This is an experimental LLM code review.

Confirmed issues with the new track color picker:

• onDismiss fires on configuration change, prematurely committing in-progress colors to native (mCategory.setCategoryTracksCustomColor, track.setColor) and delivering callbacks to a dying parent fragment. Needs an isChangingConfigurations guard, plus an explicit cancel path on btn_close.
• BookmarkCategorySettingsFragment.showTrackColorPicker() doesn't pass EXTRA_INITIAL_COLOR, so this entry point always opens at red (the other three call sites pass the current color).
• IME insets are not handled — tapping the hex EditText on a small device hides it behind the keyboard.
• hexInput.setText re-set in updateUI jumps the caret to the end on every keystroke once the user reaches a valid 6-char string.
• English source string uses British "Colour" while every other new string and the rest of values/strings.xml use American "Color"; this leaks into auto-generated translations.
• AppCompatResources.getDrawable(R.drawable.ic_plus) is unwrapped without null check — passing a Drawable? into a LayerDrawable(arrayOf<Drawable>(...)) will NPE if the resource fails to resolve.
• Cross-view echo causes redundant work — only SpectrumColorView guards against re-applying its own emission with isTracking; ColorGridView and ColorSlidersView re-run their full update path (including hex setText and shader rebuilds) on every spectrum drag sample.
• rebuildPresetViews recreates all preset views on every selectedPresetIndex change, allocating drawables and re-running FlowLayout measure/layout twice per drag through a preset color.
• ColorGridView.setColor calls invalidate() even when selection is already -1,-1, redrawing the 9*12 grid on every echoed non-matching emission.
• TrackColorPickerViewModel.loadPresets() does SharedPreferences I/O on the main thread in init { ... }; preset format also encodes a redundant alpha byte that is always FF.
• EXTRA_INITIAL_COLOR = "ExtraInitialColor" is reused as both an arguments-Bundle key and a SavedStateHandle slot — namespace it (e.g. app.organicmaps.color_picker.initial_color).
• JNI rename: the indexed nativeSetCategoryTracksColor was deleted on Android in favor of nativeSetCategoryTracksCustomColor, while iOS still uses the predefined-color path. Worth a sentence in the commit body so the asymmetry is captured.

Minor / not posted inline:

• The rebuildPresetViews long-press popup uses Color.RED for the "Delete" label without a theme attribute — fine in light theme, slightly off-brand in dark theme.
• The fragment's bottom-sheet OnLayoutChangeListener is added once and never explicitly removed; it relies on the dialog window being torn down to free it. Works today but is easy to regress if the sheet view is ever reused.
• ColorPickerState.colorChanged is persisted across process death, so even a recreated picker with no user interaction will still callback on dismiss if the previous instance had touched a control. The isChangingConfigurations fix above also helps here, but consider whether colorChanged should be reset on restore.

Tests: there are no unit tests for the view-model state transitions (selectColor, selectPreset, addCurrentColorToPresets, removePreset index-shift logic) — these are pure data and would be cheap to cover.

Generated by Claude Code

This is an experimental LLM code review.

Re-review across Android/Kotlin, Graphics, and C++/JNI/architecture lenses; nothing on iOS as this PR doesn't touch it (intentional asymmetry: Android now sets tracks via dp::Color::FromARGB through nativeSetCategoryTracksCustomColor, iOS still goes through kml::PredefinedColor via the kept overload in bookmark_manager.hpp). The split is sound and the JNI/ARGB byte order matches Android's @ColorInt (drape/color.hpp:57 extracts byte 2/1/0/3 = R/G/B/A, which is exactly the layout of Color.argb). No layer-violation or dependency concerns at the C++ side.

Inline comments above cover the issues I'd want addressed before merge. The only one that's a real correctness bug is the SharedPreferences write race; the others are polish (cursor-vs-displayed-pixel mismatch on Spectrum/Grid for off-rail colors, no-op listener firing from BookmarkCategorySettingsFragment after a re-tap on the selected preset, selector ring drawn outside the rounded clip, SOFT_INPUT_ADJUST_PAN panning the toolbar off-screen, hasFocus() as a too-coarse repaint guard, and a redundant linear scan inside the distinctUntilChanged predicate).

PR description-wise: 70+ files / ~3.1k lines added with no Fixes #NNN link, no test plan, and no note on the iOS/Android asymmetry the JNI rename introduces — please add at least the issue link and a one-paragraph description per docs/PR_GUIDE.md. The earlier reviewer asked for the JNI-rename note in the commit body too; both still pending.

Generated by Claude Code