Native Bluetooth write-credit allocation gate — 2026-09-12¶

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; same-workload attribution required¶

The first RC4.5.9 rebuilt full run failed NativeCreditProbeAndSynchronousWriteAllocateZeroAfterWarmup: 7,328 bytes against its unchanged zero-byte assertion. The run ended 5,537 passed, one failed, 11 existing opt-in skips. Its evidence remains in isolated_results/rc459-publication/rc459-full.trx. That original process was not profiled; subsequent observations do not supply an object trace for that event.

The test omitted the existing bounded StrictAllocationMeasurementScope, whose purpose and runtime mechanism are documented in the Nintendo allocation investigation. That prior investigation alone was not accepted as attribution for this workload.

New capture of the actual writer and credit probe¶

The opt-in diagnostic calls the same CanSubmitNativeCommand and synchronous TryWrite path, warming 256 calls and measuring 10,000 calls. The 32 fake native request slots are already populated; report history/copying are disabled. Diagnostic pressure creates a rooted 250,000-node reference graph and large arrays outside the measured calls. Each batch retains the exact zero assertion and stops on its first failure; no failure is averaged away or retried to pass.

The valid raw-boundary-03.trx/.txt capture reproduced a 3,496-byte jump at batch 17, with zero object-allocation callbacks and zero recorder overflow. The managed thread remained 4; collection counts remained 15,6,3. Native QPC window 333652282797..333652531848 contains GC-preparation SuspendStarted detail=7 at 333652325854, followed by suspension/resumption and collection-finished callbacks. The collection started before the window. See native log lines 235–252. Native Begin/End bracket a slightly wider interval than the managed counter pair; the recorder's derived gcDepth is not evidence.

Crucially, the same run's positive control copies the actual 398-byte report inside the same writer seam. It records 424 allocated counter bytes and one array object, with the allocation stack through ControlledNativeIo.TrySubmit and DualSenseBluetoothRealtimeWriter.TryWrite, and no overflow (lines 266–282). The positive control also proves that the unchanged zero assertion rejects that real report allocation. This corroborates allocation-context accounting repair, not a new writer object allocation, for the reproduced 3,496-byte observation.

Captures 01 and 02 are explicitly rejected for object-allocation attribution: their real-report positive control showed 424 bytes but no object callback. In the diagnostic PowerShell host, calling Environment.SetEnvironmentVariable(name, $null, 'Process') left a present-empty DS4W_ALLOCATION_NATIVE_SLOW_ONLY variable. The native selector tests presence, not its value. The corrected launcher uses Remove-Item Env:... and verifies absence before launch; capture 03 includes the required successful callback. All rejected evidence remains preserved.

Correction and validation¶

Only the test's warmed synchronous counter pair now uses the existing strict scope. Scope entry/exit failure still fails; the zero-byte limit remains zero. No production transport, allocation behavior, GC policy, JIT setting, timeout, or runtime configuration was changed. Raw opt-in tracing remains available.

isolated-boundary-04 exercised 256 pressure windows, each retaining the same 10,000-call workload and measuring exactly zero. Its positive report-copy control still observed 424 bytes and one object. All 259 native recorder windows (including initialization and positive-control envelopes) had zero overflow. Both tests passed. The subsequent ordinary rebuilt, no-profiler/no-diagnostic focused run passed 59/59 tests, including writer ownership, native backpressure and measurement scope lifecycle controls; no skips. Evidence: ordinary-focused.trx. The subsequent ordinary unfiltered rebuilt release run passed 5,539 tests, zero failures, 11 existing opt-in skips (5,550 total), with no profiler or diagnostic switches. Evidence: isolated_results/rc459-publication/rc459-full-final.trx, SHA-256 969701AE92C988D50B9E3FDF1BADD64DB1E4221DE2BD350B5B4CCF23BFE944C8. Independent exact-source release CI remains a separate gate.

All diagnostic artifacts are under isolated_results/issue81-native-credit-allocation/:

The existing v3 profiler binary was verified against 820E64A3AF8BFCEBD79F8F6C2D6FCF02F3D2C0EA4117CF6C8C7E4109EB74A7BD. Profiler settings applied only to new synthetic test processes; no controller apps, installations, drivers, or active hardware were touched by this investigation.