Nintendo rumble: controlled comparison, September 10¶

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

This is a private validation record, not release notes or a smoothness claim. The user extended unattended work to one hour at approximately 12:56 UTC; the current work window ends at 13:56 UTC on September 10, 2026.

Latest status¶

• Complete private D is prepared from c6a763d, with all 551 files verified; neither app is running and no installed files were changed.
• SDL-style first-active held-body output and two bounded Bluetooth lifetime/ retry corrections are implemented. The final full suite passed 5,074 with 11 gated skips, twice. Strict allocation checks remained enabled.
• The environment-gated B experiment below is historical and has been removed from current production code. The later first-active section supersedes it.
• Physical smoothness still needs a controlled hands-on comparison. A separate exact-native Bluetooth burst-coalescing defect is reproduced and remains open.

The sections below preserve the chronological evidence and its limitations.

Established starting point¶

• The September 9 installed RC4.5.4 Bluetooth left Joy-Con Heavy trace had one unchanged three-active-subframe waveform, no counter discontinuities and no intervals over 20 ms. The user still reported crackling. See the preceding evidence.
• Complete private candidate A uses DS4Windows checkpoint db78c9d and the independently built VIIPER candidate corresponding to checkpoint 9999f25. All 551 composed payload files, the runtime/dependency closure and Xbox identity were verified. It is not an incomplete publish-directory handoff.
• A changes standalone Joy-Con 2 Bluetooth held servicing to 10 ms, retaining the waveform bytes. Pro/joined Bluetooth remains 15 ms; USB remains 12 ms. Actual scheduling and input-tail measurements are still required.
• A connected a right Joy-Con at 90% battery at 07:05 local. It was closed at 07:07; the installed build subsequently connected a Pro. Therefore the user's tentative "slightly better" feedback cannot yet be attributed to a controlled before/after comparison. The earlier left Joy-Con's 10% battery and the different physical half are additional comparison confounders.
• With explicit user approval, A was resumed at 12:24 UTC using its existing isolated data and verified immutable payloads. Authenticated backend readiness and Bluetooth discovery succeeded. At the time this section was written, no controller had woken in the resumed session.

One-variable packet experiment¶

The pinned implementations do not establish one universally correct held layout:

The B experiment is disabled by default and requires both portable-lab mode and the exact cold setting DS4WINDOWS_SWITCH2_HELD_SUBFRAME_EXPERIMENT=neutral-tails. Only standalone Joy-Con 2 Bluetooth lifetimes can opt in. Only ordinary Test Heavy/Test Light preview output is transformed. The first subframe, every control/frequency field, strength setting, counter, cadence, ownership and freshness metadata remain unchanged; only the second and third subframe amplitudes become zero. The first write and held refreshes use the same transform.

Pro, joined Joy-Cons, USB, normal launches, game feedback, profile effects, source-preserved rich previews, finite one-shots, PCM, native packets and impulse/release presentations are excluded. There is no shared-codec change, amplitude boost, inferred firmware slot duration, or claim of equivalent actuator energy. This experiment may feel weaker or worse; it is not a fix until compared on the same physical controller under matching conditions.

Failing-first fixtures reproduced five expected trailing-amplitude failures with three unchanged controls passing. After implementation, 33 new cases and 188 combined targeted cases passed without skips. They include actual sink to BLE encoder bytes, cold factory configuration, finite/repeat distinctions, uncertain exact retries, Stop/no resurrection, default identity and strict zero allocation with an allocation-positive control. The subsequent unfiltered full run passed 5,040, failed 0, and skipped 11 explicitly gated hardware/cross-runtime rows. Allocation checks remained enabled. Evidence is isolated_results/held-subframe-experiment/full/held-subframe-full-reviewed.trx. No subjective or physical hardware acceptance is inferred from that result.

Measurement and acceptance rules¶

1. Identify the actual executable, transport, physical half, pair state, profile and battery before every comparison. Do not mix installed and lab results or compare a depleted half with a charged different controller.
2. Record accepted physical-report publication counters before/after idle and rumble windows. Compare deltas only within matching runtime, physical and clock generations. Cumulative maximum gaps cannot be subtracted.
3. Observe bounded host HCI windows with start and Stop inside the capture. Check waveform identity, active subframes, counter progression, write gaps, explicit neutral output, and absence of later active output. Host evidence is not radio-delivery or physical motor completion evidence.
4. Change cadence and packet shape in separate comparisons. Keep strength and carriers fixed. Ask for tactile confirmation rather than infer it from green software tests or a lower average interval.
5. Reject a candidate that worsens input tails, loses finite effects, changes unrelated feedback, or fails to stop. No new public release is authorized by this private experiment.

