[AKogut]

Closes #357

Why the counts can live on the run

A run is exactly one ingest batch: the processor upserts it by idempotency key and replaces its executions by ordinal, deleting any left over. So a run's counts are the statuses of the batch that last wrote it. The worker computes them from the batch and sets them in the same run.upsert (create and update). A redelivery therefore replaces the counts. Identity merges and splits move executions between identities, never between runs, so they cannot change a run's counts.

Changes

• run gains passed_count, failed_count, skipped_count and flaky_count; the migration backfills them from existing executions.
• The processor sets them from the batch.
• listRuns reads them instead of groupBy over test_execution.
• getRun counts the executions it already loads, one query fewer.
• @flakemetry/db exports countStatuses and runCountFields, used by the processor, and refreshRunCounts, which recounts a project's runs. The demo seed and test fixtures that write executions directly use it.
• docs/configuration.md: a run whose raw executions retention has pruned keeps its counts on the runs list, where it used to show zeros. Its detail page no longer lists executions, as before.

Verified

• Benchmark (ADR-0005), runs list p95: 4,760 → 1.9 ms at 20,000 tests, 23.5 → 0.9 ms at 2,000. Run detail 141 → 123 ms at 20,000 tests.

• Migration on the benchmark databases: 1 s at 1.5 M executions, 4 s at 15 M. 0 runs differ from a recount by status.

• Processor tests:

  • counts equal the stored executions' statuses;
  • redelivering the same run twice leaves them unchanged, and redelivering it with different results replaces them.

  Dropping the counts from the upsert's create fails the first test, and dropping them from update fails the second.

• pnpm demo seed: 24 runs, counts summing to all 201 executions, no mismatch.

• The read-paths listRuns test named "with counts" never checked them; it now does.

• Suites: queries 289, worker 125, api 121, db 19 and web 47 pass with the database; lint and typecheck pass; schema drift clean.