feat(openzaak): one-command UI-OpenZaak bridge, real BSN fix, flake mitigation

scripts/openzaak-ui-up.sh brings up the root app and the OpenZaak harness
together, wires them onto one docker network, seeds the catalogus, grants
the container-alias zaaktype scope, and verifies a real aanvraag submitted
through the UI lands in OpenZaak.

Along the way: DocumentStore.DemoOwner was reusing the seeded doctor's
11-digit BIG-nummer as a stand-in BSN, which isn't a valid 9-digit BSN shape
— OpenZaak rejects it, breaking both submit's rol-creation step and the
citizen's own applications list under Zgw:Enabled=true. Fixed to a real
elfproef-valid BSN.

Also adds mitigation for a still-unexplained per-container flake (every
outbound ZGW POST fails as if the body were empty, for that container's
whole lifetime) that correlates with host memory pressure: the script now
warns when host swap is heavily used, and an opt-in ZgwDiagnosticHandler
(ZGW_DEBUG_HTTP=1) logs Content-Length vs. actual bytes sent so the next
reproduction can confirm or rule out client-side body corruption.

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

backend/openzaak/README.md

@@ -80,6 +80,42 @@ BFF_AUTH='<the same secret>' ./verify-notificatie.sh   # proves a real delivery,
 To go back to the fast, no-notifications default: `docker compose -f docker-compose.openzaak.yml
 up -d --remove-orphans` (drops the celery worker, restores `NOTIFICATIONS_DISABLED: 'true'`).
 
