Re-review for the second upstream commit c4a14aa9 of PR #14155. AppSearch.sortedEntries now orders by localeCompare of the lowercased key and then by the desktop id, the same tie-break as the Apps comparator in rebuildDisplay (localeCompare of the lowercased labels, then the item id). So two names that collate equal come out in id order whatever order they arrive in; this supersedes the input-order wording of REVIEW-261003-JH9B and REVIEW-261003-721N. The PR test now runs node under LC_ALL=C.UTF-8, so its expectations use the root collation on any test machine, and it asserts id order for a composed and a decomposed É in both arrival orders. seq-apps-plugin-tie shows the menu and the plugin list agree on that order. B7ZA states that ties keep a fixed order by desktop id, the same as the Apps menu. The formula and the defensive violation row (REVIEW-261003-JTMM) are unchanged.

Final re-review on quattro-proof 49cd12b0, where shell/services/AppSearch.js is in the verification scope. sortedEntries carries an Implements line for SW-REQ-261003-B7ZA next to SW-REQ-261003-C9GM. With no query it orders by localeCompare of the lowercased key, and when that returns 0 by the old < and > on the key and then the name. So two different names that collate equal keep one order, and identical names keep their input order, as before. The Apps comparator in rebuildDisplay orders by localeCompare of the lowercased labels and then the id. Under C or C.UTF-8 Qt keeps UTF-16 code-unit order, the same as before the fix; LC_ALL, then LC_COLLATE, then LANG select the collation. Witnesses: the PR test runs the real sortedEntries (accented names, collate-equal names in both input orders) and the Apps comparator; seq-apps-plugin-list shows the menu and the plugin list agree. C9GM (which apps the library lists) is unaffected: only the order changes.

Re-review of the round-3 text corrections. Under C or C.UTF-8 Qt keeps UTF-16 code-unit order, the same as before the fix; this corrects the term code-point order in REVIEW-261003-AP6S and REVIEW-261003-CY43. B7ZA now states which setting selects the collation: Qt takes it from LC_ALL, then LC_COLLATE, then LANG, and LANGUAGE does not change it. Checks with the QML engine of Qt 6.11 show it: LANG=fr_FR.UTF-8 with LC_COLLATE=C.UTF-8 keeps É after Z, and LANG=en_US.UTF-8 with LC_COLLATE=fr_FR.UTF-8 puts É under E. The comparator, the formula and the defensive violation row (REVIEW-261003-JTMM) are unchanged. The witness test now also asserts the encoding, ascii and first-row inputs (15 cases).

Re-review for the c5ed4375 form of AppSearch.sortedEntries, which supersedes the sortedEntries wording of REVIEW-261003-CY43. With no query, sortedEntries orders by localeCompare of the lowercased key; when that returns 0 it falls back to the old < and > on the key and then on the name. So two different names that collate equal, such as a composed and a decomposed É, get one order whatever the input order (the PR test checks both input orders). Two identical names compare 0 at every step and keep their input order, as on a85e29ab; there is no id tie-break in sortedEntries. B7ZA now says exactly that. The Apps comparator in rebuildDisplay is unchanged from CY43: localeCompare of the lowercased labels, then the id.

Round-3 re-review on the upstream change cb981de6. The Apps comparator in rebuildDisplay orders the lowercased names with localeCompare and falls back to the id only on a tie. AppSearch.sortedEntries, which the plugin API hands out, now orders by the lowercased key and then the name with localeCompare, so a plugin sees the menu order. The restated claim matches the collation: accented Latin names sort under their base letter; with the language's own collation (ru_RU, el_GR, zh_CN) that script comes first, before the Latin names, where a85e29ab put it after Z; under en_US or fr_FR other scripts follow the Latin names; under C or C.UTF-8 Qt keeps code-point order. The real shell shows each of these orders. The node witnesses pin the comparator: seq-apps-locale-ru/el/zh collate as that locale through @locale, seq-apps-nonlatin as en-US, and the PR test's fourth case runs the real sortedEntries. The formula and the defensive violation row (REVIEW-261003-JTMM) are unchanged.

Re-review after the restatement on quattro-proof 87c3ca68. The comparator in rebuildDisplay orders the lowercased names with localeCompare and falls back to the id only on a tie, so the rows follow the collation of the locale, not a promise that every name sorts under a Latin letter. Accented Latin names sort under their base letter (seq-apps-french, seq-apps-german, seq-apps-decomposed). Greek, Cyrillic, Arabic and Han names sort after the Latin names, at base and at head (seq-apps-nonlatin). Symbol-led names now sort before digits and letters (seq-apps-symbols), the same order the real shell shows under en_US.UTF-8. Qt supplies the collation in the shell; under C or C.UTF-8 it keeps code-point order, so the change is a no-op there (real shell). The node witnesses pin the comparator, not Qt's collation. The formula and the defensive violation row (REVIEW-261003-JTMM) are unchanged.

Read shell/plugins/menu/Menu.qml on the pr-new-D tree (upstream a85e29ab plus the fix). rebuildDisplay, with no search query and the Apps menu active, collects the visible child rows and sorts them with an inline comparator. The comparator lowercases both labels and orders them by localeCompare; only when localeCompare returns 0 does it fall back to the item id, compared with < and >. So apps_rows_rebuilt leads to apps_rows_in_letter_order in the same call, before the rows go to displayModel. rebuildDisplay is the only place that fills displayModel for a non-dmenu menu, and every path that shows the Apps menu (openExistingMenu, setActiveMenu, provider refresh) calls it. With a search query the search sort applies instead, which already used localeCompare. The violation row (a rebuilt Apps list that is not in that order) needs the comparator to go back to the code-unit < and > of a85e29ab; it is dispositioned defensive. The PR's test test/shell.d/menu-apps-order-test.sh extracts this comparator from Menu.qml and runs it on French and German names and on a case tie; it fails on a85e29ab.