Fix cross-run flakiness detection overcounting passing runs

Cross-run flakiness detection now flags only the failing run as flaky, not the passing runs from the same test/commit/scheme. Previously, a test that passed 19 times and failed once on a commit could be counted as 20 flaky runs, which inflated the flaky_run_count and flakiness_rate automations and could spuriously trip the auto-quarantine rule. Repetition-based flakiness (mixed retries within a single run) is unchanged. Existing data is not backfilled, so previously inflated counts will roll out of the 7-30 day evaluation window on their own.