ADB in the car: record the first test — control works, bulk transfer stalls on NCM loss

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DDwfd6Ki8UfG4hK6W5YpJz

docs/ADB-IN-CAR.md

@@ -91,8 +91,22 @@ emulators, and a tunnel there shows up a second time as `emulator-5554`.
 - **Wireless debugging** (Android 11+, pairing codes) survives better, but it only
   runs over WiFi, so it does not use this link.
 
-## Status
+## Status (2026-10-11)
 
-The NCM link and the app over it are verified (2026-10-09). The ADB path above is
-**untested in the car**; HyperOS may need its USB-debugging security options for
-installs over TCP as it did over USB.
+Tested with the tablet on the Orin, charging through the PD hub:
+
+- **Works:** `adb tcpip 5555` on HyperOS, the SSH forward, `adb connect`, `shell`,
+  `input`, `screencap`. The app retargeted to `ws://192.168.55.1:7700/` showed
+  *USB link up, bound*, the Orin's gateway with no mock banner, and a mark reached
+  the recorder (answered *no session*, correctly, with no CAN traffic).
+- **Bulk transfer does not.** The NCM link loses packets under load: ping loss
+  rose from 0% at 5 packets/s to as much as 32% at 20 packets/s, mostly small
+  packets, with no error counted on the Orin's side. TCP backs off (RTO ~10 s),
+  and throughput collapses to **~54 kB/s**. A 20 MB push took 6 minutes. The
+  165 MB debug APK install failed, and so will `flutter run`'s first install.
+  The app's own traffic (small, 5 Hz) is unaffected so far.
+
+Until the link loss is understood, install at the desk over USB, and use the car
+path for `logcat`, `shell` and screenshots. Suspects, untested: the NCM gadget's
+packet aggregation on the Orin's 5.15 kernel, the PD hub, USB power management on
+either side. The gadget also offers RNDIS, which would be the first comparison.