fix(mcp): restore live connector preservation and bounded shutdown

[breedx]

Refreshing MCP configuration should not replace live connectors, and stopping a healthy connector should not deadlock on its own cleanup lock. This restores the narrow preservation and bounded-shutdown fix from #924 after its revert in #931.

Resubmission

GitHub cannot reopen a merged PR, so this is a new reviewable submission against current main (007c53b2, v0.0.833).

• Exactly the same three-file patch as fix(mcp): preserve live connectors and bound shutdown waiting #924, reapplied as one commit; matching stable patch IDs and byte-identical changed files verified.
• No additional fixes, dependency changes, version changes, or integration history.
• fix(mcp): await restart before rebinding the requesting agent #930's draft restart workflow depends on this unlocked shutdown contract; it remains separate.
• The independent stream callback cleanup from fix(stream): close callback coroutines when task submission fails #929 is already present in this base and is not part of this diff.

Change

• Preserve existing connector object identity and status when materializing newly discovered configuration.
• Cancel registered lifecycle tasks under the registry lock, then drain outside the lock with a five-second deadline.
• Preserve the cleanup task on timeout, repeated stop, and caller cancellation; retain registration after failed cleanup.
• Drain registered contexts concurrently through one stable, deduplicated shutdown callback.
• Keep detailed exception diagnostics at DEBUG with generic cleanup warnings.

Validation

Current upstream lock installed in a dedicated environment: Linux, Python 3.13.13, core plugins 0.0.45, Pydantic AI 2.35.0, MCP 1.27.1, and fastmcp-slim 3.4.7.

• pytest -q -o addopts= tests/mcp/test_shutdown_preservation.py: 9 passed, including real local stdio transport cleanup.
• Changed-file Ruff lint and format checks passed; git diff --check passed.
• Patch identity and public-only diff checks passed.
• Broad regression selection (tests/mcp tests/agents tests/tools tests/command_line/mcp tests/test_completions_and_small_modules.py tests/test_claude_refresh_review.py): 1,898 passed, 7 skipped, 1 warning.
• The warning is an unawaited AsyncMockMixin._execute_mock_call in the browser control tests. Running tests/tools/browser/test_browser_control.py independently on both this candidate and unpatched 007c53b2 reproduces the same warning: 26 passed, 1 warning in each lane. It was not suppressed.
• Public CI: quality and Windows encoding passed; macOS suite 1 failed, 7,854 passed, 14 skipped, 27 warnings. The sole failure is tests/command_line/test_autosave_menu.py::TestSessionBrowser::test_open_project_and_select_session expecting TODAY from entries timestamped one/two hours before datetime.now() just after midnight. The same isolated assertion fails on both unpatched 007c53b2 and this candidate. This patch changes neither the browser nor that test; no clock-dependent rerun or assertion suppression was used.
• The CI job is labelled Python 3.13, but its paths show Python 3.14.7; local checks above used 3.13.13. CI additionally reports deprecation/resource/coroutine warnings. Those logs were inspected, but only the browser warning above has a new independent local base control here; this is not an all-green or merge-ready claim.

Local tests use disposable HOME/XDG, no inherited credentials, and blocked socket connect/DNS/bind. Prior bug-regression baseline controls and independent review are documented in #924; those historical counts are not being presented as new results.

Boundaries

This does not fix restart/reload/remove ordering, pending startup ownership, late startup registration, manager/CLI completion reporting, or completion-time agent rebinding. Connector-entry containment is separate. The deadline bounds caller waiting, not process termination; operations remain on their owning event loop. No release or deployment qualification is claimed.