## Problem

The Agents panel presents a cached Claude usage percentage as a current measurement after the limits probe fails. The current collector preserves the last open limits in `bin/omarchy-agent-usage-claude`, but leaves `usageStatusText` empty when it does so.

## Proof it's real

On Omarchy 4.0.1-1, Claude Code's usage dialog showed the current five-hour session at 100% while the Agents panel showed 81% for the same account and reset time.

The record behind the panel was regenerated at 01:25:17, but its limits came from `claude-limits.json`, last fetched at 01:15:26. A forced collection reproduced the failed-probe path: the output contained `"retryAdvised": true`, retained `"percent": 0.81`, and still returned `"usageStatusText": ""`.

The panel recovered on a later successful refresh. The defect is the interval between those requests: the retained measurement is rendered without its stale state, so the user cannot distinguish it from a successful current measurement.

The relevant branch is:

```python
if probe.get("transport"):
  result["retryAdvised"] = True
if fallback:
  result["limits"] = fallback
else:
  result["usageStatusText"] = "Claude limits unavailable"
```

System:

- Omarchy 4.0.1-1
- Claude Code 2.1.251, Max 20x
- Linux 7.1.9-arch1-2
- AMD Ryzen 9 5900HX
- NVIDIA GeForce RTX 3080 Mobile

## Reproduce

1. Let the Claude collector successfully populate `~/.cache/omarchy/agent-usage/claude-limits.json`.
2. Continue using Claude so the account percentage changes.
3. Make Anthropic's usage endpoint temporarily unreachable, or encounter a transient transport failure.
4. Run `omarchy agent usage claude --force --limits-only`.
5. Inspect the output or open the Agents panel.

The collector returns the cached limits and `retryAdvised: true`, but leaves `usageStatusText` empty. The panel renders the old percentage as though the probe succeeded.

## Verify it IS a problem

With an open cached limits window and a stubbed transport failure, run the collector and assert that it both retains the cached percentage and exposes that the measurement is stale in its display-ready record.

If current upstream code already returns a visible stale status or measurement timestamp on this path—and the panel renders it—this report is invalid and should be closed.

## Done when

- A failed limits probe may retain the last known values, but the resulting record marks them as stale.
- The panel displays that state, ideally with the last successful measurement time.
- A regression test covers a transport failure with usable cached limits.
- A later successful probe clears the stale state and replaces the percentage.

## Behaviour

Given the collector has a usable cached Claude limit

When the next limits probe fails

Then the panel may show the cached value, but identifies it as last known rather than current