Three implemented surfaces, three bounded visual questions, nine candidates. Every frame is the same live component under different props, so the comparisons are real renders rather than drawings. Nothing here changes semantics, data, routes, copy, persistence, protocol or desktop behaviour — the candidates differ only in grouping, hierarchy and mark vocabulary.
The catalogue is daemon-owned and may be long, legacy or custom, so the list itself cannot be shortened, aliased, truncated or hidden. The only thing a visual refinement can change is how quickly the page answers “what will a new session actually use?”.
All three candidates are shown in the hardest state the brief names: 320 dp, streaming, with a wrapped draft staged — Stop and Send both live, Mic still in the field, and the cross-session switcher present in the accessory lane. Nothing about which controls exist changes; only how the bands are stacked. Product copy in these frames is the existing resource text verbatim.
Precedence, derivation, persistence and ordering are fixed. The question is narrow: which mark makes New Result read as its own rung on the ladder. All three candidates keep the written label and the attention tone.
Each selected candidate states the current situation in words at the top of its surface, then leaves the existing controls untouched below. The defaults summary, the composer ribbon and the written state label are the same move at three scales, which is why they cost one implementation pattern rather than three.
S2 was the shortest page on the board and is rejected because it hides controls. No selected candidate introduces a disclosure, overflow, swipe, horizontal scroll, type shrink or id ellipsis. Every list that is open today stays open, and every control keeps its 48 dp floor.
R3 makes fill carry the session ladder and V3 makes shape separate a failure ribbon from an informational one. Both read correctly in greyscale, which is the standing rule the current ring-versus-ring pairing quietly breaks.
The block is built only from existing group labels and stored values, so it introduces no string. The master should confirm the pairs render from the same resources the groups below use rather than duplicating them, and diff the ZH composer strings against the resource file.
Fast is proven here because the fixture's selected model advertises priority. Context window, custom value, per-model overrides and the agent filter are unchanged and sit below these frames; the master owes the full scroll, including the summary dropping Fast when a non-priority model is selected.
This board proves the ribbon on the four that determine structure. The master owes the full strip — idle, staged, streaming empty, streaming staged, upload, recording, transcribing, failure — at 402, 390, 320, 280 and 200%.