Dashboard 'Favorite model' counts text mentions, not actual inference model

Summary

The "Favorite model" stat on the home dashboard ("What's up next, …?") appears to count text mentions of model names inside session content, rather than the model field on assistant messages in ~/.claude/projects/**/*.jsonl.

Repro

On my machine the dashboard shows:

Favorite model: Mim* 2 Pro

But the actual model field on assistant turns across all sessions is:

| Model | Assistant messages |
|---|---|
| claude-opus-4-6 | 219,525 (84%) |
| claude-haiku-4-5-20251001 | 22,784 (9%) |
| claude-opus-4-7 | 15,589 (6%) |
| claude-sonnet-4-6 | 5,636 (2%) |

So the favorite should be Claude Opus 4.6 by a wide margin — not "MiMo 2 Pro" (which has zero inference messages and is only present as a topic in my project documentation).

Substring counts of the suspected source:

| Substring | Occurrences in .jsonl |
|---|---|
| MiMo | 6,060 (project notes about an internal MiMo integration) |
| Mistral | 65 |
| MiniMax | 11 |

The strong correlation between the "winner" and the most-mentioned non-Claude model name in user/assistant text content suggests the aggregator is grepping content rather than reading the message.model field.

Expected

"Favorite model" should aggregate the message.model field on assistant turns (or token usage), not text mentions inside conversation content.

Environment

• Claude Code: Opus 4.7 1M
• Platform: macOS Darwin 25.4.0