[AKogut]

Closes #358

getFlakinessCost, getTestLeaderboards, getDailyTrend and getProjectHealthKpis loaded every per-test daily rollup row in their window and summed them in JavaScript. Each now asks Postgres for what it shows:

	Now
KPIs	one row of sums
daily trend	sums grouped by day
leaderboards	the top limit rows of each board, joined to their identities
cost	sums by day, from which the totals follow; the tests with reruns or flaky occurrences (the only possible offenders), grouped before joining; the quarantined identities' rollups through the primary key. Offenders are still priced and ranked in the application, with the same rounding

Numbers

packages/queries/bench (ADR-0005), p95 ms:

query	2k tests, before → after	20k tests, before → after
cost of flakiness	186 → 8	2,433 → 104
leaderboards	96 → 9	1,301 → 146
daily trend	66 → 6	586 → 61
health KPIs	40 → 8	373 → 74

All four are under the 300 ms target at 20,000 tests.

Same answers

rollup-aggregates-db.test.ts seeds 120 tests over 20 days with random counts, some quarantined, plus a noisy row in another project. It compares each function with a reference that adds the raw rows up the old way.

The first version of that test failed intermittently on cost offenders. Two tests tied on both spend and rerun time, and the reference kept them in insertion order while the query returned them by id. The old code had the same problem: ties came out in whatever order the rows arrived. Rankings now break ties on test id, explicitly in both. Eight consecutive runs pass, and the existing cost tests, the full queries suite (293) and the web suite pass.