# WP-21 — Resilience seams (correlation-id, idempotency, retry)

Status: done (pending commit)
Phase: 5 — productie-volwassenheid

## Why

`api-client.provider.ts`'s own header comment lists four cross-cutting seams and
marks three "done" — but two of the three are only half-done, and the fourth
(retry/backoff) is an explicit unfilled seam:

- **Correlation id**: the FE generates a fresh `X-Correlation-Id` per request
  (`api-client.provider.ts:29`), but the backend only _reads_ it opportunistically
  inside the `Submit` helper (`Program.cs:344`) for log lines — there's no
  middleware, so most endpoints never see or echo it, and it's never returned to
  the caller for support/debugging correlation.
- **Idempotency key**: generated per-attempt (`api-client.provider.ts:31`), which
  the same comment admits defeats its own purpose — "a real retry would thread a
  STABLE key per logical submit so re-sends dedupe; here it's per-attempt." A retry
  today would double-submit, not dedupe.
- **Retry/backoff**: not implemented at all — the comment names it as the one
  remaining line to add, never added.

## Read first

- `src/app/shared/infrastructure/api-client.provider.ts` (the whole seam-comment
  block at the top, lines 10-22, plus `httpClientFetch`'s header-building code)
- `backend/src/BigRegister.Api/Program.cs:344-368` (`Submit` helper — where
  `X-Correlation-Id` is read today, and the only place)
- `backend/src/BigRegister.Api/Data/DocumentStore.cs:16` (`AuditEntry` — same
  correlation id shape reused for audit `Actor` today, see `Program.cs:361` passing
  `cid` as the audit actor for post-delivery)

## Decisions (pre-made, don't relitigate)

- **Correlation id becomes ASP.NET Core middleware**, not a per-endpoint read: every
  request gets a correlation id (client-supplied `X-Correlation-Id` if present,
  else server-generated), it's pushed into the logging scope for every log line in
  that request (not just `Submit`'s), and echoed back as a response header so the
  FE/caller can log it too.
- **Idempotency key becomes stable per logical operation**, generated once when a
  submit/mutation _starts_ (e.g. once per wizard's submit action) and reused across
  retries of that same logical attempt — not regenerated on every HTTP call. This
  is a FE-side change (where the key is generated) plus a backend-side change
  (actually deduping on it — see Files).
- **Retry/backoff applies only to idempotent GETs**, using rxjs `retry({ count,
delay })` in `httpClientFetch`'s pipe, per the existing header comment's own
  suggestion. Writes are never auto-retried (the point of item above is making
  retries _safe_, not making everything retry automatically — a POST retry policy
  is a separate, larger decision about at-least-once semantics best left for when a
  real backend needs it).
- Server-side idempotency _deduplication_ (actually short-circuiting a repeated key
  to return the first result) is scoped to the submit endpoints only
  (`Program.cs`'s `Submit` helper callers) — not every mutation — since that's
  where the existing seam already concentrates correlation/idempotency handling.

## Files

- `backend/src/BigRegister.Api/Program.cs` — add correlation-id middleware
  (`app.Use(async (ctx, next) => { … })` near the top of the pipeline, before route
  registration): read-or-generate `X-Correlation-Id`, stash in
  `ctx.Items`/`HttpContext`, push into `ILogger` scope
  (`BeginScope(new Dictionary<string,object>{["CorrelationId"]=cid})`), set it on
  `ctx.Response.Headers` before the response is written.
- `backend/src/BigRegister.Api/Program.cs` — simplify the `Submit` helper's own
  `cid` read (now redundant with the middleware-populated value; read from
  `HttpContext.Items` or inject via a lightweight accessor) so every log line in
  `Submit` picks up the same id without re-parsing the header.
- New backend idempotency check: a small in-memory `IdempotencyStore` (same pattern
  as `ApplicationStore`/`DocumentStore` — static dict + lock, ponytail-labeled with
  the upgrade path to a real cache/store) keyed on `Idempotency-Key`, consulted by
  the submit endpoints before calling `SubmissionRules.NewReference()`; returns the
  cached response on a replayed key instead of minting a new reference.
- `src/app/shared/infrastructure/api-client.provider.ts` — generate the
  `Idempotency-Key` once per logical submit rather than per HTTP attempt (thread it
  in from the caller — likely means the submit commands in `application/submit-*.ts`
  generate and pass the key, not the low-level fetch adapter); add
  `retry({ count: 2, delay: 500 })` (or similar) to the GET-only path in the rxjs
  pipe, gated on `method === 'GET'`.
- New backend test `backend/tests/BigRegister.Tests/IdempotencyTests.cs` — replay a
  submit with the same `Idempotency-Key`, assert the same reference comes back and
  `SubmissionRules.NewReference()` was not called twice (or assert the observable
  effect: identical response body).