The private observer now has separate strict parsers for a Joy-Con-shaped 36-byte event402 and a Pro-shaped 52-byte event402. The latter records the two protocol groups separately. Synthetic malformed/layout/counter/neutral tests pass and independent review found no concrete capture blocker. Both modes retain the 25-second owned-session deadline, bounded rows, no raw ETL/payload, no controller writes and explicit host-only limitations. Missing/unknown events are inconclusive, not evidence that rumble was absent.

Resumed A: measured host output, before the first-active-frame change¶

The resumed mapper is PID 416, broker PID 13920, from the complete A folder. It connected a standalone Bluetooth right Joy-Con 2 at 12:33 UTC and a Bluetooth Pro at 12:34 UTC, both reporting 90% battery and using the isolated ds profile with virtual DualSense output. No physical left Joy-Con was connected in these captures.

Both protocol groups retained the same waveform throughout each capture: three active subframes, maximum amplitude code 453, no counter discontinuity, no observed neutral packet, and no ETW event loss. These are host-stack observations, not actuator timing or proof of a firmware timeout. The user's small perceived skips remain unexplained by these measurements alone.

During the roughly 40.7-second Heavy input comparison, clock/runtime/physical generations remained unchanged. Pro accepted 2,729 reports with 17 publication gaps above 20 ms; right Joy-Con accepted 2,691 with 25 above 20 ms. Neither had a gap above 50 ms in that delta window, and neither recorded a busy publication drop. Earlier cumulative maximum values are not this window's maximum.

The automatic Light-click capture (pro-light-A-0741) had no matching output; the user's subsequent manual click produced the measured Light capture above. The unattended Joy-Con-shaped observation (joycon-A-unattended-1305) also had no matching packets. These are inconclusive for the corresponding motor, not evidence of output loss. The intended Stop was not captured in either primary Pro window. A later unattended UI Stop attempt did not change the displayed Stop Light state or establish a neutral output, so physical Stop acceptance is explicitly not verified here.

Requested first-active-frame implementation¶

After asking to focus on cadence while away, the user explicitly requested the SDL active-frame change as well. Shape and timing remain separate changes. The earlier B build at 7011175 was fully composed and prepared but never launched; its gated preview experiment is not a hardware-tested result.

The production candidate applies first-active plus zero-amplitude tails only at the Switch 2 delivery sink, for ordinary canonical body-compatible Apply frames. All four arbitration origins are eligible. It retains the current first subframe, control fields, gain, lifetime/sequence metadata and cadence. Source-preserved/native/PCM/one-shot/adaptive designs, impulse and release presentations, raw nonzero trigger lanes, Stop and Neutral are excluded. The shared compatibility-group builder is deliberately unchanged because it also provides body samples for richer DualSense mixing.

This is a BLE-shaped adaptation, not byte-identical SDL USB: the USB donor zeros the entire trailing words, whereas the BLE donor retains neutral carrier fields. Our actual sink coverage is Pro USB and standalone/joined Joy-Con 2/Pro BLE; this change does not create a Joy-Con USB transport.

Other donor mechanics were reviewed, not blindly copied: SDL USB services rumble from its input-read loop with a 12 ms gate and a 24 ms joined-child workaround; the BLE donor uses a 10 ms update pump. DS4Windows keeps its independent output-only maintenance worker, immediate Stop/new-frame admission, and exact uncertain-payload retry/counter ownership. No donor comment establishes the playback duration of an individual subframe.

The failing-first run caught 18 expected old-layout failures with 30 unchanged controls passing. After enabling the sink policy, all 48 focused cases passed. The first full run caught two older integration helpers that explicitly required three identical active frames. Their body-only assertions now check one active frame plus exact zero-amplitude, carrier-preserving tails; rich frame assertions were not weakened. The reviewed full run passed 5,055, failed 0, and skipped the same 11 gated hardware/cross-runtime cases. Strict zero-allocation and allocation-positive controls remained enabled. Evidence: isolated_results/held-rumble-production/full-reviewed/first-active-full-reviewed.trx.

