TestWebApplicationFactory instances shared the same in-memory database
because EF Core's UseInMemoryDatabase("InMemory") returns the same
backing store to every DbContext that asks for it under the same
connection string, in the same process. Whichever fixture started
first defined the state, and every subsequent fixture inherited it,
making tests silently order-dependent and flaky.
Fix:
- Yavsc.Tests.Shared/InMemoryDatabaseName: helper that suffixes the
in-memory connection string with a per-fixture GUID.
- TestWebApplicationFactory: instance GUID + ConnectionStrings__
YavscConnection set as an environment variable in the constructor
and cleared in Dispose, so each factory gets its own backing store.
Env var is needed because IdentityServer8.EntityFramework exposes
ConfigureDbContext as Action<DbContextOptionsBuilder> with no
service-provider access, so the connection string is captured at
registration time. AddEnvironmentVariables is the last provider in
the config pipeline and wins regardless.
- WebServerFixture: process-static GUID (WebHostFixture is a
per-process singleton by design, so the test collection shares one
store; the GUID still isolates from TestWebApplicationFactory).
- AddIdentityDBAndStores: read the connection string at DbContext
construction time via the (sp, options) overload of AddDbContext,
so test fixtures can override it via the host's IConfiguration.
IdentityServer stores cannot do the same without subclassing the
framework's DbContexts; the env var path is the documented escape
hatch in HostingExtensions.AddIdentityServer.
- UsesInMemoryProvider: StartsWith instead of equality, so
'InMemory-{guid}' is still recognised as an in-memory connection
string.
Regression sentinel in
Controllers/TestWebApplicationFactoryIsolationTests: two factories
seed a marker client in the first, the second must not see it.
Suite: 45/45 over 3 stable runs, 13-15s each.
The bool Comment on CircleAuthorizationToBlogPost was dead code:
never read or written by any caller in src/, no UI exposure, no
behavioural semantics. The wire DTO (CircleAuthorization in
Yavsc.Abstract) doesn't carry it, no reader consumes it, and the
PostIt client builds its payload without it.
What changes:
- src/Yavsc.Server/Models/Access/CircleAuthorizationToBlogPost.cs:
remove the property.
- src/Yavsc.Blogs.Tests/BlogAclApiTests.cs: drop 'Comment = true'
from the existing test payload and trim the now-inaccurate XML
doc comment ('CircleId + BlogPostId + Comment' -> 'CircleId +
BlogPostId'). Also adds a new [Fact] pinning the prod bug
reported on 2026-08-21 (HTTP 500 'BlogPostId is unknown' when
PostIt POSTs the bare { circleId } shape). That test stays red:
the real fix for the 500 is in PostIt (payload needs blogPostId)
+ on the wire DTO + server-side validation, and lives in a
follow-up commit.
Migration:
- src/Yavsc.Org/Migrations/20260820232152_DropCommentFromCircleAuthorizationToBlogPost
drops the boolean 'Comment' column on CircleAuthorizationToBlogPost.
The generated scaffold also wanted to drop three 'ClientId1'
shadow FK columns on ClientScopes / ClientRedirectUris /
ClientGrantTypes (from leftover HasOne<Client>() overrides in
ApplicationDbContext.OnModelCreating); those were removed from
the .cs to keep the migration scoped to this fix. Cleaning up the
shadow property declarations themselves is left as a separate
task.
The ModelSnapshot still reflects the shadow 'ClientId1' columns
intentionally: they exist in the prod database today (all NULL),
and EF will rescaffold a drop migration for them on the next
'migrations add' regardless. No data loss.
LoadClientAsync(id) on ClientController used to trip an
IndexOutOfRangeException at the InMemory shaper for any
.Include() of one of three Client navs: RedirectUris,
AllowedScopes, AllowedGrantTypes. Five other Client navs (with
the same EF shape and the same application-level config) worked
fine.
Bisection pointed at the InMemory provider; that hypothesis was
wrong. The real cause is in ApplicationDbContext.OnModelCreating:
yavsc was redeclaring the HasOne<Client>().WithMany(...).
HasForeignKey(e => e.ClientId) for all eight Client* navs. The
same relation is already declared (more completely, with
.IsRequired().OnDelete(DeleteBehavior.Cascade)) by
IdentityServer8's ConfigureClientStore via ModelBuilderExtensions.