+## Testing the Angular UI against this harness
+
+The base harness above and the app's own root `docker-compose.yml` are independent projects on
+purpose (see the top of this file) — this is the opt-in bridge between them, for when you want
+to click through the real UI and see an aanvraag land in a real OpenZaak instead of just
+running `dotnet test`. One command from the repo root:
+
+```bash
+scripts/openzaak-ui-up.sh
+```
+
+It brings up the root app (so its docker network exists), brings up this harness plus
+`docker-compose.openzaak.bff.yml` (gives this harness's `web` service a dotted alias,
+`openzaak.local`, on its own network — the root project's `api` container joins THIS network,
+in the opposite direction from the notifications overlay below, to avoid a real alias
+collision: the root project's frontend service is also called `web`. The alias needs a dot
+because Django's URLValidator rejects a bare hostname in a URL field; this environment's
+rootless Podman also can't route container→host-port traffic through `host.docker.internal`,
+so container-to-container is the only reliable path either way — see that file's header
+comment for the full, empirically-confirmed reasoning), seeds the catalogus, additively
+replaces the zrc authorization grant to match the alias (ZGW authorization is scoped by the
+*exact* zaaktype URL string, not just the resource; see `scripts/openzaak-ui-up.sh`'s own
+comment for why this is a replace, not an add), and brings the root app back up pointed at
+OpenZaak (`docker-compose.openzaak.yml` at the repo root). A final self-check submits a
+throwaway aanvraag and confirms it actually lands in OpenZaak, restarting `api` (up to 5
+times) if not — see that script for a caveat about an intermittent per-container networking
+flake this environment can hit under memory pressure (the script now warns if host swap is
+already high going in; `ZGW_DEBUG_HTTP=1` on `api`, see `docker-compose.openzaak.yml`, logs
+diagnostics to help nail the cause next time it reproduces). Prints the URLs to check
+afterward and the teardown commands.
+
+Two caveats, both non-fatal (WP-60 catches and flags rather than surfacing an error):
+**only `herregistratie` has a seeded zaaktype** here, so submit that wizard to prove a real
+write; and **no Documenten content is seeded**, so a document upload's ZGW half no-ops (pick
+"per post" in the wizard's document step, or ignore it).
+
 ## Tear down
 
 ```bash

backend/openzaak/docker-compose.openzaak.bff.yml

@@ -0,0 +1,38 @@
+# Opt-in overlay, layered ON TOP of docker-compose.openzaak.yml (never alone):
+#
+#   docker compose -f docker-compose.openzaak.yml -f docker-compose.openzaak.bff.yml up -d
+#
+# Gives this harness's `web` (the OpenZaak API) an extra, dotted hostname alias
+# (`openzaak.local`) on its OWN network, so the root project's `api` container (joined in via
+# `docker-compose.openzaak.yml` at the repo root, as an EXTERNAL network) can reach it. See
+# `scripts/openzaak-ui-up.sh` for the one-command version that brings both projects up
+# together, seeds the catalogus, and grants the extra authorization scope this alias needs.
+#
+# Three real things this works around, each discovered empirically (curl against the
+# running containers), not guessed:
+#
+# 1. Why container-to-container instead of `http://localhost:8000`: this dev environment's
+#    rootless Podman drops container→host-port traffic through `host.docker.internal`
+#    (confirmed for the WP-58 notifications overlay's celery worker — DNS resolves it, every
+#    TCP connect times out).
+#
+# 2. Why the ROOT project's `api` joins INTO this project's network (below), not the other way
+#    around: the root project's frontend service is also called `web`. Docker Compose always
+#    adds a service's own name as a network alias on every network it joins — so if THIS `web`
+#    joined the root project's network, "web" would resolve to two different containers there.
+#    Only `api` crosses into this network, under its own already-unique name.
+#
+# 3. Why the alias has a dot in it (`openzaak.local`, not e.g. `openzaak`): Django's built-in
+#    URLValidator rejects a bare, dotless hostname in a URL field (it special-cases exactly
+#    "localhost"; anything else needs a dot or to be a valid IP). OpenZaak's `zaaktype` field
+#    (and others) run through this validator — confirmed with a POST referencing
+#    `http://<dotless-alias>:8000/...` failing "Voer een geldige URL in" (enter a valid URL)
+#    before any authorization check even runs.
+services:
+  web:
+    environment:
+      # Django rejects any request whose Host header isn't in ALLOWED_HOSTS.
+      ALLOWED_HOSTS: localhost,127.0.0.1,web,openzaak.local
+    networks:
+      default:
+        aliases: [openzaak.local]

backend/src/BigRegister.Api/Data/DocumentStore.cs

@@ -36,8 +36,16 @@ public sealed record AuditEntry(DateTimeOffset At, string Action, string Documen
 /// </summary>
 public static class DocumentStore
 {
-  /// The single seeded user (the demo has no real auth; ownership = this id).
