Applications, documents (+ audit log) and the brief move off static in-memory
Dictionaries onto a real SQLite file via EF Core, so demo data survives a
process restart or `docker compose restart api` for the first time. The three
stores (ApplicationStore/DocumentStore/BriefStore) keep their exact public
signatures and static-class shape — no DI, no async ripple into Program.cs's
minimal-API handlers — each method just opens a short-lived AppDbContext via
Db.Create() under the same lock it already had. Opaque nested shapes (a
wizard's draft snapshot, a brief's sections/placeholders/status) are stored as
JSON text columns rather than redesigned into relational tables, matching the
existing "don't interpret it" posture.
Found two things the WP's own text got wrong, corrected in
docs/backlog/WP-22-durable-persistence.md's Deviations section: SeedData never
seeded these three stores (only the read-only BRP/DUO-mimicking GETs, which
stay in-memory) so there's no seed step; and no new docker-compose volume is
needed since the existing bind mount already covers the SQLite file — verified
against this environment's real podman-backed compose stack, not just by
reading the file.
Also: pinned SQLitePCLRaw.bundle_e_sqlite3 to 3.0.3 (EF Core Sqlite's own
transitive default bundles a pre-3.50.2 SQLite with a known high-severity
memory-corruption advisory); found and fixed a real xUnit test race where
concurrent test-class hosts stomped a shared static connection-string field,
fixed by disabling cross-class test parallelization rather than adding DI the
stores don't otherwise need.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Correlation id becomes real ASP.NET Core middleware instead of a per-endpoint
read: every request gets one (client-supplied or generated), it's echoed as
an X-Correlation-Id response header, and pushed into the logging scope so
every log line for that request carries it — not just the Submit helper's,
verified against LogBrief which never threads it explicitly.
Idempotency-Key moves from per-HTTP-attempt (defeating its own purpose) to
per-logical-submit: runSubmit mints one key and threads it through a small
bridge (withIdempotencyKey/currentIdempotencyKey) since the NSwag-generated
client has no per-call header hook. Backend gains an IdempotencyStore that
short-circuits a replayed key to the first call's result instead of minting
a second reference — scoped to the Submit-helper endpoints per the WP's own
decision.
GET requests now retry transient failures (rxjs retry({count:2, delay:500}));
writes never auto-retry. Proven with a fake-HttpClient spec
(api-client.provider.spec.ts) rather than a manual network-tab check — the
WP's suggested `?scenario=error` check turned out not to exercise a real
network call at all (the interceptor throws before calling next()), so the
automated test is the actual proof.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
