container to reach
+  `selectielijst.openzaak.nl`. It is a stable public reference API (the same one OpenZaak uses
+  in production) but it is a network touchpoint, and a CI environment without egress would need
+  a local Selectielijst service or a recorded fixture. `OZ_SELECTIELIJST` overrides the base URL.
+- **Cost:** the lane needs the stack up first; CI runs it as a dedicated job (Docker + dotnet +
+  python3), separate from the fast lanes.
+
+## Alternatives considered
+
+- **Full Testcontainers graph** — rejected: re-implements the compose stack in C# (brittle,
+  duplicative) for no hermeticity gain, since the Selectielijst dependency remains.
+- **Single OpenZaak container (sqlite/locmem)** — rejected: diverges from the real
+  PostGIS-backed, Redis-cached deployment; the Zaken API is a geo API and the divergence would
+  undermine the contract the test exists to verify.
+- **Mock OpenZaak / record-replay** — rejected: that is what the existing stubbed-handler unit
+  tests already do; it cannot exercise the real contract, and would not have caught the chunked
+  body bug.