[xvcezrqpbxtpzblrt]

Problem

• The context size had two controls side by side: a dropdown and a free-entry box that both showed 512K. Nothing said which one won.
• The KV format (Context memory) was only reachable in Expert mode, although it is the setting that decides whether a long context fits at all.
• The steps between 384K and 512K were missing: on a 16 GB card those are the only points between "fits comfortably" and "fits with nothing to spare".

Result

• One control: the dropdown lists the sizes and ends with Custom; only then does the free-entry box appear. 448 is read as 448K, 448K and 1M are understood, 458752 as tokens, and the value is clamped to the model's ceiling with the result shown back.
• Steps are doubled: 8/16/24/32/48/64/96/128/192/256/320/384/448/512K (plus 768K/1M where the catalog allows).
• Context memory sits next to the context size, in the main form, showing the format the installed model really uses (read from its run configuration).
• The card under the version list says when the chosen size will not fit, in the same words for every model.

Validation

node --test tests/native-fit.test.mjs tests/deepseek-card.test.mjs → 18 pass, 0 fail.

Screenshots taken of the real page (before/after): the duplicate field is gone, the four fields are one column, and Context memory shows K8 / V4 matching the running engine.

Limits

• The ceiling is the catalog's largest listed size (512K in the app) or the DeepSeek version's own limit; per-model maxima are not exposed by Control yet, so a 256K model still shows 448K/512K in the list.
• app.css in this branch also carries the row alignment for the model form only.