AdaIN convolution runspace scheduling audit

A PowerShell ForEach-Object -Parallel prototype split one bounded AdaIN convolution by output channel. Each worker retained the scalar accumulation order and returned a private output segment; an analytic gate was bit-identical to the serial operator. The mechanism was traced to PowerShell 149ab5cd6cad34869177f86ef9a3da8414f85dc6, src/System.Management.Automation/engine/InternalCommands.cs (ForEachObjectCommand and using-value admission) and src/System.Management.Automation/engine/hostifaces/PSTask.cs (PSTaskPool).

On this laptop, a 64-channel, 121-frame, 11-tap same-input comparison with four runspace workers measured 5,848.8 ms serial and 19,386.2 ms parallel. That is 0.30x serial throughput, not a speedup. The prototype was not promoted into the reference or product path. This single fixture does not rule out other parallel schedules or a PowerShell-authored emitted backend; it does rule out claiming this channel-per-task implementation accelerates the current correctness gate.

The attempted complete stock-layer phoneme-to-PCM scalar gate was stopped after prolonged CPU activity without a stage result or output PCM. It is not a failed numerical comparison and not a passed full-forward gate.