[beinan]

When a persistent native publisher cannot claim its table for 30 seconds, it currently exits as a failure and earns at least 120 seconds of additional backoff despite doing no merge work. The executor now returns a typed pre-claim busy outcome; the CLI emits an attempt-bound JSON marker and exit 75, and the supervisor retries confirmed busy outcomes after 2–5 seconds without adding failure debt.

Claim RPC errors and unknown outcomes remain failures. A claim accepted across the retry deadline is delivered without cancellation. The supervisor joins the child, validates the exact marker and absence of claim/work evidence, preserves previous failure counts and durable deadlines, and retains the full operator/owner/attempt CAS. Old binaries and partial-work failures retain their existing behavior. Deployment requires both the updated binary and supervisor; production has not been changed.

Validation:

• 16 supervisor process/real-etcd and marker-contract tests passed; eight Rust executor regressions passed.
• Real-etcd actual-executor regression passed: a live claim and four pending WAL generations survived the full 30-second busy boundary; after release, merge completed and all four payloads matched.
• Actual compiled CLI plus tracked supervisor and disposable etcd passed: 30.14s busy returned the matching attempt marker and exit75, recorded a ~4.5s deferral, retained previous failure debt and the competing owner, and joined on shutdown. This was local binary staging, not a production or Kubernetes rollout.
• Formatting, Ruff and diff checks passed. CI results are recorded on this PR.