Captures from macOS 26.5 (Apple Silicon MacBook), device "AGI Bench 1.2.5" - a
BLE HID composite (keyboard + consumer + mouse, HID over GATT, nRF52840) paired
over Bluetooth. Same firmware and descriptor that drops Shift on iPhone.

hid-descriptor-raw.txt       ReportDescriptor hex pulled from ioreg for the device
hid-descriptor-decoded.txt   same descriptor decoded item by item (report IDs 1/2/3)
macos-raw-hid-reports.log    IOHIDManager capture while the device typed "A" then "xAx".
                             The "A" arrives as one report: 01 02 00 04 ... i.e. report
                             ID 1, modifier 0x02 (left shift), keycode 0x04. Each report
                             shows up 3x because macOS exposes the device as multiple
                             IOHIDDevice services.
macos-eventtap.log           CGEventTap for the same keystrokes: flagsChanged keycode=56
                             SHIFT=1, then keyDown keycode=0 with SHIFT=1 - macOS applies
                             the modifier and TextEdit shows "A" / "xAx". On iPhone these
                             same reports produce "a" / "xax".
two-service-split-firmware.patch  the isolation experiment: same keyboard-only service
                             that works when it's alone, plus a separate mouse-only HID
                             service on the same device. Cursor works on iPhone, Shift
                             still dropped.
capture-tools/               the small tools used above (IOHIDManager monitor, event tap,
                             descriptor decoder, and the CoreBluetooth helper that sends
                             the device its "type" command).