Heartbeat hides current agent start-failure reasons until blocked

[clubanderson]

Finding

Fleet can render a generic down — restart needed agent problem even while the spoke already knows the current start failure reason. Example from fleet triage: projectbluefin/common showed 1 agent(s) down — restart needed, but spoke logs showed copilot hung with no output/no CLI prompt.

Expected

Heartbeat should send the current per-agent mgr.StartFailureState before the blocked threshold, including reason, consecutive count, lastAt, and last exit code/signal when available. Hub verdict text should render starting failed ×N: <reason> (last <ago>) and use that as the blocker instead of the generic down/restart text. Absent fields remain backward compatible.

Tests

• Pre-threshold start failure renders the precise start failure reason.
• Threshold/blocked behaviour remains compatible.
• Absent new fields keeps old generic verdict behaviour.

Source: fleet triage 2026-09-04, recommended fix B2.