fix(session-manager): chown new session dirs when host runs as root

[netadmincmh-hash]

Summary

Linux installs that run NanoClaw as root with the data directory on a network filesystem hit an unrecoverable container spawn loop. Two constraints collide:

1. The agent image ships with USER node (uid 1000) and Claude Code refuses to run as root:

   --dangerously-skip-permissions cannot be used with root/sudo privileges
   

   So --user 0:0 on docker run is not a workaround.

2. The host process writes inbound.db and the session-folder scaffolding as uid 0. On a network filesystem (NFS in my case) the container's uid 1000 cannot write outbound.db or touch the heartbeat file. bun:sqlite surfaces this as:

   Fatal error: attempt to write a readonly database
   

The container exits code=1 microseconds after agent-runner startup, the host sweep retries a few times, then marks the inbound message completed without ever sending a response.

What this changes

initSessionFolder chowns each freshly-created session directory to 1000:1000 when process.getuid() === 0. No-op for any non-root host UID — those paths fall through to the existing --user $hostUid:$hostGid mapping in container-runner.ts. execFileSync('chown', ...) is best-effort: if it fails, the agent fails later with the clearer SQLite error and the sweep retries until the operator notices.

Test plan

• pnpm run build passes (no new deps, no type changes).
• On the affected host (root + NFS-backed /pods), reproduced the spawn loop on main. With this patch applied, chown -R 1000:1000 <session-dir> runs at session create, the container's node user can write outbound.db, and Telegram round-trip completes (verified end-to-end).
• Doesn't run when host UID != 0, so no behavior change for the normal Mac/Linux-as-non-root install.

Reproducer

# On a host running as root, with /pods on NFS:
sudo -i
cd /pods/nanoclaw-v2
systemctl start nanoclaw-v2-<slug>
# Send any inbound to the bot. Container exits ~370ms after spawn.
journalctl -u nanoclaw-v2-<slug> | grep "readonly database"

Notes / things to discuss

• The hard-coded 1000:1000 matches the image's USER node directive but isn't future-proof if the image ever changes UID. If you'd prefer, this could be made configurable via env var (e.g. NANOCLAW_CONTAINER_UID:GID) or read from the image at startup. Happy to fold in either approach.
• Existing session dirs created before the patch is applied won't be chowned automatically — operators on affected setups will need a one-time chown -R 1000:1000 data/v2-sessions/.
• This was discovered finishing a v1→v2 migration; full incident notes in the migration record (separate from this PR).

🤖 Generated with Claude Code