select: run on_dismiss however the popup closes

[huacnlee]

Follow-up to #2973, which is already merged. That PR is good work — this closes the gaps a post-merge review turned up.

on_dismiss was skipped by the accessible close

Select's Cancel handler runs on_dismiss before asking the controlled open state to close. The on_a11y_action(Click) path added in #2973 called on_open_change(false) alone, so a consumer wiring on_dismiss saw Escape and an outside click but not a screen-reader user pressing the same control to close.

That is not hypothetical: crates/shell/src/materialize/components/select.rs:240 forwards on_dismiss to JS as onDismiss, so a shell app silently lost the callback on exactly the interaction #2973 was written to support.

Both paths now share one close closure, so they cannot drift apart again:

let close: ActionHandler = Rc::new({ /* on_dismiss → on_open_change(false) → focus trigger */ });

Also in this patch

• zh-CN docs. select: expose committed values and accessible activation #2973 changed website/docs/components/select.md and website/base/primitives/select.md; both zh-CN counterparts exist and were left behind. Synced.
• SearchableListItem docs. The rule that the accessible value reports title() while display_title() stays presentational was documented only on the site. Moved onto the trait, where someone implementing display_title will actually read it.
• A test assertion that asserted nothing. projects_application_owned_accessible_state checked disabled.is_expanded() == Some(false) on a Select that was never opened, so it passed for a reason unrelated to disabled. It is now opened, and asserts a disabled control still reports the state it is in.

Testing

cargo test -p gpui-base -p gpui-component --lib   # 777 + 418 pass
cargo clippy -p gpui-base -p gpui-component --all-targets --locked   # clean
cargo fmt --all --check   # clean

GPUI exposes no way to dispatch an accessibility action from a test — Window::handle_a11y_action is pub(crate) — so every_close_dismisses_before_it_closes covers the shared close path through Escape, the route a test can reach. I confirmed it fails (["open", "close"] vs ["open", "dismiss", "close"]) when the on_dismiss call is removed again.

Left open deliberately

Two findings from the same review are not in this PR, because both are judgement calls rather than defects:

1. The accessible value and the drawn trigger disagree during a search. accessibility_value() reads the committed state.selection; display_title() reads the list cursor, which set_query clears. Verified on main:

   after set_selected_value("Rust") then set_query("Go"):
     a11y value = "Rust"     trigger draws = <placeholder>
   

   I think the committed value is right for assistive technology and the display is the bug — a Select with a committed selection should not visually revert to its placeholder while you type in the popup's search box. That is pre-existing behaviour, so fixing it is a separate change. Worth deciding whether to fix it or stop documenting the mismatch as intended.

2. Placeholder as the accessible value. With nothing selected the AX value becomes the placeholder, so AT reads "Programming language, Choose a language". GPUI has aria_placeholder (gpui-pre-0.3.2, src/elements/div.rs:1391), which is the semantically correct slot — that would also keep name, placeholder and value as three separate things, the way an_explicit_accessibility_label_does_not_replace_the_placeholder already insists.

AI assistance

The post-merge review that found these and this patch were produced with Claude Code. I reviewed the diff and the checks above before opening this.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EC7fHTjQ6nGPKqq4WdryL8

[huacnlee]

Raising the severity on this after tracing the consumers.

gpui_base::Combobox renders a gpui_base::Select and forwards on_dismiss to it (crates/base/src/combobox.rs:126,141). The styled Combobox uses that callback to commit the pending selection (crates/component/src/combobox.rs:944):

// This combobox commits its pending selection when the popup
// closes, so it listens for dismissal rather than Confirm.
.on_dismiss(move |_, cx| {
    confirm_state.update(cx, |state, cx| {
        cx.emit(ComboboxEvent::Confirm(state.selected_values()));
    });
})

So before this patch, a screen-reader user closing a Combobox through the accessible activation never got ComboboxEvent::Confirm — their selection was silently discarded, while Escape and an outside click committed it normally. That is data loss on the exact interaction #2973 was written to enable, not just a missing callback.

crates/component-shell/src/shell/delegate_combobox/mod.rs:213 consumes that event, so the loss reached the shell delegate too.