fix(freshness): output-floor on cycle-SUM over fresh_days window (#1945, #1947)
Adversarial review found the single-newest last_success_work basis fragile:
a load cycle = MULTIPLE done-runs (objective per group_name, kn per region +
resume), so the newest run alone misrepresents the cycle. Replace with
SUM(work_col) FILTER(done) over the source's own fresh_days window.
- recent_output = SUM(COALESCE(work_col,0)) FILTER done in fresh_days window
(make_interval secs => fresh_days*86400; no :: cast; NULL-counter -> 0)
- downgrade when status==ok and recent_output < min_output_rows (strict <)
- registry floors kept: kn_flats=50000 (healthy ~376k, broken sum <=3670),
objective=1000 (SUM 7d ~946k); comments rewritten to cycle-sum basis +
kn_flats zombie-resume KNOWN LIMITATION
- window=fresh_days keeps the #1947 aging-FP fix (7-8d run stays in window)
- tests: low-output-cycle->failed, above-floor->ok, boundary==floor->ok,
weekly-aging->ok, age-stale precedence, registry floors
Verified on prod: domrf_kn_flats latest snapshot 2026-06-22=9 flats vs healthy
2026-05-17=376604 (broken 5+ weeks); objective SUM(rows_lots,7d)=946264.
Light1YT
c646a71001