Async context entry failures leave earlier connection contexts unclosed #97

MCPAdapt.__aenter__ acquires connections into an AsyncExitStack before adapting the tools. If a later connection or tool adaptation raises, context entry fails and Python does not call __aexit__, leaving the already-entered connection contexts without their normal cleanup.

I reproduced this locally using two synthetic MCP stdio servers followed by a failing setup step. Both failure cases miss the expected reverse-order exits on the current main revision (538cd856); a successful entry/exit control passes. A focused local fix keeps the stack in an async with block during setup and transfers it with pop_all() only after adaptation succeeds. All three regressions then pass, and the full suite passed with 48 tests and one optional audio skip. A separate AI reviewer checked the patch and additional cleanup controls.

This overlaps the __aenter__ code touched by #76, but concerns cleanup when entry fails rather than adding fault-tolerance behavior.

Before opening a PR, I checked the documented 95% coverage requirement. An isolated run using the Makefile's default collection scope reports 86% overall; production-only coverage is 77%, versus 76% on the unchanged baseline. That coverage run had 47 passes and two skips (optional audio and an existing HTTP fixture startup timeout). The changed entry path is covered. Would you prefer a focused cleanup PR with those reported coverage results documented, or separate coverage work first?

Prepared with AI assistance. Testing used local synthetic processes and temporary data with external networking disabled.