The 18 new composed scheduling rows exercise the real worker and sink with a controlled physical-writer clock at 10/12/15 ms. They cover service duration, early/late callbacks, competing publishers, bounded contention retries, exact uncertain retries and Stop sealing rearm. They establish no new timing defect and are not measurements of Windows radio or actuator scheduling.

At 13:07:35 UTC the old A mapper/broker processes were terminated after the observed UI Stop Light and close inputs did not take effect. Exact retained PID/start-time/path/hash checks scoped termination to that private session. No installed files changed. This ended the old host output stream; it is not a clean Stop test or physical neutral acknowledgement. No new build has been launched, and no user wake/manual test is requested while the user is away.

Native DualSense rapid-fire: new cross-process software evidence¶

The bounded synthetic harness ran against the actual A mapper assembly (FA2F033F7A5375B662D8DC6885C05B24681A7FBACE0AC9AF26397F93CBA84B45) and reviewed VIIPER binary (DC2D47B49F94FA827903FD24F18AE0289EBE682F09D6EF67A09EE7A6A8005EB7). It uses the authenticated production client, USB/IP input commands, V5 reader, four-slot native lane and 64-slot physical ring, terminating at a synthetic USB physical-writer hook with no HID handle or Windows controller attach.

All 32 broker-accepted commands reached that hook exactly once, in order, through a deliberately observed full queue held for at least 40 ms. Short stops, duplicates, independent trigger/LED validity and truncated trigger groups passed. There were zero compatibility fallbacks and callback failures. The feedback workers joined, the private device was removed, the owned broker exited, and its generated key was deleted. Result: Desktop/Controller-Diagnostics-2026-09-10/native-burst-A-1312/SUMMARY.json. Harness assembly hash: CB4B9776200039C3A19CCE347617D05D15914B26EFE1026D84DF74CD351853E5.

This is stronger cross-process software evidence for the prior burst fixes, not a GTA V Enhanced gameplay test, physical actuator confirmation, Bluetooth transport test, media-fairness result or arbitrary-overload guarantee. These private fixes remain outside published RC4.5.5.

Confirmed BLE write-lifetime defect, separate from perceived crackle¶

The adapter formerly passed its 100 ms cancellation token into the WinRT write and marked output idle as soon as that managed wait ended. The pinned CsWinRT bridge cancels its managed task separately from native completion. Read-only metadata inspection corroborated that mechanism in our compiled SDK projection. Thus timeout could admit another send or teardown while the original native write was still pending. The captured steady rumble windows do not establish that this timeout path caused their tactile skips.

The correction retains the uncancelled operation and copied payload while bounding the output caller's wait to 100 ms. An exact retry consumes the late receipt without duplicating the native call; another payload cannot overtake a pending operation. Admission and its drain obligation share the short lease state lock, but neither platform entry nor waiting holds that input-facing lock. Teardown's public observer stays bounded while actual resource release waits asynchronously for native completion. No global WinRT semantics or cadence constants changed.

Eight reproductions failed before the correction. The initial combined lane passed 135/135; six additional immediate-completion/admission-race/physical- counter cases brought the focused set to 14/14. Independent concurrency review found no concrete blocker. The unfiltered combined full run passed 5,069, failed 0, and skipped the same 11 gated cases. Evidence: isolated_results/rumble-completion/full/completion-full.trx, SHA-256 BD46AE162CD7B8F91A4510D572EA603F424346693CFA74F3EC35245ECCFCA725.

The real WinRT adapter already allocates detached buffers and projected tasks; this change replaces the per-call CTS/timer with a drain completion source and an observer for incomplete operations. No real-adapter zero-allocation or net allocation improvement is claimed. The existing strict fake-lease hot-path allocation tests remained enabled and passed.

Complete portable C and repeated native burst check¶

The complete private C candidate was built from clean source fd4ac13c4b7887f56517fa17162d3f8ab609cac9, with the first-active-frame and native Bluetooth write-lifetime fixes. Its 551-file manifest passed the full runtime/dependency/Xbox-identity closure verifier. It is prepared, not launched; no new hardware or tactile result is implied. Location: Desktop/DS4Windows-Haptics-Lab-20260910-C/portable/DS4Windows.

