Stress: RSS memory tracking in YCSB heavy run — assert growth stays bounded

Evidence: Three production instances swapping after weeks of sustained load (CDN edge cache 533/537 MB swap, global prerender 536/537 MB swap, redirect cluster 353/512 MB swap). Inject RSS sampling every 30s into YCSB heavy scale run; assert RSS at end is within N% of post-warmup baseline. Confluence: Section 5 ST-6

## Existing coverage in harper-pro
The `integrationTests/stress/` suite in harper-pro already covers several memory-under-load scenarios:
- `backlogRecovery.test.mjs` — asserts per-peer backlog memory is capped and catch-up RSS stays bounded
- `slowConsumerBackpressure.test.mjs` — checks RSS stays under 1500 MB under sustained write churn

This issue should focus on the YCSB-level regression gate (sampling RSS during the standard 1M-record benchmark run, not cluster-level tests) to catch step-function leaks in the write path.