[LurigeLars]

Local MCP discovery and reconnections launch short-lived proxy processes. Each process previously created and removed its own Docker backend, causing repeated container creation and teardown even when no GitHub tool was being used.

Reuse one local backend per pinned image and non-secret configuration. Concurrent starters resolve the same container, client exit leaves it running, and a stopped backend is started again without recreation. A separate label keeps legacy PID-based orphan cleanup from removing the shared backend. The DPAPI credential path, per-request authentication, loopback binding, container hardening, tool policy and cloud gateway remain unchanged.

Validation:

• 35 static security and local forwarding/lifecycle unit tests passed.
• A real Docker regression passed with four concurrent client processes, subsequent reconnection, client exit, backend stop/restart and MCP tools/list.
• The pinned upstream write-tool inventory regression and PowerShell syntax checks passed.
• Local Gateway Security CI now runs the real Docker lifecycle regression.

Previous image/config backends require explicit retirement after their clients have exited. This change prevents Docker churn from short-lived clients; it does not claim to fix or identify Codex's internal refresh trigger.