A targeted fix release. The headline is the ASAN-confirmed root-cause fix for the chronic wizard step-transition crash family that has driven roughly half a dozen patch attempts since v0.99.34. With AddressSanitizer wired into the Pi cross-build, two distinct heap-use-after-frees were caught within minutes of human interaction with the wizard — replacing weeks of victim-site whack-a-mole.
-
Wizard step-transition heap corruption — root cause (#880, #871, #870, #872, #873, #874, #875, #877) —
WizardConnectionStep::auto_probe_timer_is a one-shot LVGL timer; LVGL deletes one-shot timers internally after the callback returns, butattempt_auto_probe()only nulled the member pointer on its non-early-return path. When the timer fired and early-returned (no IP yet), the member was left pointing into freed memory. On back-nav from the Connection step,cleanup()calledlv_timer_set_cb(nullptr)against the freed timer, corrupting LVGL's timer linked list and the heap region behind it. Downstream effects manifested as crashes inlv_event_mark_deleted,trans_anim_start_cb,lv_draw_sw_blend_image, and the wider LVGL anim/event/render paths. The member is now nulled at the start of the timer callback, before any path can early-return. ASAN-confirmed. -
Test Connection click crash — the libhv WebSocket
onopenlambda capturedconst char* urldirectly, but callers pass a stack-localstd::string's.c_str(). The lambda fires on a libhv worker thread after the caller's scope unwinds, and spdlog's format-arg processing strlen'd the freed buffer. Lambda now captures by string-copy. ASAN-confirmed. -
Connection step IP/port pre-fill — fields appeared empty on every visit (instead of showing 127.0.0.1/7125 or the saved config) and lost user-typed values on back/forward navigation.
init_subjects()now seeds buffers only on first init so user input survives re-visits, andcreate()drives pre-fill from the buffer (the source of truth) rather than the subject (which can be transiently empty during XML widget construction). -
AD5M Wi-Fi step on first launch — Forge-X firmware doesn't auto-load the Realtek RTL8821CU USB driver, and wpa_supplicant only starts after the user has supplied credentials via the stock UI. Fresh installs running HelixScreen as the launcher therefore never enumerated
wlan0and reported "no WiFi hardware found." Platform hooks now insmod8821cu.koand startwpa_supplicantinplatform_pre_start(), both idempotent. No-op on AD5M boards without a USB Wi-Fi dongle. -
Wizard subtree purge regression — the v0.99.45 fix to cancel style transitions during cleanup recursed into the root container and stripped its flex layout, collapsing new step content into a thin column. Root container now only gets
lv_anim_delete; descendants still get the full style strip. -
AMS
set_slot_infoclobberedmapped_toolwhen the caller didn't pass it; persists across calls now. Horizontal card scrolling on the AMS panel also restored. -
AddressSanitizer build for Pi — new
make pi-asan-dockertarget produces a fully-instrumented binary (LVGL, helix-xml, libhv, helix-screen) with-static-libasanfor self-contained deployment. Output tobuild/pi-asan/so it doesn't clobber regular builds. Strip is forced off; debug info is split into a separate.debugfile kept on the build host for symbol resolution. -
Wizard step-transition stress harness — Catch2 fixture (
tests/unit/test_wizard_step_stress.cpp) drivesui_wizard_navigate_to_stepprogrammatically in 2↔3 bounces and full-sweep patterns, configurable viaWIZARD_STRESS_ITERATIONS. Tagged[wizard][stress][.ui_integration]— hidden from default test runs.