The redundant mapping on three specific entities — ClientScope,
ClientRedirectUri, ClientGrantType — interacts with the InMemory
provider's shaper in a way that throws IndexOutOfRange. Removing
the redundancy fixes it.
This commit also walks back b12c272d:
- Drops .AsSplitQuery() from LoadClientAsync (no longer needed
for the InMemory shaper, and the Postgres path it was a hedge
against was a false alarm — there is no Postgres production
bug here, only an InMemory shaper quirk surfaced by the
redundant mapping).
- Removes the 9 Bisect_*_alone tests that were the artefact of
the provider-hypothesis phase. They pointed at the right
entities but for the wrong reason.
- Keeps EditRedirectUris_GET_after_add_lists_both_uris as the
end-to-end regression sentinel: with the fix in place, it
loads a Client with two RedirectUris and asserts both are
rendered. Without the fix, it fails with IndexOutOfRange.
LoadClientAsync chains 9 .Include() calls on dbContext.Clients.
On Postgres (and InMemory for some IdentityServer8 nav types), the
resulting cartesian product trips the query shaper with
IndexOutOfRangeException at IncludeCollection materialisation time.
Bug reproduces in production on the Blog admin pages that load a
Client by id.
AsSplitQuery() rewrites the load as 9 separate SELECTs joined by
client id, which sidesteps the cartesian explosion and any shaper
ambiguity between Claims/Properties/ClientSecrets (which share
Type/Value column names across some IdentityServer8 versions).
Tests:
- EditRedirectUris_GET_after_add_lists_both_uris: end-to-end
reproducer that adds a second RedirectUri via POST then re-GETs
the editor. Guards the fix on the integration path.
- Bisect_*_alone: nine unit tests that exercise the same
.SingleOrDefaultAsync(c => c.Id == id).Include(nav) on the
InMemory provider, one nav at a time. Pinpointed three
problematic navs (RedirectUris, AllowedScopes, AllowedGrantTypes)
on InMemory; kept as a regression net for any future shaper
regressions on the InMemory provider (not the Postgres path).
TestWebApplicationFactory used ASPNETCORE_ENVIRONMENT=Development, which
caused Program.Main's AddConfiguration("org") to load the tracked
appsettings-org.json (the reference file with the
'*** via dotnet user-secrets ou variable d'environnement ***'
placeholder connection string). Npgsql then failed to parse that
placeholder during host startup, failing six integration tests
(observed 2026-07-11: System.ArgumentException on
NpgsqlConnectionStringBuilder.set_Item).
Switching the test host to a dedicated Testing environment makes
AddConfiguration("org") pick up the new optional
appsettings-org.Testing.json file as the last source in the chain
(JSON → env vars), which overrides YavscConnection with the
InMemory marker and the Smtp section with the test stub values.
The .gitignore exception whitelists this file explicitly: it is a
configuration source for the test host, not a secrets file.
The WebServerFixture path is unchanged — it owns its
WebApplicationBuilder and adds the same in-memory override via its
BuildApp hook.
GET /BlogSpot/Details/1 retournait 500 (avec un 500-sur-500 sur
la page d'erreur elle-même) parce que DisplayTemplates
/ApplicationUser.cshtml faisait `var avuri = "/Avatars/" +
Model.UserName + ".s.png"` : avec <Nullable>enable</Nullable>,
Razor émet un null-check implicite sur Model.UserName et lève
NullReferenceException quand l'auteur n'a pas de UserName
posé (donnée héritée, user partiellement initialisé).
- UserDisplayHelpers.AvatarSrc : helper statique pur dans
Yavsc.Abstract.Identity qui retourne
YavscConstants.DefaultAvatar pour user null / UserName vide
ou whitespace, et un path /avatars/<name>.s.png sinon.
- ApplicationUser.cshtml : utilise le helper.
- Tests : 4 cas (null, vide, whitespace, valide) dans
Yavsc.Org.Tests/NonRegression.
Bonus : le path d'avatar passe de "/Avatars/" (S majuscule,
ne résolvait pas dans le middleware de fichiers statiques) à
YavscConstants.AvatarsPath ("/avatars" minuscule), pour fermer
l'autre trou que centraliser le calcul permettait de fixer
proprement.
