OpenCode heap-snapshot isolation

Steps:

1. Freeze OpenCode PID, start identity, cgroup, snapshot metadata, and hashes—restarts invalidate attribution.
2. Keep the service and artifacts read-only: never signal, restart, force GC, use the network, access production data, or load snapshots wholesale in JavaScript/Python.
3. Run bounded streaming C/C++ analysis outside opencode.service with MemoryHigh≤640M, MemoryMax≤768M, and swap disabled.
4. Monitor RssAnon + VmSwap, [heap], isolate committed prefixes, 256 KiB mapping count, and cgroup anon/file/swap—RSS alone hides paging.
5. Stop all live captures once reclaim, swap, PSI, or memory.events high begins; preserve active work and use a planned restart for immediate allocator-high-water reclamation. Gotchas:

• Snapshot files can remain cgroup-charged as file cache after descriptors close; this raises memory.current, not anonymous RSS.
• Separate systemd /tmp namespaces can hide output; use a verified shared path. Verify:
• [ ] Identity and hashes remain unchanged.
• [ ] Report bounds, peak RSS, cgroup evidence, exact counts, limitations, and confidence.