fix: drain builder before mutating worker runtime

[sunwoo162]

Summary

• transfer the runtime bridge into a per-run staging directory instead of the live path
• request Builder drain and wait for the active busy cycle to finish before touching the shared checkout
• remove stale busy markers only when their recorded PID is no longer alive
• smoke-test a next bridge candidate and atomically promote it to the live path
• always remove the drain marker and staging artifacts on exit/signals

Root cause

Production Live E2E run #45 was interrupted by worker deployment: pm2 began deleting bloom-builder-worker roughly 100ms before the active runtime bridge exited with code null. Stage 1 (#193) added worker-side drain/busy markers and is already deployed.

Verification

• pnpm test:production-runtime (57/57)
• pnpm run build:bloom-worker
• git diff --check

Rollout

After merge, verify Deploy to Server and Deploy Bloom Worker both succeed and confirm the worker deploy logs show Builder drain request/completion, runtime bridge smoke, and both worker processes online.