Add macOS support (Core Audio backend + headless player), and make the circuit tests assertive

[ledogar]

This adds macOS support alongside the existing Windows app rather than changing it. LiveSPICE.sln and every WPF/VST project are untouched; the macOS-specific projects build from a separate LiveSPICE.Core.sln. The Windows CI job in this PR passes unchanged, so the risk to existing paths should be close to zero.

It's two separable commits, and the first is useful on its own even if you don't want the audio work.

1. Portable core build, and tests that actually assert (c9d8b65)

The core is already portable — Util, Audio, Circuit, ComputerAlgebra are netstandard2.0 with exactly one Windows dependency: the kernel32!MoveFileEx P/Invoke in Schematic.Save, which is now File.Move(overwrite: true) under net8.0 (the in-source comment anticipated this). Tests was pinned to net6.0-windows only because of the System.Drawing plotting sidecar, now #if PLOTTING-guarded and excluded off Windows.

The more valuable half is the test change. test previously passed if nothing threw — it never compared against Tests/Stats/*.csv. There's now a --check mode that compares every variable's mean/min/max/rms against those baselines and exits non-zero on mismatch, wired into CI on both platforms.

Making that work required recovering how the baselines were generated, which wasn't recorded: --sampleRate 44100, all other options default. At any other rate they don't reproduce — Simulation.TimeStep is Solution.TimeStep * oversample = 1/SampleRate regardless of oversample, so the sample rate alone sets the grid. I recovered it by fitting the input-node row that's common to all 49 files. Worth recording somewhere permanent; it's now in the CI invocation and a doc comment.

With that, on windows-latest all 49 circuits match the committed baselines with max deviation exactly 0.

2. Core Audio backend and headless player (a29ff31)

CoreAudio/ implements the Audio contract with AUHAL via P/Invoke to the system frameworks, mirroring how WaveAudio P/Invokes winmm — no third-party native dependency, nothing checked in. It follows the existing patterns: Driver per Asio/Driver.cs (one device each, per-device try/catch), Stream per Asio/Stream.cs (render callback, not WaveAudio's polling thread).

A few decisions worth reviewing:

• One AudioUnit drives everything — the output-bus render callback pulls input from bus 1 first, so there's a single clock and no ring buffer. This requires input and output on the same device, which Audio.Device.Open already assumes.
• Non-interleaved float32, so each AudioBuffer maps 1:1 onto Audio.Util.LEf32ToLEf64/LEf64ToLEf32. Those helpers are unit-stride only; interleaved would need a new de-interleave helper.
• Output is clamped before conversion — LEf64ToLEf32 doesn't clamp and neither does Core Audio.
• The Driver constructor no-ops off macOS, so this builds and is inert on Windows.

LiveSPICE.CLI/ is a headless player: list, tone, render, play, loopback. It doesn't reuse LiveSPICEVst/SimulationProcessor.cs — that code is genuinely portable, but reusing it meant either duplicating it or restructuring a Windows-only csproj I can't build. If you'd prefer it factored into a shared project, that's an easy follow-up and probably the right long-term shape.

Its SimulationHost does add two things the VST path lacks, which may be worth porting back: it forces the Linq expression tree to compile on the build thread (Simulation.Run compiles on first call, so today the VST takes a multi-millisecond stall inside the audio callback after every pot change), and it recovers from SimulationDiverged the way LiveSimulation does.

Verification

On macOS 26.5 / Apple Silicon:

• render of Passive 1stOrder Lowpass RC matches the analytic RC transfer function to 0.0002% in magnitude. The residual 0.29° phase lag is the half-sample delay from oversample output averaging (predicted 0.335°).
• Big Muff Pi renders 30.4% THD with odd harmonics dominant, as expected for symmetric diode clipping, from a 0.008% THD input.
• BlackHole loopback returns the played tone at full amplitude, rms exactly A/√2 — verifying enumeration, format negotiation, the render callback, both conversions and teardown.
• Live play produces exactly 282 × 512 frames in 3 s at 48 kHz, no dropouts, clean stop. Tone from real speakers confirmed by ear.
• Circuit suite passes 49/49 on macOS against the Windows-generated baselines, 45 of them to ≤1e-12.

That last point is the one I'd highlight: it means Expression.Compile() codegen, the Newton iteration and the Gauss-Jordan solver agree across OS, CPU architecture and SIMD width. Vector<double> is width-invariant here because the row operations are elementwise, so NEON at width 2 and AVX2 at width 4 give bit-identical results.

Known limitations

• Pro Co Rat needs a loose tolerance (1e-1, documented in Test.cs). It's bit-identical run-to-run on each platform, but hard clipping in feedback deterministically amplifies ulp-level libm differences into a ~4e-2 trajectory shift. Windows matches its baseline exactly; macOS doesn't. I didn't want to hide this behind a global tolerance.
• Input-only devices throw NotSupportedException pointing at Aggregate Devices. Duplex and output-only work. Mac built-in mic and speakers are separate Core Audio devices, so they need an aggregate.
• No GUI. This is headless only — the WPF UI remains Windows-only. PR Add Linux GUI, audio, and LV2 plugin port #272 looks like the better starting point there, and Avalonia runs natively on macOS.
• CoreAudio isn't in LiveSPICE.sln. Happy to add it if you want it built on Windows too (it's inert there).

Related: #262, #263, #202, #51.

🤖 Generated with Claude Code

[ledogar]

Apologies, PR created in error