Faster, more stable flaky-alert validation under load

The server’s flaky-alert validation no longer scans every test run to determine which triggered test cases are proven on the default branch. A dedicated precomputed marker table is now used, replacing the previous per-evaluation full-table scan that was the single most CPU-expensive ClickHouse query on busy projects, reading millions of rows per evaluation and running roughly ~1.5s under contention (vs ~100ms idle). Flaky-alert evaluations should now respond more quickly and more consistently during traffic spikes, while behavior stays identical: a test case is validated exactly when it has had a successful, non-flaky run on the default branch.