S2/S12 startup diagnostic checkpoint (2026-10-04)

Recorded for #5551; the roadmap body keeps a one-line pointer here.

When the managed Effect runtime cannot publish its startup locator, the failure now travels through the existing typed startup envelope and reuses the shared filesystem/lock codes (io_is_directory, mutation_lock_timeout, Windows io_permission_denied for an occupied locator directory) instead of surfacing only as runtime_exited_before_ready from an unhandled listen-callback rejection. The bounded message carries no raw Node error, locator path, token or stack trace.

Real fixtures occupy the locator with a directory and hold a live mutation lock. Both fail exactly once with the exact code, leave the foreign occupant and its lock untouched, remove the runtime's own start lock, and resume normal ping/shutdown once the injected fault is gone. An exit without a typed envelope still reports runtime_exited_before_ready.

This closes the reproduced missing diagnostic only. It does not attribute the unrelated Linux unexpected-exit CI failure, qualify Linux/Windows CI as passing, or claim installation, release or strict publication readiness. The operator-facing boundary is in the installation guide.