Switch to 35B i1-Q4_K_M (same compression, faster load); size GPU reservations from measurements

Refine UI: typography, in-panel progress, inline errors, token tooltips

Simplify UI, show token coding costs, and skip automatic decode verification

Increase input limit to 8192 tokens and 64 KiB

Use GGML BF16 cache enum supported by the binding

Add published Qwen3.5 Base Q8_0 arithmetic compressor