fix(integrations): prevent Hermes bm mcp process leaks

[a-m-brewer-marve]

Summary

• reuse a process-local BasicMemoryProvider singleton during Hermes plugin registration
• shut down any existing _BmMcpActor before provider reinitialization
• add regression tests for repeated register() and repeated initialize() lifecycles

Why

Fixes/mitigates #1017. In long-running Hermes gateway/dashboard processes, repeated plugin registration or provider initialization can otherwise create multiple bm mcp subprocesses that stay alive and retain memory.

Tests

• uv run --with pytest pytest integrations/hermes/tests/test_provider.py integrations/hermes/tests/test_commands.py -q
• uv run --with pytest pytest integrations/hermes/tests -q
• uv run ruff check integrations/hermes/__init__.py integrations/hermes/tests/conftest.py integrations/hermes/tests/test_provider.py
• uv run ruff format --check integrations/hermes/tests/conftest.py integrations/hermes/tests/test_provider.py

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 320f6c68ef

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Reset per-session state before reusing the singleton

When Hermes starts another session in the same process, this singleton can be initialized again after the prior session has captured turns. shutdown() only clears _actor and _initialized, while _session_note_id and _first_user_msg remain set; the next session's first sync_turn() will therefore take the append path in _capture_turn() and write into the previous session transcript, and its summary can link/use the previous session state. Please clear the session-scoped fields when reusing the provider for a new session.

Useful? React with 👍 / 👎.

[phernandez]

Thanks for the contribution. Two contributor gates are still blocking this PR:

• CLA: Please complete or re-authorize the contributor agreement so the CLA check turns green.
• DCO: Every commit needs a Signed-off-by trailer. Please sign off the commits (for example, by rebasing with --signoff) and update the branch.

Once both checks are green, maintainers can take over any remaining implementation changes needed to get this merged.

[phernandez]

Thank you for digging into the repeated Hermes initialization leak and for adding lifecycle tests—the diagnosis was useful. We have now merged #1059, which fixes #1017 by reusing the provider actor across repeated initialization while explicitly resetting per-session capture state and guaranteeing actor/thread/child-process cleanup on failed startup and shutdown. We chose that lifecycle-scoped approach instead of a process-wide provider singleton because the singleton could leak provider configuration and session state across independent registrations. Since the issue is now fixed and tested in the merged implementation, I am closing this PR as superseded. We appreciate the contribution.