V8 heap-snapshot RSS attribution

Trap: treating one V8 snapshot, a growing JavaScript cohort, or falling RSS as whole-process memory evidence looks right because each provides concrete counts. Fix: reject a cohort as the multi-GiB cause when its shallow or retained scale is orders of magnitude smaller; reconcile every isolate with RssAnon + VmSwap, allocator arenas, WASM/direct mappings, external stores, workers, and cgroup anon/file/kernel/swap. Snapshot capture can leave profiler workspace and file cache behind; an RSS fall offset by swap is paging, not release. Prove allocator retention only with isolated pre/post capture, malloc_trim(0), and a matching no-snapshot control.