Improve CLI error telemetry

[zahlekhan]

Summary

• add bounded failure_stage, error_class, and error_code properties to existing CLI failure events
• attach an ephemeral cli_run_id and command to every event from one invocation
• classify dependency failures across peer conflicts, registry access, network failures, and install scripts
• identify lockfile/workspace configuration and package engine/version compatibility failures
• attach bounded process metadata (duration_ms, exit_code, and failure_signal) to skill-install, dependency-install, and dev-server processes
• distinguish Cloud-auth method, OIDC, user-info, organization, key-mint, and HTTP failures
• emit cancellation events separately for prompts and interrupted child processes
• keep commands/generate.ts unchanged; the shared top-level handler emits bounded fields without raw messages

Why

The CLI already had useful command and stage-specific events, but top-level create failures discarded CreateError.stage, dependency and auth failures lacked actionable categories, process failures lacked consistent exit metadata, cancellations could look like failures, and generate failures sent truncated raw messages that could contain paths.

This patch extends the existing event model. A compact process runner streams output unchanged while retaining a bounded 16 KiB diagnostic tail locally for allowlisted classification; that diagnostic text is never emitted.

Validation

• pnpm --filter @openuidev/cli build:cli
• pnpm --filter @openuidev/cli lint:check
• pnpm --filter @openuidev/cli format:check
• git diff --check

Notes

• Existing event names and funnel ordering are preserved.
• Error messages, paths, stack traces, command output, and generated output are not sent.
• The implementation intentionally omits the generalized lifecycle/run-ID framework from the earlier version of this PR.
• The CLI was not published and live PostHog delivery was not tested.

[vercel]

@zahlekhan is attempting to deploy a commit to the thesys-devs Team on Vercel.

A member of the Team first needs to authorize it.