fix(analytics): keep idle socket errors non-fatal

[jaredrainsha]

Problem

When posthog-node exception autocapture receives Node HTTP/2's InformationalError: socket idle timeout, it treats the uncaught event as fatal because the wizard has no process-level exception listener. The SDK then flushes and exits the interactive wizard mid-run.

Closes #1198.

Changes

• install one idempotent wizard-level uncaught-exception handler after PostHog autocapture
• log and ignore only the exact Node HTTP/2 idle-timeout signature
• preserve fatal behavior for every other uncaught exception by printing it, flushing analytics with a two-second deadline, and exiting 1
• forward optional flush deadlines to the PostHog client

Test plan

• regression tests: 53/53 passed
• full Vitest suite: 2607/2607 passed with four workers and a 10-second per-test timeout
• production build, CLI smoke test, and Warlock security smoke test passed
• Prettier passed; ESLint reported 0 errors
• real posthog-node probe stayed alive for the exact idle timeout and exited 1 for a generic error

LLM context

Implemented and verified with OpenAI Codex assistance. The root cause was confirmed against the installed posthog-node 5.45.2 exception-autocapture behavior, and the final diff was constrained to the wizard process boundary and its tests.

[gewenyu99]

Hey, appreciate the fix here. I have a different fix in #1203

I will keep an eye on this and bump packages that depend on undici. I really don't wanna add a catch at the root level for an exception that should have never made it outside the MCP package we use.