• DS4 assembly: E7534285182DAEC6B6F3703B29FBE2DA752A63359FED1D118BE838FD2405588D.
• Candidate broker: DC2D47B49F94FA827903FD24F18AE0289EBE682F09D6EF67A09EE7A6A8005EB7.
• Private manifest: 122C634B9267A1F8A0A6D11DFBEF1EC84907B17023570EC422227E1C174BDF72.

The same native cross-process harness passed against C: 32 accepted commands, 32 ordered synthetic USB physical publications, zero compatibility fallbacks or callback failures, and a measured forced-full stall of 55.9739 ms. Owned workers and broker exited and the private device was removed. Evidence: Desktop/Controller-Diagnostics-2026-09-10/native-burst-C-1323/SUMMARY.json. The Bluetooth-helper and PCM/gameplay limitations of the A run also apply to C.

Reference cadence mechanics differ beyond nominal interval constants. SDL USB's shared HID rumble thread adds a 10 ms sleep after a write; the BLE donor's 10 ms gate runs inside its update pump, not an autonomous 100 Hz timer. Our worker anchors its next held refresh to successful write start plus the interval, subtracts service time once, and does not emit catch-up bursts. No firmware subframe playback duration or periodic silent gap was established.

Actual-worker cadence isolation, without a physical transport¶

The separate hash-pinned harness executed C then A for about 30 seconds each. It composes the real runtime, preview renewal, lifetime, sink and maintenance worker, substituting only a successful generation-bound BLE lease. The dormant worker is replaced with the same worker wrapping the actual service delegate for bounded observations. It requests and releases the same scoped 1 ms timer period as app startup; both APIs returned success. No controller, broker, UI, Windows transport or input slot is opened. Raw heavy strength is byte 100, not the UI's 100 percent setting. Each condition lasts 7.5 seconds.

Every phase had zero write intervals above 20 ms, zero neutral payloads, counter discontinuities, waveform changes, overflow or worker failures. C's Pro groups each had one active subframe; its right Joy-Con group had one. A retained three per group. Publishers joined and fake lifetimes retired; no writes were observed in the subsequent 40 ms. That bounded observation and idle wrapped service are not proof of joining every Timer callback.

These short sequential trials do not establish a performance advantage for C, radio/firmware continuity, tactile smoothness or behavior under arbitrary load. They did not reproduce a preview-renewal or composed scheduling stall under the tested quiet and neutral-publication conditions. The full WinRT adapter, GATT completion and physical input workload are intentionally absent.

Evidence (folder labels are not authoritative timestamps; JSON records UTC): Desktop/Controller-Diagnostics-2026-09-10/actual-schedule-C-1334/SUMMARY.json (13:32:48-13:33:18 UTC) and Desktop/Controller-Diagnostics-2026-09-10/actual-schedule-A-1335/SUMMARY.json (13:33:29-13:34:00 UTC). Harness SHA-256: 6BB2FE21855FFB5A092A98A8AE844FDF013BE3C6F08989B67690F8E27F3CCCA9.

Open native Bluetooth ordering defect reproduced independently¶

The C USB burst result does not cover the physical Bluetooth helper. A separate isolated fixture invokes the real HelperHost.ReceiveGameStateAndTemplate, actual pacer loop and physical writer, with synthetic native I/O and no device handle. Both exact commands are admitted while presentation is paused, then the real presenter is started. C reproduces two ordering failures:

• Rumble pulse 91/173 followed by Stop: only the zero-motor Stop is submitted.
• Right-trigger A followed by B: only B is submitted.

Both spaced positive controls submit the two expected reports in order. The helper's pendingControllerState accumulator merges the commands before presentation; the existing validity composer intentionally replaces an earlier valid field with a later valid field. This is a confirmed native-command loss boundary, not proof of the GTA reporter's cause or of USB PCM behavior.

Evidence: isolated_results/dualsense-native-bt-ordering/baseline-c.jsonl, hash-pinned to the same C assembly above. These deliberately failing diagnostic cases are isolated from the normal test suite; they must not be represented as fixed because the existing main suite is green.

