fix: pause builder claims during deploy drain

[sunwoo162]

Summary

• add filesystem-backed Builder drain/busy state
• mark every Builder claim/execution cycle busy before checking drain
• pause new claims while a deployment drain marker exists
• always clear busy evidence after success or failure

Root cause

Production Live E2E run #45 failed at 2026-09-02T23:36:06.743Z with an active runtime bridge returning exit code null. Deploy Bloom Worker run 33695422781 started deleting bloom-builder-worker at 23:36:06.643Z, about 100ms earlier.

Verification

• node --test bloom-worker/run.policy-test.js (8/8)
• pnpm test:production-runtime (56/56)
• pnpm run build:bloom-worker
• pnpm run build:bloom-web
• pnpm run harness
• git diff --check

Known unrelated baseline

• pnpm test:bloom-runtime has an existing lunaServerRuntime policy failure: candidate must use an instance-scoped runtime environment file. The same failure was reproduced on clean origin/main with this patch stashed.

Rollout

Do not merge while a Builder run is active. This is Stage 1: deploy the worker-side drain protocol first. A follow-up deploy-workflow change will stage the bridge and drain/wait before mutating live runtime files.