Runner heartbeat-only jobs can block daemon rotation indefinitely

[chubes4]

Problem

A runner.exec workload can remain running indefinitely when its wrapper continues emitting heartbeats even though the child workload makes no meaningful progress. Because the job remains active, a daemon with binary_hash_mismatch is not restartable, runtime generations cannot drain, and new Lab work is rejected with daemon lease is not fresh.

Observed evidence

• Homeboy: 0.304.0+3e0ece8b6f14
• Runner: homeboy-lab
• Job ID: d911d948-062f-4ffb-b089-1671290d2177
• Durable run ID: blocks-engine-663-standalone-demo-94c26a4d
• Source workload: Improve Figma image projection and responsive layout Automattic/blocks-engine#663
• Started: 2026-07-22T12:48:09.521Z
• Still reported active after more than three hours
• Process tree: 8 processes, about 657 MB RSS, only about 33 seconds cumulative user CPU after roughly three hours
• Logs after initial activity contained only periodic phase=heartbeat events
• homeboy runner status homeboy-lab reported daemon_freshness.fresh=false, stale_reason_code=binary_hash_mismatch, restartable=false, 33 draining generations, and this job as the final active pin
• New Agent Task Cook admission failed before provider execution with daemon lease is not fresh

Reviewer reproduction/inspection commands:

homeboy runs show blocks-engine-663-standalone-demo-94c26a4d
homeboy runs evidence blocks-engine-663-standalone-demo-94c26a4d
homeboy runner job logs homeboy-lab d911d948-062f-4ffb-b089-1671290d2177
homeboy runner status homeboy-lab

Expected contract

• Runner liveness should distinguish wrapper heartbeat from child workload progress.
• A heartbeat-only workload with no meaningful process or output progress should transition to a diagnosable stalled state under bounded policy.
• Stalled jobs should expose deterministic cancellation/recovery guidance and should not pin daemon generation promotion forever.
• Active, genuinely progressing jobs must remain protected during daemon rotation.

Impact

This blocked Homeboy issue #8395 itself from entering provider execution and required manual operator diagnosis. The failure is circular: the starvation fix cannot be cooked on the runner because a heartbeat-only job is currently starving runtime promotion.