Joy-Con 2 counter-reset disconnect: live evidence, 2026-09-08¶

Historical record. This preserves the source, tests and limitations recorded at the time; it is not current setup guidance. See the documentation index.

Incident and capture¶

The user reported a left-only disconnect at 19:11 and a joined-pair disconnect at 19:33, then authorized a process dump and requested a fix. All times below are machine local (UTC-05:00). The running portable process was DS4Windows PID 34488, version 5.0.4.93-Bluetooth-Audio-Lab, launched September 7 at 23:51. VIIPER PID 30276 also retained its original launch time. Both were responsive.

The app log records the joined pair ready at 19:20:32, then its virtual Xbox 360 disassociated at 19:33:14.342, USB/IP detached at .418, and output removed at .493. VIIPER records client stream closure followed by explicit remove commands, not a preceding broker error. Windows System logs contain no event in the queried incident window. These logs alone did not establish the cause.

An elevated, hidden dotnet-dump collect --type Heap -p 34488 captured the still-running app at 19:38, without restarting either app, reconnecting a controller intentionally, or sending controller commands. Capture side effect: the heap snapshot paused the target long enough for its USB Pro to time out at 19:38:39; normal USB discovery reattached it at 19:38:42.997. This later diagnostic-induced pause is not either earlier Joy-Con incident. No second live dump was taken. The dump is 405,359,237 bytes and remains local only on the Desktop under Controller-Diagnostics-2026-09-08. It contains process memory and must not be committed or uploaded as a public fixture. Only the non-secret protocol fields below are recorded in source.

Exact retained evidence¶

The latest joined owner has runtime generation 23, pair epoch 24, left device / transport generations 19/20 and right 21/22. Its right input owner ended with SinkFailure; the right drain pump recorded SinkRejected and one rejection. The left owner was subsequently stopped normally. Neither Windows input lease had disconnectObserved; both input overflow counters were zero. The runtime had bluetoothDisconnectRequested=0.

The input queues retain body bytes and timestamps after resetting head/count. Matching the right input session's last completion timestamp to retained ring index 9 recovers the exact rejected packet. Index 8 is the last accepted right observation, also retained by the joined mapper:

Host arrival advanced normally by 14.8609 ms on the same exact transport lifetime. The joined sink's runtime-admission wait count is zero. Its final generic failure fields were cleared by successful terminal cleanup; those zeros do not mean the earlier input succeeded.

The same dump independently retains the previous standalone left failure: device / transport generations 13/14, one input rejection, SinkFailure, no queue overflow, no Windows disconnect observed. Its standalone sink preserves lastJoyConMappingFailure=BackwardOrOutOfOrder. The session's retained counter high-water is 1,431,651; ring index 0 matches its last completion timestamp 2,296,057,651,119 and contains counter 12 (0C 00 00 00).

Root cause and correction scope¶

The physical report counter can reset without a transport disconnect. The session already records a modular discontinuity diagnostically, but the Joy-Con profile admission (also used by the joined coordinator before pair staging) treated it as out-of-order input and rejected the publication. The input owner's exception boundary then retired the controller; joined ownership necessarily retired both halves.

This is the same class already handled for ProController2 by exact-lifetime host-arrival ordering. Apply that policy to validated Joy-Con 2 Bluetooth Common05 reports across session, profile admission and pairing. Retain counter values/classification for diagnosis, and advance the diagnostic baseline after a discontinuity. Do not weaken framing, descriptor identity, transport-generation, host-QPC monotonicity, or unsupported-transport checks. Do not drop button transitions, add buffering, increase queue limits, or guess a firmware counter modulus from two observations.

Regression coverage must include both captured reset values; standalone left and right; joined reset on either half; recovery on subsequent input; stale host timestamps and lifetimes still rejected; independent report rates; and zero-allocation input processing. One-shot retained failure details should survive cleanup so a future incident does not require reconstructing a heap.

