Cut the dead weight the repo audit found: -1 dependency, -7 files, no behaviour change

[rghvgrv]

All five findings from the whole-repo over-engineering audit, in one branch. No behaviour change — 552 tests pass and the emitted migration SQL is proven byte-identical.

Closes #158 — drop the Pgvector dependency

Nothing in src/ used it; embeddings were replaced by in-database trigram matching. Ten migration files kept it alive purely through a CLR type parameter.

table.Column<Vector> / Property<Vector> → <string>. The type: "vector(1536)" string that actually reaches Postgres is untouched.

Proof the DDL is unchanged — generated before and after the change:

dotnet ef migrations script --idempotent -o before.sql   # on 928e3a7
dotnet ef migrations script --idempotent -o after.sql    # after
diff before.sql after.sql
→ IDENTICAL

Frozen history stays frozen in the sense that matters: a database that already ran these migrations is unaffected, and no new migration is needed. CREATE EXTENSION vector still runs, which is why Neon is still the right host (DEPLOYMENT.md §1).

Closes #159 — dead code and dead config

Cut	Why it was dead
IConnectivityService.ConnectivityChanged	raised, zero subscribers — view models poll IsConnected deliberately, since they're transient and the service is a singleton. Took ConnectivityService's entire IDisposable with it.
IFxRateService.GetRateAsync + ReadCachedAsync	superseded by GetRatesAsync; only a test helper called it
ExtractFirstMessage(IApiResponse)	zero callers
SuggestedLogoUrl	bound in no XAML — see the test note below
RenewalScan:UtcHour	describes a background job that no longer exists
Logging:LogLevel	never read; UseSerilog((_, cfg) => ...) discards the IConfiguration argument

SuggestedTier was left alone. It's a three-value enum driving one null check, but changing it is a behaviour question rather than dead weight, and the properties aren't cleared in lockstep — worth its own pass.

Closes #160 — three identical auth result classes → one

LoginResult, RefreshResult and ChangePasswordResult were byte-identical. Now one AuthResult record.

ResetPasswordResult stays separate on purpose: a reset issues no token pair, and folding it in would make "succeeded with no tokens" representable on the three calls where it must never happen.

Closes #161 — two single-implementation interfaces

IBurnRateCalculator and ISubscriptionMatchService sat beside their only implementation, in the same assembly, with no test fake. Concrete types are registered directly.

Every other interface in SubVora.Application stays — they cross into Infrastructure, which is what keeps EF Core out of that project.

Closes #162 — the config example that never existed

README, CLAUDE.md and docs/debug/ANDROID_DEVICE.md all tell you to copy appsettings.Development.example.json. It had never been committed — a broken first step on every fresh clone.

Values match docker-compose.yml, so the documented local loop works unedited. Jwt:Secret is blank because it has to be generated. The dev database password is the same throwaway already in tracked docker-compose.yml, audited is_secret: false in .secrets.baseline; the file scans clean.

Verification

dotnet test tests/SubVora.Api.Tests             → 147 passed
dotnet test tests/SubVora.Infrastructure.Tests  → 103 passed
dotnet test tests/SubVora.Mobile.Tests          → 264 passed
dotnet test tests/SubVora.Application.Tests     →  38 passed
dotnet build src/SubVora.Api -c Release         → 0 warnings, 0 errors
                                          total → 552 passed, 0 failed

The two Testcontainers suites are what make the Pgvector change safe to trust: every one of their test classes calls Database.MigrateAsync(), applying the full migration history from scratch against a real pgvector/pgvector:pg16 container — including the CREATE EXTENSION vector in AddSubscriptionCatalog and the drop in ReplaceCatalogEmbeddingWithTrigram. MigrationDriftTests separately asserts HasPendingModelChanges() == false, so the designer-snapshot edits did not desync the model.

.secrets.baseline is updated for line-number drift only — removing using Pgvector; shifted four existing entries by one line. No new detections; regenerated through the hook, not bypassed.

Test changes, and why they aren't a loss of coverage

Three assertions on SuggestedLogoUrl went with the property. They were proxies for "the match was adopted" — and the five SaveAsync_* tests asserting request.CatalogId already pin that far better, because they check what actually leaves the device rather than an intermediate field nothing renders. One test was renamed from AcceptingAMatch_AppliesNameCategoryAndLogoTogether to ...AppliesNameAndCategoryTogether to match what it now asserts.

FakeFxRateService.SinglePairCalls went the same way: with no single-pair path left on the interface, there is nothing for it to count. The BatchCalls == 1 assertion that carries the N+1 regression guard is untouched.