Goose model picker: Grok 4.6 missing thinking levels and correct context window

### Before filing

- [x] I searched [open and closed issues](https://github.com/block/berd/issues?q=is%3Aissue) for duplicates.
- [x] I reproduced this on the [latest release](https://github.com/block/berd/releases).
- [x] This is one bug, not several bundled together.

### Closest existing issue

#141 is adjacent (composer usage tooltip capped at 128k). This report is not that meter. It is Grok 4.6 on the Goose harness: the model picker does not show thinking-level controls or the correct context window, while Goose CLI does.

### What's broken

On Goose in Berd, `grok-4.6` (xAI) appears in the model picker, but the picker does not surface that model's thinking-level options or the correct context window. The same Goose 1.47.0 CLI does recognize Grok 4.6: it lets me choose a thinking level and shows the correct context window.

### Steps to reproduce

1. Launch Berd 0.6.2.
2. New chat, Goose harness, xAI / `grok-4.6` (`xai_oauth`).
3. Open the agent/model picker and look at thinking-level options and the context window shown for that model.
4. Compare with Goose CLI 1.47.0 against the same model: CLI offers thinking-level choice and displays the correct context window.

### What you expected to happen

Berd's Goose model picker should expose the same Grok 4.6 thinking levels Goose CLI does, and should show the same context window Goose CLI does for that model.

### What actually happened

The picker is missing Grok 4.6 thinking-level information and does not show the correct context window. I had to patch Goose 1.47.0 locally so the CLI path was right. I did not write down the exact wrong window Berd displayed.

Stock Goose 1.47.0's xAI known-model list still defaults to `grok-4.5` and does not treat `grok-4.6` as an xAI reasoning-effort model. The model can still show up from the live xAI list, which is how Berd can list `grok-4.6` without the thinking/context metadata.

### How often does it happen?

Every time — reliably reproducible

### Berd version

0.6.2

### Operating system

macOS (Apple Silicon)

### Model and provider

Grok 4.6 (`grok-4.6`) via xAI (`xai_oauth`) on the Goose harness

### Relevant log output

```
[renderer] [reasoning-effort] snapshot dispatch {"modelId":"grok-4.6","providerId":"xai_oauth","hasModelSnapshot":true,"hasReasoningEffortSnapshot":true,"reasoningEffortConfigId":"thinking_effort","reasoningEffortCurrentValue":"max","reasoningEffortOptionCount":5,"configOptionCount":4,"configOptionIds":"provider,mode,model,thinking_effort","configOptionCategories":"mode,model,thought_level"}
```

These lines are from a machine that also has a local Goose 1.47.0 patch. I am not claiming they are the broken unpatched state. No other relevant log lines for the wrong context-window number.

### Screenshots, recordings, or other context

`berdctl info models --harness-id goose` does list `grok-4.6` / Grok 4.6 / `xai_oauth`. That call only returns `model_id`, `name`, and `provider` — no context-limit or thinking-level metadata.

Goose CLI 1.47.0, same machine, does let me choose thinking level and shows the correct context window. Berd 0.6.2 ships its own `goosed` sidecar, which may not be that CLI.

I am not asking Berd to take my local Goose patch. I want the picker to show Grok 4.6 thinking levels and the correct context window, however that gets wired.

Start at the Goose model picker and the `berdctl info models --harness-id goose` entry point, then trace how the bundled `goosed` sidecar supplies model metadata. Compare that path with Goose CLI 1.47.0 for `grok-4.6` and its xAI reasoning metadata. Done means the picker exposes the available thinking levels and displays the correct context window for this model.

Written by the indexing model from the issue text.