DS4 profile-output allocation gate — 2026-09-14¶

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

Release held on the original failure¶

The first final RC4.6 run failed WarmPolicyTransitionsAndWriterCompositionAllocateNothing (0) with 816 bytes against the unchanged zero-byte assertion. Evidence remains in isolated_results/rc46-publication/full-final.trx. That process was not profiled; passing repeats and earlier allocation-counter investigations were not accepted as attribution for this observation.

The failing USB case ran at 09:56:39.150–09:56:39.184. The eight enabled Xbox process-interop tests ran later, at 09:57:45–09:57:54, so their execution did not cause this earlier failure. The USB writer bypasses the new Bluetooth fresh- motor admission branch. Source review found no warmed allocation site, but that alone did not establish the cause.

Same-writer capture and positive control¶

An isolated harness loads the actual built test assembly and instantiates its private RecordingDevice, with report copying disabled. Pre-created delegates invoke the actual profile-policy, rumble and writer methods. Each window keeps the original 2,000 warmup / 20,000 measured Step calls: enable, publish nonzero rumble, Pump, disable, Pump, Pump. Construction and reflection occur before measurement; no hardware, installed app or production setting is used.

The diagnostic roots a 2,000,000-node BCL graph and requests a forced nonblocking collection before the counter window. Pressure allocations remain outside the measured workload. It preserves and stops at the first failed strict-zero window rather than averaging it away or retrying until success.

raw-usb-02 reproduces a 2,760-byte increase at batch 13 with zero object callbacks and zero recorder overflow; the managed thread remains 1. The native QPC envelope 608040659818..608040780160 contains GC-preparation SuspendStarted detail=7 at 608040755726, followed by suspension/resumption (native log lines 192–208). Collection counts remain 16,15,9: the collection started before this window. Native Begin/End bracket a slightly wider interval than the managed counters. The recorder's experimental gcDepth is not used.

The same process enables the actual recording sink's report Clone, after warming its List capacity, and retains the report. This positive control records 88 bytes and one array object, with a stack through System.Object.MemberwiseClone and DS4Device.sendOutputReport, with no overflow (lines 209–225). The exact-zero condition rejects this real allocation.

This is same-workload evidence for GC allocation-context accounting repair, whose runtime mechanism is documented in the Nintendo allocation investigation. It is not a retroactive object trace for the original 816-byte observation. The first raw-usb-01 capture had 256 zero windows and a valid positive control, but no failing window; it was not sufficient for attribution.

Narrow test correction¶

scoped-usb-03 repeats the same pressure, warmup and measured calls using the existing StrictAllocationMeasurementScope around only the warmed counter pair. All 256 windows measure exactly zero bytes and zero objects. The real report-clone positive control still records 88 bytes and one object. All 258 native envelopes, including initialization and the positive control, have zero overflow. Scope entry/exit failures are not ignored.

The ordinary test now uses this existing bounded scope and retains its exact zero-byte assertion. Three ordinary positive controls exercise the actual recording writer for USB, Bluetooth and full-duplex audio-owner cases and require the same zero-byte gate to reject a real report clone. The focused rebuilt suite passes 48 tests, zero failures. The subsequent ordinary, unfiltered rebuilt x64 run passed 5,717 tests, zero failures, three gated live-audio skips, including the eight opt-in isolated process-interop cases. Evidence: isolated_results/rc46-publication/full-final-confirmed.trx. No profiler or diagnostic-pressure switches were used in that full run. Exact-source CI and release-package validation remain separate gates.

No production code, GC policy, runtime/JIT configuration, allocation limit, feedback encoding, transport cadence or input timing was changed to clear this failure. The temporary no-GC reservation belongs only to the synthetic test process and surrounds a bounded synchronous counter window.

Reproduction and evidence¶

Harness and captured files are in isolated_results/rc46-allocation-audit/. After building Audit.csproj in Release, the relevant fresh-process commands were Run-Capture.ps1 -Name raw-usb-02 -Mode raw -Transport 0 and Run-Capture.ps1 -Name scoped-usb-03 -Mode scoped -Transport 0. Existing evidence paths cannot be overwritten; reruns must use new names. The launcher verifies the existing v3 profiler hash and removes the slow-only selector entirely. Profiling applies only to its new synthetic child, not installed applications.