[lellomele]

Summary

• Let the Start Menu accept input while shortcuts are still being preloaded. Resolve frequent programs from UserAssist in their existing order, then apply new-program highlighting when preload is ready. Update the open menu and Programs tree without rebuilding either list or moving the selection.
• Allocate temporary icon image lists on first use and release retired bitmaps after the item workers have stopped.
• Validate cached PIDLs, string lengths, bitmap dimensions and blacklist records. Replace the previous cache only after the temporary file has been written and flushed successfully.
• Clear the disk cache without invalidating entries still referenced by the menu or its workers. Report whether the operation succeeded and rebuild the cache on the next launch.
• Clean up partial search-worker initialization and allow item workers to exit promptly on shutdown.

This is based on 9dfa7fa. It keeps the index checks from #2549 and the cache writer's exclusion of temporary items and icons from #2550. Cache clearing also waits for an in-flight writer before deleting the file.

Validation

• Built Setup|x64 for both master and this branch with MSVC v145. Both builds completed with zero errors and warnings.
• Compared the two builds in Windows Sandbox with their DLLs loaded in Explorer. In two pairs per cache state, first keyboard input was 390–487 ms earlier with an empty cache and 181–224 ms earlier with a populated cache. All runs accepted six navigation keys, found the test shortcut through search, highlighted it and unloaded the DLL.
• Repeated the comparison with the menu's accessibility provider enabled. All nine runs, including the warm-up, exposed the same three frequent entries in the same order from four seeded entries that shared one duplicate app ID.
• Earlier Windows Sandbox tests on the v4.4.198 development builds verified selection and frequent-program order, cache clearing through Settings, regeneration on restart, and preservation of a locked cache file. The focused startup comparison across six pairs per cache state measured median paired reductions of 261 ms with an empty cache and 110 ms with a populated cache. A separate warmed comparison of lazy icon lists measured a median reduction of 26 GDI objects.