Rolling flaky and reliability alerts no longer double-count re-inserted runs

Fixed a bug in the rolling flaky-test and reliability alert evaluation that caused healthy tests to be incorrectly auto-skipped or muted. The aggregation was counting re-inserted runs multiple times in the window, which inflated flakiness rates and deflated reliability rates exactly for the runs a threshold reacts to. In one customer case this led to 615 genuinely reliable tests being auto-skipped. Each run is now counted once per window, so flakiness_rate, flaky_run_count, and reliability_rate evaluations reflect actual run history rather than duplicate inserts. This affects all rolling-window flaky-test monitors (default window of 100 reads).