A correct fix needs a bounded exact-native admission/acknowledgement domain, unchanged local latest-state behavior, immutable head/retry ownership, both idle 0x31 and media-piggyback coverage, and lifecycle cancellation. Merely blocking the shared helper command reader on a full FIFO risks starving Clear/Stop, templates and media. This larger correction is not included in C.

Narrow idle native Bluetooth retry correction after C¶

A fifth isolated C case reproduced a separate retry stall: one accepted native command encounters an occupied oldest physical-writer slot, that credit later returns, but the helper emits nothing for one second. Its rejected-write path had manufactured controllerStateReportsAhead = 1 without any queued media, so the command could only recover when another source command or media arrived. Evidence: isolated_results/dualsense-native-bt-ordering/baseline-c-idle-retry.jsonl.

The correction grants the one-media-frame fairness yield only when an actual speaker report is queued at the head. With no queued media it leaves the current ordering boundary unchanged, normally zero for an idle claim. In particular, it must not clear a newer microphone boundary published concurrently during physical I/O. This is separate from the still-open exact-native coalescing defect.

Five normal-suite tests exercise the actual helper/pacer/physical writer with synthetic native I/O: recovered idle credit, real queued-media fairness and subsequent piggyback, Clear into a new epoch, Stop joining before late credit, and preservation of a concurrent two-media-frame microphone boundary. They passed 5/5. The unfiltered Release x64 suite passed 5,074, failed 0, and skipped the same 11 gated cases (total 5,085). No real controller was opened for these tests. The complete C candidate predates this narrow correction. An independent repeat of the full suite produced the same counts. Durable TRX: isolated_results/dualsense-native-bt-ordering/full-suite/native-idle-retry-full.trx, SHA-256 27C3762397622F12BC955FC493C0F1F635EE15757A8608CEE159060B49640E44.

Final private D candidate and work-window handoff¶

Complete portable D contains all three corrections from this window, including the narrow idle retry fix, built from clean source c6a763ddf7ecfa151fd055f10249e23519c89b4c. Independent review found no concrete blocker in the final idle-retry guard, its concurrent microphone-boundary preservation, or the first-active policy's rich/impulse/finite protections.

• Portable: Desktop/DS4Windows-Haptics-Lab-20260910-D/portable/DS4Windows.
• App SHA-256: F80F1378A80F927E8C730DBE77EFCDC1A5DAFD01F2BEF80A4AEB77EF04331803.
• Broker SHA-256: DC2D47B49F94FA827903FD24F18AE0289EBE682F09D6EF67A09EE7A6A8005EB7.
• Manifest SHA-256: 4B60033271DE26D3E517D97727D5A5F37B4537E5EB56069C474ACF31BB50DF10.

Both full composition and the separate preflight verified all 551 files, including the bundled broker, self-contained runtime and Xbox identity. Fresh isolated data was prepared without launching either application. D is a private test candidate, not a published release or a claim of production haptics parity. The composition's base ZIP is an intermediate with the released broker and is not the final lab or a handoff artifact; only the final portable directory contains the pinned candidate broker.

The pinned D binary also passed the same cross-process native USB burst check: 32 accepted commands, 32 synthetic physical publications, zero compatibility fallbacks or callback failures, with an observed full-queue stall of 42.8868 ms. Workers joined, private device removal succeeded, the owned broker exited, and the generated key was removed. Evidence: Desktop/Controller-Diagnostics-2026-09-10/native-burst-D-1351/SUMMARY.json. This still does not cover the confirmed Bluetooth native-command coalescing defect, physical GTA gameplay, PCM fairness, or actuator smoothness.

No installed files, drivers, profiles or public releases were changed during this window. No controller wake or further manual test was requested after the user left. All owned captures and synthetic runs ended; both applications are left closed. Hands-on comparison of D's rumble remains outstanding. The next software task is the explicitly bounded native-command credit/ordering design, not an unvalidated FIFO or a claim that the GTA report is fully resolved.

Later Bluetooth source follow-up¶

The subsequently requested Bluetooth command-ordering work is recorded in the dedicated validation ledger. It corrects the identified native coalescing boundary and passes the expanded 5,185-test suite. D and public downloads still predate that source checkpoint; hardware GTA/Bluetooth confirmation remains outstanding.