-  public const string DemoOwner = "19012345601";
+  /// The single seeded user (the demo has no real auth; ownership = this id) — a real,
+  /// elfproef-valid 9-digit BSN (src/app/shared/kernel/bsn.ts's own checksum), distinct from
+  /// SeedData.Registration.BigNummer ("19012345601", 11 digits — the seeded doctor's BIG-nummer,
+  /// a different Dutch identifier scheme). Previously this constant reused that BigNummer value
+  /// as a stand-in BSN, which is invalid Dutch-BSN shape: harmless against the local store, but
+  /// a real OpenZaak instance rejects it outright — GET /api/v1/applications 500s (`inpBsn` query
+  /// filter validation) and every submit's rol-creation POST fails (`inpBsn` max_length) once
+  /// Zgw:Enabled=true. Not "111222333" or "999888777" — both already mean a different fixture
+  /// identity (the OpenZaak-harness/unit-test caller, and ApplicationTests' "other citizen").
+  public const string DemoOwner = "123456782";
 
   private static readonly object _gate = new();
 

backend/src/BigRegister.Api/Program.cs

@@ -61,9 +61,18 @@ if (zgw.Enabled)
   // WP-60: a bounded client timeout matters once ZgwHttpClient retries — without one, the
   // sources' sync-over-async call (no CancellationToken threaded through) could block a
   // thread-pool thread for HttpClient's 100s default times 3 attempts.
-  builder.Services.AddHttpClient<IZaakSource, OpenZaakZaakSource>(c => c.Timeout = TimeSpan.FromSeconds(15));
+  var zaakClientBuilder = builder.Services.AddHttpClient<IZaakSource, OpenZaakZaakSource>(c => c.Timeout = TimeSpan.FromSeconds(15));
   // WP-51: the documents (Documenten API / DRC) seam — same pattern as IZaakSource above.
-  builder.Services.AddHttpClient<IDocumentSource, OpenZaakDocumentSource>(c => c.Timeout = TimeSpan.FromSeconds(15));
+  var documentClientBuilder = builder.Services.AddHttpClient<IDocumentSource, OpenZaakDocumentSource>(c => c.Timeout = TimeSpan.FromSeconds(15));
+
+  // Opt-in diagnostic for the still-unexplained per-container flake (see
+  // scripts/openzaak-ui-up.sh's header comment) — off by default, zero cost unless set.
+  if (Environment.GetEnvironmentVariable("ZGW_DEBUG_HTTP") == "1")
+  {
+    builder.Services.AddTransient<ZgwDiagnosticHandler>();
+    zaakClientBuilder.AddHttpMessageHandler<ZgwDiagnosticHandler>();
+    documentClientBuilder.AddHttpMessageHandler<ZgwDiagnosticHandler>();
+  }
 }
 else
 {

backend/src/BigRegister.Api/Zgw/ZgwDiagnosticHandler.cs

@@ -0,0 +1,25 @@
+namespace BigRegister.Api.Zgw;
+
+/// <summary>
+/// Opt-in only (wired in <c>Program.cs</c> behind <c>ZGW_DEBUG_HTTP=1</c>) — chases the
+/// still-unexplained flake where a freshly-(re)started `api` container has every outbound ZGW
+/// POST fail with what looks like an empty body reaching OpenZaak (see
+/// <c>scripts/openzaak-ui-up.sh</c>'s header comment). Logs the one signal that would actually
+/// distinguish "client built an empty body" from "something ate it after send": the declared
+/// Content-Length vs. the byte count actually read from the request right before it goes out.
+/// A mismatch here would prove client-side corruption; agreement would point downstream instead.
+/// </summary>
+public sealed class ZgwDiagnosticHandler(ILogger<ZgwDiagnosticHandler> logger) : DelegatingHandler
+{
+  protected override async Task<HttpResponseMessage> SendAsync(HttpRequestMessage request, CancellationToken cancellationToken)
+  {
+    if (request.Content is not null)
+    {
+      var bytes = await request.Content.ReadAsByteArrayAsync(cancellationToken);
+      logger.LogInformation(
+          "ZGW {Method} {Url}: Content-Length={ContentLength} actualBytes={Actual}",
+          request.Method, request.RequestUri, request.Content.Headers.ContentLength, bytes.Length);
+    }
+    return await base.SendAsync(request, cancellationToken);
+  }
+}

backend/tests/BigRegister.Tests/ZgwDiagnosticHandlerTests.cs

@@ -0,0 +1,28 @@
+using System.Net.Http.Json;
+using BigRegister.Api.Zgw;
+using Microsoft.Extensions.Logging.Abstractions;
+
+namespace BigRegister.Tests;
+
+/// <summary>
+/// The opt-in diagnostic handler (see ZgwDiagnosticHandler's own doc comment) must be inert on
+/// the request/response — it only observes. This doesn't catch the flake itself (that needs a
+/// real repro with ZGW_DEBUG_HTTP=1), just proves the hook doesn't alter what's sent or break
+/// the response passthrough.
+/// </summary>
+public class ZgwDiagnosticHandlerTests
+{
+  [Fact]
+  public async Task Passes_request_and_response_through_unchanged()
+  {
+    var stub = new ZgwStubHandler(_ => """{ "ok": true }""");
+    var diagnostic = new ZgwDiagnosticHandler(NullLogger<ZgwDiagnosticHandler>.Instance) { InnerHandler = stub };
+    using var client = new HttpClient(diagnostic);
+
+    var response = await client.PostAsJsonAsync("https://oz.example/zaken", new { foo = "bar" });
+
+    Assert.True(response.IsSuccessStatusCode);
+    Assert.Single(stub.Requests);
+    Assert.Contains("\"foo\":\"bar\"", stub.BodyOf("https://oz.example/zaken"));
+  }
+}