SDL3 migration, Phase 2: event loop restructuring

[Nidrax]

Summary

Part of #73. Closes #262. Stacked on #268 (Phase 1) — base branch is nidrax/sdl3-phase1-core-port, not trunk, so the diff here is just the event-loop work; will re-target trunk once #268 merges.

Ports SDL2EventQueue::getEventInternal's ~17-case outer event switch (plus the nested 15-case SDL_WINDOWEVENT switch, now flattened into top-level SDL_EVENT_WINDOW_* cases), the keycode/modifier tables, set_input_rect, and SDL_EventState from the SDL2 backend into she/sdl3/she.cpp.

Mechanical renames: ~25 event-constant renames, ~64 SDLK_a..z → SDLK_A..Z, 10 field-access changes (keysym.sym/keysym.scancode → key.key/key.scancode, drop.file → drop.data, ...), 7 signature changes (SDL_StartTextInput/SDL_StopTextInput/SDL_SetTextInputArea now take an SDL_Window* — added a public SDL3Display::window() accessor for this).

Three real bugs fixed, not just renamed (as flagged in the issue):

1. Dropped the SDL_free(drop.file) call — SDL3's drop.data is owned by SDL and freed with the event, so keeping it was a double free on every drag-and-drop.
2. SDL_SetHint(SDL_HINT_KEYCODE_OPTIONS, "unmodified") before SDL_Init — SDL3 otherwise applies Shift/Caps Lock to the reported keycode (e.g. Shift+1 → SDLK_EXCLAIM, not in keyCodeMapping), silently dropping Shift/Ctrl+Shift shortcuts.
3. Cast the now-float mouse/wheel coordinates back to int explicitly where they're stored in she::Event's gfx::Point — SDL2 had these as ints, so the old brace-init would otherwise be a narrowing-conversion compile error.

Also fixed while touching this code (per the issue's "worth fixing" note): pressed-modifier state is now reset on SDL_EVENT_WINDOW_FOCUS_LOST, so Ctrl/Alt no longer gets stuck after Alt-Tab (this bug also exists in the SDL2 backend, left alone there).

New SDL3 event types with no SDL2 equivalent (window enter/leave-fullscreen, occlusion, HDR/ICC/display-scale changes, display hotplug, mouse/keyboard hotplug, clipboard update, drop begin/complete/position) are silently swallowed via range checks (SDL_EVENT_WINDOW_FIRST..LAST, SDL_EVENT_DISPLAY_FIRST..LAST) plus a few explicit cases, so they don't flood the "Unknown event" log.

Out of scope, per the issue's boundaries: SDL_SYSWMEVENT has no SDL3 equivalent reachable from this file (removed outright; replaced by SDL_SetWindowsMessageHook/SDL_SetX11EventHook), so tablet/WM-info handling (EasyTab, nativeHandle()) remains a separate phase — nothing here depends on it.

Test plan

• cmake -DUSE_SDL2_BACKEND=off -DUSE_SDL3_BACKEND=on builds clean.
• Launched bin/besprited under a bounded timeout: window/renderer/theme load as before, and — unlike Phase 1, where the process needed SIGKILL because there was no event loop to observe anything — the process now exits cleanly within the timeout with zero "Unknown event" spam in the log.
• Manual mouse/keyboard/drag-and-drop interaction — not exercised in this sandboxed session (no way to drive real input events here); worth a manual pass before merge.

🤖 Generated with Claude Code