refactor: consolidate the duplication from the repo review (DRY) (#470)

[Maxaubert]

Closes #470

Consolidates the duplication the 2026-10-13 repo review found (section "DRY and duplication" and related). Refactors only; each finding was re-checked against current main first.

What changed

1. Exe folder and %LOCALAPPDATA%\Wind: cached wind::ExeDir() and wind::UserDataDir() in config_path.h replace the copies in ResolveIniPath, ResolveLogDir, main.cpp, WindTray (AppDir/g_appDir removed), WindConfig (ExeDir, the WebView2 folder) and SessionKeepPath/RunningModelPath.
2. trim: one pure wind::TrimWs in ini_edit.h replaces the five copies (config, ini_edit, profiles, tray_items, flyout_tools).
3. UIAccess token read: ProcessHasUiAccess() in the new Win32 header uiaccess_token.h (with the banned-probe note), used by the Refuse to run without UIAccess, with a message that says how to fix it #460 start gate and the transform engine.
4. MPO: ReadOverlayTestMode() in mpo.h is the one OverlayTestMode read (it keeps KEY_WOW64_64KEY and the type check). MpoDisabledInRegistry and DetectMpoDisabled use it, and the startup log still shows the raw value.
5. Kernel object names: kInstanceMutexName, kQuitEventName and kIniWriteMutexName sit next to the tray constants in tray_ipc.h, and all three binaries use them. installer/app.nsh, tools/testenv/lib.ps1, tools/gpu_ab.ps1 and tools/zoom_response_ab.ps1 keep their literals, each with a comment pointing to the header.
6. Global / UI-only keys: kUiOnlyKeys in profiles.h is now the only C++ list. IsGlobalProfileKey, MakeLiveText, StripUiOnlyKeys and the test all loop over it. A new fixture, tests/fixtures/global_keys.txt, is checked from both sides: by a doctest against kUiOnlyKeys + profile, and by a new node test against the UI's GLOBAL_KEYS (it runs under npm run test:rules).
7. Profile switch and relaunch: src/profile_switch.h now holds SwitchProfile, LaunchWindExe, RevertModelAfterFailedRelaunch and WriteSessionKeep/ClearSessionKeep. The tray switch, the Settings switch and the tray engine pick share them, and each surface keeps only its own messages and logs.
8. Render path: RLog, ULog and DiagLog are now one-line macros over wind::Log, so no call site changes and lines are no longer formatted twice. CompileHlsl (hlsl_compile.h) is the one D3DCompile wrapper; each caller keeps its own flags.
9. build.bat: a :core_exe <define> <manifest> subroutine builds Wind.exe for both the app and uiaccess targets. check now also compiles src\config_ui\main.cpp (into its own object folder).
10. Version: tools/version.ps1 (Get-WindVersion) is dot-sourced by release.ps1, installer_check.ps1 and both workflows. The alpha and release workflows were changed together. A new tests/test_version.cpp checks WIND_VERSION_STR against the three numbers. It is a test and not a stringize, because the .rc files consume the literal.
11. testenv: sweep.ps1, shimmer_ab.ps1 and ab_acryl.ps1 now use lib.ps1's Update-IniKnobs, Restart-WindClean and Stop-Wind instead of their own in-place ini rewrite and fixed 2 s quit. The settle time is unchanged.
12. Theme data: WindConfig's ThemeBackground now reads the generated kPalettes[i].dark.menu instead of four hand-copied colours (same values). A new doctest checks each palette's menu colour against that theme's --bg in ui/src/design/themes.css. The kUiPalettes vs kPalettes id test already existed. flyout_icons.h is left hand-copied: the review says to generate it only if the icon set grows.

Docs: CLAUDE.md (the "add new UI-only keys" rule, the check target) and chapters 01, 02, 05, 08, 09 and 11 now point at the new helpers.

Behaviour notes (intended, minor)

• The tray's profile switch now writes session.keep before a relaunch, as Settings already did; the copies had drifted. After a switch the session equals the profile, so the next start comes out the same either way.
• When LOCALAPPDATA is unset (in practice it never is), session.keep and running.model now fall back to %TEMP%\Wind like the ini, logs and WebView2 folder, where before they used .\Wind.
• WindConfig now launches Wind.exe with the exe folder as its working directory and checks that the file exists first, like the tray already did. Wind sets its own working directory anyway.
• A failed render-engine shader compile now logs the compiler's error text (before, it failed silently).

Verified

• build.bat check: exit 0, now including WindConfig's host.
• build.bat test: exit 0, 632/632 test cases, 98195/98195 assertions. That includes the new fixture, version and themes.css tests.
• build.bat (Wind.exe + WindTray.exe): exit 0. build.bat uiaccess: exit 0, and the embedded manifest reads uiAccess="true" ("false" for the plain build).
• npm ci + build.bat config: exit 0.
• npm run test:rules: 2/2 pass (keybind rules plus the new global keys test). npx playwright test: 205 passed.
• build.bat installer: exit 0, all 63 installer checks OK. The install smoke test was skipped because the shell is not elevated. WindUA.exe was produced the way release.ps1 does it (uiaccess build renamed), without running release.ps1.
• . tools/version.ps1; Get-WindVersion prints 0.29.3, which matches src/version.h and WIND_VERSION_STR. All touched .ps1 files parse, and both workflow YAMLs load.
• Not deployed and not run. Wind.exe needs a field test with the signed UIAccess build.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TdjopXZfsUiesYxYLYHEdH