v0.5.17

[0.5.17] - 2026-10-23

Fixed

• Teardown can no longer hang forever on an unobservable process tree, and the last reference is no longer dropped after the retries fail (issue #13 follow-up). tearDown() now bounds both connection.closed and waitForProcessTreeExit() with killGraceMs + shutdownTimeoutMs; a timeout becomes a failed teardown that the reap queue retains instead of a permanently pending fiber. Instances whose three reaps fail move to a failedReaps set that disposeAll() retries, runOnce() clears a previous action's idle timer before borrowing the instance (so it cannot fire under a live action), and LspConnection terminates an already-spawned handle when the provider drops a piped protocol stream. The failure log now says the process tree may still be alive.

  This is plugin-side defense and observability, not a claim that the Windows orphan root cause is fixed: the provider's job-owner SIGKILL escalation is a separate host fix (see the issue discussion).