Usage dashboard 'Favorite model' on 'All' filter reflects recent usage only, not all-time

Describe the bug

The usage dashboard's "Favorite model" stat displays a model that is dominant only in the last 2–3 weeks, not over the full account history, despite the "All" time filter being selected. The "All" filter appears to use a recency-weighted or windowed calculation rather than true all-time aggregation.

Again, I am not estimating — I have a complete local database built from my JSONL session files and can count records precisely.

To reproduce

1. Open the claude.ai usage dashboard
2. Select the "All" time filter
3. Note the "Favorite model" displayed (Opus 4.7 in my case)
4. Switch to the "Models" tab and observe the actual session distribution across the full history
5. The model shown as "Favorite" in step 3 is not the plurality model in step 4

Expected behavior

With "All" selected, "Favorite model" should reflect the model with the highest usage across the entire account history.

Actual behavior

"Favorite model" reflects only recent usage — approximately the last 2 weeks. Opus 4.7 is shown as "Favorite" despite being adopted only on approximately April 17, 2026.

Actual all-time session counts from local JSONL records:

| Model | Sessions |
|-------|----------|
| Sonnet 4.6 | 123 |
| Opus 4.6 | 120 |
| Haiku 4.5 | 48 |
| Opus 4.7 | 43 |
| Sonnet 4.5 | 31 |
| Opus 4.5 | 43 |

Opus 4.7 is fourth by session count. Sonnet 4.6 is the plurality model by sessions; Opus 4.6 leads by total tokens (15.1M vs 3.6M for Opus 4.7). Opus 4.7 does not win on any reasonable definition of "favorite" over the full period.

On the meaning of "All"

My account's first activity was November 20, 2025. "All" should mean all sessions since account creation, not a recent window. If the stat is intentionally recency-weighted, that should be disclosed in the UI (e.g., "Favorite model (recent)").

Additional context

Related to but distinct from #53469, which identifies a separate aggregation bug (text-mention counting vs. actual message.model field). Both bugs produce incorrect "Favorite model" output via different mechanisms — #53469 is a wrong aggregation method; this is a wrong time window.

Both this bug and #53469 may be partially downstream of the session ingestion failure reported in #55212 — if the majority of sessions are missing, the model distribution is wrong regardless of how "favorite" is calculated.

The usage dashboard was released approximately April 13, 2026.