fix: close finalized LCM engine clones

[bennybuoy]

Summary

• register a supported on_session_finalize plugin hook
• resolve the exact per-agent LCM clone by durable session ID
• shut down only that finalized clone, leaving other active clones and the plugin prototype untouched
• preserve interactive CLI /new, which reuses its agent/context engine

Why

Hermes deep-copies the registered context-engine prototype for each AIAgent. In long-lived Desktop and gateway processes, retiring an agent does not currently guarantee that the clone's SQLite-backed LCM helpers are closed. Finalized sessions can therefore leave Store, DAG, and lifecycle descriptors open.

The plugin already owns an exact weak session-to-engine registry and an idempotent shutdown() boundary. This change connects those existing pieces through Hermes' supported hard session-boundary hook; it does not patch Hermes core or sweep unrelated engines.

Verification

• focused plugin lifecycle, hook registration, and packaging matrix: 33 passed
• 12-cycle Linux FD soak while retaining references to every finalized clone: post-warm-up LCM FD count returned from 13 to the steady baseline of 10 after every cycle
• Ruff: passed
• Python compilation: passed
• git diff --check: passed

Scope

This closes plugin-owned LCM SQLite resources. It does not claim to close Hermes-owned state.db connections.

Automated review

GitHub Copilot reviewed the identical diff in bennybuoy/hermes-lcm#24, because fork contributors cannot request reviewers on the upstream repository. Its initial test-helper suggestion was addressed in 085d6ee; a second review of that commit covered both changed files and generated no new comments.

[bennybuoy]

Superseded by #528, refreshed onto current main with updated lifecycle/packaging coverage and explicit compatibility validation against #501.