Issue 81: separate native-feedback pacing from local settings¶

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

This follows 47cc945 (conservative pending-repeat suppression). It does not replace ordered feedback with latest-value state and does not relax unknown trigger-byte preservation.

Source change¶

• Native Bluetooth feedback has a separate 500 Hz / 2 ms submission ceiling. Local profile/settings updates retain their 200 Hz / 5 ms policy.
• Native deadline waits use the existing high-resolution timer, with stop/new work interruption and a final readiness recheck. They add no spin window.
• Due media, startup/prime ordering, microphone barriers and retained-command retries still take priority where required. Only accepted writes consume a native command or advance its submission timestamp.
• Native-bearing physical writes are explicitly tracked, including native state piggybacked onto media. Exact completion, failure and lifecycle rules apply to those writes. The production pending allowance remains the existing 32 slots; it is not reduced to one outstanding command.
• An optional smaller native-credit limit exists in the internal writer seam for adversarial tests. Delayed Windows IRP completion is not a Bluetooth or actuator acknowledgement; a one-credit production gate could worsen delay.
• No changes to input mapping, Nintendo translation, PCM gain/codec, persistent rumble mode, trigger ownership, descriptor-sized writes or VIIPER protocol.

Automated evidence¶

isolated_results/issue81-native-rate/ contains the local TRX evidence.

• native-rate-first.trx: exposed an old deadline assertion and an unintended early completion poll of unrelated media. The deadline regression now covers both old/new rates and credit exhaustion. The writer now skips completion polling when there is no pending native write; the original Busy/media fairness test passes unchanged.
• native-rate-focused-green.trx: 527 passed, zero failed/skipped.
• native-rate-full-green.trx: 5,538 passed, zero failed, 11 existing gated skips (5,549 total). Allocation assertions remained enabled.
• Fixed-deadline pressure tests retain every distinct rumble command, A/B/A trigger transition, final zero and all 128 finite PCM frames, including zero frames. Offered 250 Hz under a 500 Hz budget and offered 500 Hz under a 1,000 Hz test budget pass. The old 200 Hz budget under offered 500 Hz demonstrates accumulated queue delay as a positive bottleneck control.
• Initial synthetic run: offered 250/budget 500 had 1.506 ms maximum effect age and queue high-water 1; offered 500/budget 1,000 had 0.623 ms maximum age and high-water 1. Old budget 200 accumulated 1,564 ms final-stop age. These are scheduler/test-hook observations, not measured Bluetooth or game latency.
• Writer tests cover credit exhaustion, media progress, 0x31 and native-bearing 0x36, sync/async completions, failed/short completion, cancellation/disposal, unchanged default allowance and zero-allocation warmed readiness checks.

Validation boundary¶

The 500 Hz value is a submission ceiling, not a guaranteed sustained service rate or hardware capability claim. Hardware comparison and GTA reporter confirmation remain pending at this checkpoint. No release or issue closure is authorized by these automated results alone.

Connected Bluetooth comparison¶

After the source checkpoint, complete portable RC4.5.8 and candidate packages passed the independent 553-file / 297-dependency verifier. Both used the same released VIIPER binary and a cloned profile selecting virtual DualSense. Only the test copy's output type and idle timeout were changed; the Roaming profile and installed files were preserved. The initial installed session reached its existing 900-second idle timeout before the first replay; the user reconnected the controller after the baseline portable was ready.

The probe opens only an explicitly inventoried virtual DualSense HID interface whose parent chain has the USB/IP UDE service/hardware identity. It sends one second of modest, distinct body-rumble commands at 250 Hz, then an explicit zero. It never directly writes the physical controller. Physical submission traces are captured from the existing helper's bounded diagnostic arrays using non-suspending, read-only inspection. Both 0x31 and native-bearing 0x36 carriers are included; media copies of an already accepted amplitude are not new effects.

An initial probe using Sleep(1) ran at only about 67 Hz on this host. Its 68 commands including stop were preserved, but it is not the pressure test. The diagnostic scheduler was corrected to use a high-resolution waitable timer without spin or global timer-resolution changes; its no-HID timing check produced 250 wakes in 1,000.47 ms. Both comparison replays below have 250 body commands plus the stop, and comparable roughly 4.1 ms source intervals.

The baseline physical command interval was about 5 ms and accumulated delay; the candidate followed the offered roughly 4 ms stream and drained. Media submission intervals remained about 10 ms with occasional 20 ms gaps in both replay windows. No short physical-write completions were observed. The actual physical HID descriptor write length is 547 bytes; logical 78/398-byte packet lengths must not be mistaken for the Windows write size or changed to force higher throughput.

Evidence is under the Desktop DS4Windows-Issue81-Lab-20260912 directory: baseline/samples/burst250-highres, candidate/samples/burst250-highres, and their inventories, replay requests, identity records and live metrics. Offline analysis lives in isolated_results/issue81-live-metrics/. Candidate archive SHA256: 3B335B43C741FFBCBEF0A0C4B3D51699402A7C95DE1E3A9F0FCFC24A72FFB8A0.

This establishes the measured body-rumble transition sequence and a substantial host-queue improvement on the connected Bluetooth controller. It does not measure radio arrival, actuator response, trigger payload fidelity on hardware, or GTA's perceived smoothness. Trigger/opaque-byte preservation is covered by the source regression suite, not this rumble-only probe. The first input histogram windows span unequal idle periods, so they cannot establish input latency parity. GTA reporter confirmation remains necessary; no release or issue closure follows automatically from this result.

Two additional candidate replays also preserved 251/251 measured amplitude transitions each. Median source-start to submission was 0.1677/0.1658 ms, maximum 0.4521/0.5298 ms, and final neutral 0.1511/0.1765 ms. Each peaked at one awaiting command and drained to zero. The first repeat's observation window ends immediately before the deliberate second replay, avoiding mislabeling that later authorized rumble as a resurrected effect.

Candidate-only idle/load input checks retained identical process, device and histogram identities across six snapshots, with exact bucket/count deltas. Intervals above 5 ms were 3.59% idle / 3.95% load in the first pair and 4.96% idle / 3.87% load in the second. No new interval above 50 ms was observed; binned p50/p95/p99 upper bounds remained 2/5/20 ms. This is a bounded gross-stall check with no consistent load-correlated worsening, not baseline parity. Heap scans took 544–648 ms and approximate midpoint windows were 1.85–2.04 seconds, so neither precise one-second isolation nor improved input throughput can be claimed. Evidence: candidate/samples/input-windows and samples/final.