[BenjaminMichaelis]

Summary

Follow-up to #1267. Adds the real-SQL-Server integration tests for the reading profile back, now self-provisioned with Testcontainers instead of an environment variable and a CI service container.

Stacked on #1267: the base is that PR's branch, because these tests exercise code that only exists there. Retarget to main once #1267 merges.

Why a separate PR

In #1267 these tests required ESSENTIALCSHARP_TEST_SQLSERVER and failed when it was missing in CI. The SQL Server service container and that variable existed only on the push-only "Build and test (Ubuntu)" job, so all 8 tests failed on every OS in the PR's CI matrix (8 of 540 per job, nothing else failing).

What it covers

Against a real SQL Server (SQLite serializes writers, so it cannot prove these):

• concurrent writes for a new user and for an existing user lose no updates
• the AddUserReadingProfile migration applies, reverts and reapplies; no pending model changes; expected columns
• foreign key violation (SQL error 547) and cascade delete on user removal
• reset only removes the caller's row

How it works

• Testcontainers.MsSql 4.16.0; one container per test class with a pinned image (2022-CU20-ubuntu-22.04), disposed in the class teardown.
• No env var, no workflow change.
• Skip vs fail: when no Linux container runtime can be used (local, and the macOS/Windows jobs, whose hosted runners cannot run Linux containers) the tests are skipped with a reason. On GitHub Linux CI (GITHUB_ACTIONS=true and RUNNER_OS=Linux) a failure to start the container fails the tests, so coverage cannot silently disappear.

Testing

• All 8 tests ran and passed locally against a Podman-backed container (about 60s); full suite 505 passed, 0 failed.
• Not verified: the skip path (a missing runtime could not be reproduced locally) and the first real run on GitHub Linux CI. Draft until CI confirms.

Review notes

• Linux CI adds roughly 40-60s for the image pull and container start.
• The loud-fail condition relies on RUNNER_OS; an explicit workflow env var would be the alternative.
• Off Linux CI the catch is deliberately broad, so a non-Docker startup error would skip with its message shown instead of failing.