test: marvel describe team surfaces RoleHealth during backoff

[arcaven]

What to test

marvel describe team should surface per-role crash-loop state — RestartCount, LastRestartAt, BackoffUntil — so an operator can see why a role isn't spawning when they expect it to. Critical for debugging saturation or long-backoff scenarios.

Test plan

1. Apply a crashloop manifest (role that exits quickly, restart_policy: always).
2. Wait through at least 2-3 restarts.
3. marvel describe team <name> — verify the output includes each role's RoleHealth.
4. Trigger saturation (add max_restarts: 2 and re-apply, or wait long enough) and re-run describe team — verify BackoffUntil is far-future.
5. On a fresh team that has never crashed, verify describe team does NOT show a zero-valued RoleHealth block (shouldn't surface noise).

Pass

• Per-role section includes RestartCount, LastRestartAt, BackoffUntil.
• During active backoff, BackoffUntil renders as a human-readable timestamp (absolute or relative-duration, but not a raw UTC integer).
• Roles with no crash history show nothing or an explicit (never restarted) marker — not a zero-valued struct.
• Saturation state is visibly distinct from a normal backoff window.

Fail

• No RoleHealth in output.
• Raw internal struct fields leak (unformatted durations, UTC-vs-local confusion, zero values shown as real data).
• Stale values after the state clears.

Related: PR #21.

[arcavenai]

FAIL on alpha-20260419.034700.20d90ac.

After ~5 restart cycles on a crashloop role, marvel describe team crashloop/flaky returns:

{
  "CreatedAt": "2026-04-19T04:25:37.030491451Z",
  "Generation": 1,
  "Name": "flaky",
  "Roles": [{
    "DangerousPermissions": false,
    "HealthCheck": {"FailureThreshold": 1, "Timeout": 2000000000, "Type": "process-alive"},
    "Identity": "",
    "MaxRestarts": 0,
    "Name": "exit-fast",
    "Permissions": "",
    "Persona": "",
    "Replicas": 1,
    "RestartPolicy": "always",
    "Runtime": {...}
  }],
  "Shift": {
    "OldGeneration": 0,
    "Phase": "",
    "RoleIndex": 0,
    "Roles": null,
    "StartedAt": "0001-01-01T00:00:00Z"
  },
  "Workspace": "crashloop"
}

Failures vs pass criteria

• ✗ No RoleHealth in output. RestartCount, LastRestartAt, BackoffUntil — none present. Operator can't see why a role isn't spawning.
• ✗ Raw internal struct, zero values as real data. The Shift block shows for a team with no active shift — all zero-valued fields, including StartedAt: "0001-01-01T00:00:00Z" which is Go's zero time. That's exactly the "raw internal struct fields leak" antipattern called out in the issue.
• ✗ Raw timestamp formats. CreatedAt renders as UTC RFC3339Nano; Timeout as a raw nanosecond integer (2000000000) — fine for JSON, rough for a human.

What exists that isn't being used

The RoleHealth state clearly exists server-side — visible indirectly through events (restart #N, next backoff=... in daemon log lines) and behavior (backoff spacing). It just isn't exposed via describe team.

Observations worth noting

• Session-level describe session does surface RestartCount, FailureCount, LastHeartbeat, HealthState, ContextPercent (verified earlier today). But those are per-instance values — they read 0 because the session object is recreated on each restart. The team-level counters (RoleHealth, keyed by workspace/team/role) are where the accumulated state lives and that's what isn't shown.
• For operator UX, a Markdown-ish or columnar renderer (similar to kubectl describe) would be more useful than raw JSON. -o yaml and -o json as format flags could keep both audiences happy.