Compound controls: pickers, wheels and steppers

A control iOS reports as several nodes, or as one node hiding several. Three of these came out of the first run against a real Clock alarm and none was reachable from a fixture.

• A Picker is not a PickerWheel, and mapping both to one role cost the selection. A picker’s rect is the union of its wheels, so it is concentric with whichever wheel is in the middle, and _dedupe_colocated read that pair as one control reported twice. The container won, because it carries an accessibility id where a wheel carries only a value, and the middle column’s selected option vanished: a date picker read =6 and =AM either side of a wheel showing nothing. The container is now its own role and collapses like any other wrapper. The same bug is latent wherever a container’s centre lands on a child, which is why the fixture has three wheels rather than one. Measured with the ranking change below: the oracle series unchanged on all thirteen tasks, the golden flows 8306 device tokens to 8361, +0.7%, and the resolution-tier distribution identical, so two capabilities cost under a percent between them.
• An accessibility id is not a label, and treating them as equal deleted a control. _beats asked “does this node have text” as a boolean, so a Stepper named interval_stepper outranked buttons labelled “Increment” and “Decrement”. A stepper’s rect is exactly its two buttons side by side, its centre lands on the boundary they share, and Rect.contains is inclusive at an edge, so both children were coincident with the container and both lost. The digest showed one node with an id and nothing to tap, whose centre happens to sit on Increment: raising a value silently worked and lowering one was unreachable at any price. Ranked now, label above id above nothing.
• A picker wheel hides every option but the one it is showing. There is no options list in the tree at any price, so nothing can compute how far to turn, and set_value moves it a row at a time and reads it back. The version before it sent every non-switch role to send_keys, which types into whatever holds keyboard focus: a wheel holds none, so the call succeeded, changed nothing, and reported success. Segmented controls and steppers are not set this way at all, because iOS already reports their parts and the digest already shows them; set_value on one says so and names the part.
• A wheel’s value is a spoken phrase, not the thing it looks like. A real Clock alarm reports 5 o’clock, 36 minutes and PM, with a typographic apostrophe. set_value("7") cannot match at any tier, which is why failing lists every option it saw: that list is the only way a caller learns the spelling. Measured: 12 steps and 30.7s to walk a twelve-hour wheel once.
• Apple wraps a picker twice, and the outer wrapper is a Cell. Collapsing the Picker is not enough. The Cell has the same rect, so it is mutually centred with the middle wheel of a three-column time picker, and with picker missing from ROLE_PRECEDENCE the tie went to the cell: the minutes column vanished while the hours and the meridiem either side of it survived. A role absent from that tuple is not neutral, it is beaten by every wrapper that has a role on it.
• A drag cannot move a UIPickerView by a known amount; a tap can. UIPickerView decelerates a flick through however many rows the momentum carries. Measured on an iPhone 17 Pro Max: a fifth of the wheel’s height over 0.2s moved four rows, slowing it to 1.6s and shortening it to a tenth still moved three, and the count varied between runs. Since an hour wheel wraps, a nondeterministic multi-row step cannot reach two thirds of the options at all: stepping by four from 5 o’clock reaches 1 and 9 and nothing else, forever. A tap 30 points from the middle selects the neighbouring row exactly, twice out of two, and at 55 points it selects two rows. So rows are about 30 points and the band ends near 45. set_value on a neighbour went from 61.4s of random walk to 9.6s.
• A wheel that wraps has no end to detect. “Stop when the value stops changing” never returns on an hour wheel. The loop stops on a value it has already seen, which covers both a wrap and a finite wheel run onto its own end, and a wheel that returns to where it started is known to have been walked in full, so the reverse pass is skipped.
• A simulator’s Settings is a reduced build, and it has no wheel in it. On iOS 27, iPhone 18 Pro, there is no Date & Time pane under General, no Sounds pane, and no Clock app installed at all, so there is no stock picker wheel or slider reachable from Settings. The compound-control integration tests skip on that build and say why, and tests/integration/test_device_picker.py carries the wheel proof instead, against a real iPhone. Three of the bugs listed above came out of its first run and none of them was reachable from a fixture, which is the rule this file already states and which held again.