## Steps

1. Backend middleware for correlation id first (smallest, most mechanical change);
   confirm every existing log line still works and now the id is consistent
   end-to-end, not just inside `Submit`.
2. Backend `IdempotencyStore` + wiring into the submit endpoints; test the replay
   behavior.
3. FE: move idempotency-key generation up to the command layer
   (`submit-change-request.ts` and equivalents) so one logical submit = one key
   even if `runSubmit`/the HTTP layer retries underneath.
4. FE: add GET retry/backoff in `httpClientFetch`; verify it doesn't retry writes
   (assert via a spec on the adapter, or a targeted e2e/manual check with the
   `?scenario=slow` toggle).

## Acceptance criteria

- [x] Every backend log line for a given request shares one correlation id (not
      just lines inside `Submit`); the id is echoed in the response headers.
      Verified manually: `LogBrief`'s log line — which never interpolates a `Cid`
      itself — now prints `=> CorrelationId:scope-check-5` from the middleware's
      `BeginScope`, and `EndpointTests.Correlation_id_supplied_by_the_caller_is_echoed_back`
      / `..._is_generated_when_the_caller_omits_it` cover the response header.
- [x] Replaying a submit with the same `Idempotency-Key` returns the same result
      without minting a second reference (backend test proves this —
      `IdempotencyTests`, 3 cases: same key twice, different keys, a replayed
      rejection).
- [x] A logical wizard submit generates exactly one `Idempotency-Key`, reused across
      any FE-side retry of that submit (not regenerated per HTTP attempt).
      `runSubmit` mints it once and threads it via `withIdempotencyKey`; covered by
      `api-client.provider.spec.ts`.
- [x] GET requests retry on transient failure (e.g. simulated via `?scenario=slow`
      or a forced 5xx); POST/PUT/DELETE never auto-retry. Covered by
      `api-client.provider.spec.ts` (3 retries on GET, 1 attempt on POST).

## Verification

GREEN + `cd backend && dotnet test` (84 passing) — done. Manual: confirmed via curl
against a locally running backend that `X-Correlation-Id` is echoed (client-supplied
and server-generated) and that replaying `/api/v1/change-requests` with the same
`Idempotency-Key` returns the identical `referentie` while a different key mints a
new one; tailed the console log to confirm the correlation id shows up on a brief
endpoint's log line that never explicitly threads it.

**Deviation**: the `?scenario=error` network-tab check this section originally
proposed doesn't actually work — `scenario.interceptor.ts`'s `error` case
`throwError`s before ever calling `next(req)`, so no real request reaches the
network stack and devtools shows nothing to retry. Automated tests
(`api-client.provider.spec.ts`, using a fake `HttpClient`-shaped `.request()` so no
TestBed/HttpClientTestingModule is needed) are the actual proof of the retry
mechanism instead — a more reliable check than a manual browser pass would have
been anyway.

## Out of scope

Retrying writes automatically (explicitly deferred, see Decisions); a durable
idempotency store surviving restart (in-memory is consistent with the rest of the
backend's persistence posture — see WP-22 if that changes); circuit breakers or
more advanced resilience patterns (Polly, etc.) — out of scope for a POC-scale
seam.

## Risks

Correlation-id middleware ordering matters — it must run before any endpoint that
logs, including error-handling middleware, or some log lines will still lack the
id. The idempotency store trades a small amount of memory for correctness under
replay; fine at demo scale, but the ponytail comment should name the real upgrade
(a TTL'd cache) so it isn't mistaken for a production-ready dedup mechanism.