Tout billet a un auteur, tout commentaire a un auteur. On aligne la
base sur ce contrat (Postgres) en droppant les orphelins existants
puis en remplaçant les FK en cascade par des FK Restrict.
- ApplicationDbContext: fluent pour BlogPost.Author et Comment.Author
en DeleteBehavior.Restrict.
- ApplicationUser: ajoute la nav inverse BlogComments (manquait,
EF aurait sinon créé une shadow FK).
- Migration 20260711173717_EnforceBlogAuthorFKs: Up purge les
Comment/BlogSpot dont l'AuthorId n'existe plus, log le volume,
puis drop+add des FK. Down laisse la cascade (état pré-migration).
Le code applicatif (BlogSpotService.Details) s'appuiera sur cette
contrainte dans un commit séparé.
Extract the kid calculation out of LoadSigningCredentialsInner
into a new internal static HostingExtensions.ComputeKid(string),
and cover it with five focused unit tests in
Yavsc.Org.Tests.ComputeKidTests.
The kid is the bit of signing-credential metadata that ties a
JWT to the right key in the JWKS. Without it, resource servers
(Yavsc.Blogs, Yavsc.Api) fail signature validation with IDX10500
'The signature key was not found', as fixed in 2c6d1157. That fix
inlined three lines of thumbprint-truncation logic at the top of
LoadSigningCredentialsInner, but left the calculation untested.
The tests in this commit pin its shape, value, stability, and
uniqueness, so a future refactor (e.g. switching from SHA-1 to
SHA-256, or moving to X509CertificateLoader for SYSLIB0057) has
to update them deliberately instead of silently changing the
JWKS key id.
Concretely:
- InternalsVisibleTo("Yavsc.Org.Tests") in AssemblyInfo.cs
gives the test project access to the new internal method
without forcing LoadSigningCredentialsInner to leak
further.
- ComputeKid(string) is the single source of truth for the
16-hex truncation; the production call site in
LoadSigningCredentialsInner now reads
'var kid = ComputeKid(certPath);'.
- The inline comment block is updated to say SHA-1 (which is
what X509Certificate2.GetCertHash() actually returns) instead
of the previous SHA-256 claim. The behaviour is unchanged.
- ComputeKid uses X509CertificateLoader.LoadCertificateFromFile
rather than the obsolete 'new X509Certificate2(string)' ctor
(SYSLIB0057); same on-disk behaviour, no obsolete warning.
Tests cover:
- 16-char upper-case hex output matching the first 16 hex
chars of the cert's GetCertHash();
- stability across repeated reads of the same cert;
- distinctness between two independently generated certs;
- the SHA-1 size of the underlying thumbprint (20 bytes), so
a future switch to SHA-256 forces a test update;
- CryptographicException propagation for a missing cert file
(Assert.ThrowsAny to stay portable across the Linux OpenSSL
and Windows leaf exception types).
IdentityServer8 was emitting JWTs without a 'kid' header and
serving the JWKS without per-key identifiers, because
LoadSigningCredentialsInner constructed RsaSecurityKey /
ECDsaSecurityKey objects without an explicit KeyId. Resource
servers (Yavsc.Blogs, Yavsc.Api) cannot match a token to a key
in the JWKS without one, so every signature validation failed
with 'The signature key was not found' (Microsoft.IdentityModel
IDX10500).
Root cause: SigningCredentials were built directly from the
BC-parsed key parameters, bypassing the X509Certificate2 path
IdentityServer normally derives the kid from. The fix derives
a stable KeyId from the certificate's SHA-256 thumbprint
(truncated to 16 hex chars) and sets it on both SecurityKey
variants before constructing SigningCredentials.
The thumbprint-based kid is stable across process restarts as
long as the cert doesn't change, and changes naturally on
LetsEncrypt renewal (~90 days), which is the right behaviour:
old tokens age out, resource servers refresh their JWKS cache
to discover the new kid.
Production rollout: redeploy Yavsc.Org and re-login (or let
the refresh-token path rotate) so newly issued tokens carry
the kid. Pre-restart tokens will continue to be rejected with
IDX10500 until they expire or are refreshed.
