The complete fix needs a small companion change in InfiniTime: InfiniTimeOrg/InfiniTime#2447.

The bug

LVGL is not thread-safe, but InfiniSim drives it from two threads:

• the DisplayApp task, LVGL's normal owner (lv_task_handler, screen loads), and
• the SDL main thread, whose refresh_screen() creates and deletes the "Screen is OFF" overlay objects and runs lv_task_handler while the display task sleeps.

The wake/sleep handoff between the two corrupts the LVGL heap. It reproduces easily. Toggle sleep/wake with the right mouse button every ~600 ms, and an AddressSanitizer build crashes within about 4 cycles:

ERROR: AddressSanitizer: heap-use-after-free
READ of size 8 ... in get_property_index lvgl/src/lv_core/lv_style.c:996
    ... _lv_disp_refr_task <- lv_task_handler <- DisplayApp::Refresh   (thread "displayapp")
freed by thread T0 (main) here:
    ... lv_obj_del <- Framework::refresh_screen <- Framework::refresh <- main
previously allocated by thread T0 here:
    ... lv_obj_set_style_local_bg_color <- Framework::refresh_screen  (screen_off_bg)

In a normal build the same corruption surfaces later as a SIGSEGV in the style walk during refresh (_lv_style_get_int reading a freed style map). The existing screen_off_delete_cb double-free guard covers one symptom but not the underlying race.

The fix

A std::recursive_mutex behind LVGL_GUARD() (sim/displayapp/LvglGuard.h), with the two sides locking differently on purpose:

• the DisplayApp task takes it blocking, since LVGL is its whole job;
• the main thread takes it with try_lock (LVGL_TRY_GUARD()) and skips the overlay frame if the display task holds it. The main loop also pumps SDL input, so it must never park on a lock held by another thread. Doing so would turn any stall inside the display task into a whole-process wedge, dead input, rather than a frozen watch screen. A skipped overlay frame is harmless; the loop runs again about 30 ms later. I hit exactly this while testing a blocking first version: the process stayed alive with the main thread in futex_wait.

The other half of the race is the DisplayApp task itself, which is InfiniTime code. InfiniTimeOrg/InfiniTime#2447 adds a no-op displayapp/LvglGuard.h to InfiniTime and takes LVGL_GUARD() around the two LVGL regions of DisplayApp::Refresh, not across the blocking queue wait. On hardware the macro compiles to nothing, since LVGL has a single owner there. In the sim, this repo's header shadows it via include-path order (the same mechanism the sim already uses for nrf_log.h), so both threads serialize.

Without the InfiniTime side, this change alone is safe but only partial: it serializes the overlay operations against the sim's own off-screen lv_task_handler, not against the display task.

Verification

• Stress: 300 wake/sleep cycles under ASan, zero reports, sim alive. Unfixed, it crashes at cycle 4.
• This branch compiles standalone against the vendored InfiniTime submodule (cmake && make), no new warnings.
• Normal interactive use, and alarms and notifications firing from sleep, are unaffected.