Board: Kenik KG-5430TAS-ILC (Xiongmai OEM, sold in Poland). PCB marking IPM714513S-B VER.B; LED board JF-1R1W-B1EC VER.A1; OEM name IPC714557. Stock firmware V500.R013.E506.04AA.C0023.B002 (2025-02-19), vendor update ...B010 (2025-12-24). Novatek NT98566 (NA51089), 128 MB RAM, 16 MB SPI NOR (nvt,nvt_spinor at f0400000, 64K erase, no JEDEC id exposed). Runs OpenIPC nt98566-nor-lite, so the YAML's board.vendor/version is our firmware and its partition map is the split layout. main-app is our board daemon, /usr/bin/nightwatch (motion + IR lamp); majestic provides neither here. Stock map: loader 64K, fdt 128K, uboot 256K, linux 2M @0x070000, rootfs0 1M, app 12M @0x370000, mnt 512K, cfg 64K. Sensor: ipctool scans it as SC501AI, but the vendor firmware for this board ships nvt_sen_sc500ai with sen_sc500ai.cfg (i2c_id 0, rst 0x42, addr 0x30) and isp_sc500ai_* profiles. With the generic cfg OpenIPC ships (i2c_id 1, rst 0x44) there is no video. ipcinfo --short-sensor hangs or answers sc501ai here, so the load script gets SENSOR=sc500ai. Both names are plausible. Hardware facts in no datasheet: - Lamps are PWM channels 9 (IR) and 10 (white) via /proc/pwm/cmd, not GPIOs. An enabled channel is lit whatever its duty; only closepwm turns it off. The IR-cut filter is GPIO 170 (night) / 67 (day), pulsed. The PWM pads are not muxed by the kernel or the DTB: the vendor app writes i2c 0x1, sensor 0x220, pwm 0x21000000, pwm2 0x1000000 to /proc/nvt_info/nvt_pinmux/pinmux_set before the sensor driver loads. Without that the PWM block runs, every register reads correctly, and no lamp lights. - No photocell. The LED header is labelled CDS, but the ADC channels read flat in light and dark, and the vendor's sw_cds is computed from AET/AWB statistics. switchBy: picture therefore behaves oddly; sun plus coordinates is the reliable schedule. - No U-Boot environment anywhere: printenv is compiled in (713/8188), saveenv does not exist, and cfg is a provisioning block (magic 0x87654321), not an env. fw_printenv can never succeed - and anything that waits on it hangs load_novatek, so no modules load and the camera answers ping with no video. - Linux sees ~50 MB of the 128 (72 MB is the media zone); 2880x1620 needs ~35 MB of pools and is boot-racy, 1080p is stable. - The RTC is crystal-less ("RTC s/w reset timeout"). With CONFIG_RTC_SYSTOHC the driver spins one core at 100% system time; without it the board is fine. Platform facts, by cost of discovery: - majestic has no audio-capture backend here: 35 hd_* imports, zero hd_audiocap_*. A second process cannot supply it, because HDAL is single-process: the first ioctl of hd_common_init() (0xc014491f on /dev/isf_flow0) resets the shared kflow state and kills the streamer. What works is an in-process plugin (dlopen /usr/lib/novatek.so) with hd_audiocap_*, plus two ABI traps - libhdal is uClibc (32-bit off_t) under musl, so mmap offsets arrive as garbage (map /dev/nvtmpp yourself), and hdal_flow_log_p faults until the plugin exports a no-op. - Motion detection cannot be made to work: md_rects_recv_total stays 0 with someone waving in frame, kdrv_md loads but is never opened, HDAL reports max_out_count=1 with no MD capability bits, and kdrv_ive will not load (unknown symbol kdrv_ive_waitdone_nonblock). Expect to detect elsewhere. - The watchdog range is 1..89 s; a larger request is rejected by the core rather than clamped, and majestic then disarms - with its 300 s default there is no watchdog at all. - OSD is offered by the config schema but not implemented in the Lite build ("this build has no OSD"), and enabling it triggers a pipeline rebuild that can fail in the SDK teardown, leaving video dead while HTTP still answers (200, 0-byte JPEG) until the streamer is restarted. Not included: a UART boot log (ours was lost in a cleanup; the board's 3-pad UART is easy to probe, so one can be supplied on request) and a flash dump.