v0.5.3

Fix: the published model table now matches the route it describes.

On the VPS the gateway offered minimax-m3.1-flash-free a 3,491,182-token window against 1,048,092 the route had served, deepseek-v4.1-flash-free 1,233,222 against the 326,000 measured, big-pickle a 32,000-token output where the route named 131,072 in a refusal, and nemotron-3.5-super-vl-free an 8,192-token output for a model whose 1,000,000-token ask was answered 200.

• An encoded attachment is no longer counted as prompt text: a data: URI, or a bare base64 field under a type, mime_type or media_type, is skipped. A 10 MB screenshot became 3,498,676 "tokens" of base64, and that floor was published back as the model's window.
• The estimate scale moves to 3, which drops the brackets measured by the counting that included attachments — those windows were sizes the route was never asked for.
• A completion ceiling the route stated itself (route or route-twin in the table, or a refusal seen at runtime) now outranks a smaller models.dev claim; a larger claim still wins.
• With no paid twin, models.dev's row for the route id itself and the canonical aggregate are both read and the larger wins (NVIDIA's canonical 8192 vs the route row's 131072; minimax's canonical 524288 vs the row's 512000).
• The discovery cache version moves to 3, so a listing cached with the old limits is fetched again.
• bun test 775 pass; CI green.