fix: clean up timed-out bundle MCP children

[onepequity]

Summary

• close bundle MCP transports when startup times out so orphan stdio children do not survive failed handshakes
• serialize duplicate bundle MCP startup attempts by server identity to avoid concurrent races against single-writer backends
• add regression coverage and lifecycle documentation for timeout/disposal behavior

Why

A stdio MCP server that blocks during startup can time out in OpenClaw while its child process continues running. For local single-writer backends like PGLite-backed gbrain, those orphaned children can hold locks and make subsequent startup attempts worse.

Validation

• npx vitest run src/agents/pi-bundle-mcp-runtime.test.ts — 8 passed
• full commit hook attempted pnpm check, but local checkout is missing madge, so check:madge-import-cycles fails with ERR_MODULE_NOT_FOUND: Cannot find package 'madge'

[onepequity]

Closing per repo-boundary policy: Nathan-local OpenClaw runtime changes are carried through openclaw-ops vendor patches, not the OpenClaw repo. Patch queue source of truth: openclaw-ops/vendor-patches/openclaw/.