[R-3][Medium] Multi-day soak / longevity test

[v-code01]

From the review (https://github.com/v-code01/ledge/blob/main/AUDIT.md R-3). Only minutes-long soaks have run; long-run memory/stability unproven. Closes it: a multi-day soak with RSS/heap tracking under steady create/commit/GC/clone load; assert no unbounded growth. Severity: Medium.

[v-code01]

Added a longevity/memory soak harness (soak/longevity.sh, commit 0d82352) — steady workspace-churn load with RSS sampling and a warmup-skipped trend report. Bounded by default; R-3 proper is the same script run for days.

Baseline (soak/results/2026-07-02-longevity.txt): 2675 workspaces / 6 min, 0 push errors, RSS 31.7→41 MB. Investigated the ~12% growth — not an unbounded leak: per-workspace in-memory state is freed on reclaim (lease map remove on tombstone; workspace refs deleted from the ART on release), but the lease/ref WALs are append-only and only compact at 64 MiB, so a bounded run captures the pre-compaction rising edge. Definitive no-leak evidence still needs the multi-day run (crosses compaction → RSS sawtooths) + heap profiling on a real host — that's the remaining R-3 work, which needs wall-clock, not code. Harness + interpretation are in place.