Reference alignment¶

The locally pinned Switch2Connect 61ac6642ce12fe7217e38a860b14863b18ca7e28 input notification callback (src/controller.py, around line 4000) builds controller input directly from arriving notifications rather than requiring this raw counter to advance. SDL-hifihedgehog d98c5804a9d20b0d96e993741797878c86b8f1e1, src/joystick/hidapi/SDL_hidapi_switch2.c:1188, timestamps delivered controller state with SDL_GetTicksNS() before dispatching mini/combined Joy-Con and Pro buttons/axes. Sensor timestamp interpretation is separate. These source paths support using arrival order for presentation; the captured packets on this machine, not an assumed donor counter modulus, establish the reset defect.

Implemented counter policy and offline validation¶

The shared Switch2CounterSequence.UsesArrivalOrdering policy now covers Pro USB/BLE Common05 and Joy-Con 2 left/right BLE Common05. It excludes unknown models, Joy-Con USB and dedicated reports 07/08. Both Joy-Con profile-admission counter gates use the shared policy; session/replay and raw-stick calibration already consume that policy. The pair reducer itself has no raw-counter gate and retains its generation, clock and skew checks unchanged.

The raw modular sequence/delta remains observable, including backward jumps. Accepted arrivals advance its next diagnostic baseline; there is no guessed reset threshold, special treatment of 1.431 million, or inferred modulus.

Switch2JoyConCounterResetTests initially reproduced the defect: 10 failures and 2 passing safety cases in switch2-joycon-counter-reset-before.trx. The failing cases included standalone and joined publication on both halves, all four standalone holding layouts, and subsequent session/replay baseline classification. These fixtures contain only the recovered counter pairs and right-side QPC values; button/stick data and successor reports are synthetic.

After correction, the expanded 17-case fixture passed, including arbitrary forward/backward jumps, duplicates, full uint wrap, repeated resets and resumed advancement across all four supported model/transport combinations. A warmed joined-runtime test processed 20,000 discontinuity-bearing inputs with zero managed allocation. Both standalone and joined paths remain Active and publish button release/next input instead of requesting terminal teardown.

The broader source/profile/coordinator/replay/calibration suite passed 147/147, no failures/skips, in switch2-joycon-counter-reset-green.trx (Release/x64). An intermediate run had 145 passes and two obsolete calibration expectations that still demanded counter rejection; those now verify accepted reset samples while retaining rejection of duplicate host timestamps. Host-QPC regression, stale generation/descriptor, unsupported report and wrong-epoch tests remain enforced.

This capture proves the two observed failures, and offline regressions exercise their corrected path. A live patched-build endurance run is separate acceptance evidence and has not yet been performed.

Retained failure evidence and normal lifecycle diagnostics¶

Both standalone and joined Bluetooth input sinks now retain their first input rejection before successful half-loss or terminal cleanup changes their ordinary status fields. The snapshot records the model and sink/profile/pair failure classifications. If a sink dependency throws, only its exception type and HResult are retained. The runtime separately records the first swallowed gyro observer, Report subscriber or queued-action exception using the same bounded metadata; existing rejection and terminal-neutral behavior is unchanged.

A once-per-owner lifecycle observer queues a normal diagnostic off the input thread. It reports side/model, runtime generation, end reason, pump failure and the preserved cause for unexpected disconnect, queue overflow, sink rejection or worker failure. Ordinary intentional Stop with no pump fault remains quiet. No successful report formats a log entry or allocates a failure snapshot. The snapshot stores no raw packets, controller addresses, bond keys, exception messages, exception objects, file paths or stack traces. The first evidence is not cleared after the message is queued or after cleanup succeeds.

Switch2InputFailureEvidenceTests: 12 passed, no failures, in switch2-rumble-maintenance-complete-final.trx (Release/x64). Coverage includes concurrent first-fault retention; deferred, once-only reporting and reason filters; privacy of exception messages; exceptional Report callbacks on both standalone halves and the joined sink; cleanup/recovered-input preservation; and separate gyro/queued-action fault stages. These results are offline source validation; deployment and a live endurance test remain separate evidence.