Implement the VST3 factory so the plugin actually loads

[godofecht]

Includes the docs/examples work it is based on, plus the factory implementation.

The bug

danzig's example plugin could not be loaded by any host.

fn factory_countClasses(_: *anyopaque) callconv(.c) i32 { return 1; }
fn factory_getClassInfo(_: *anyopaque, index: i32, _: *anyopaque) callconv(.c) i32 {
    _ = index;
    return 0;          // 0 is kResultOk: claims success, writes nothing
}

A host is told one class exists, asks for it, and gets an uninitialized struct back with a success code.

Root cause was two things

The stub factory was the visible bug. The bundle also exported no bundleEntry. macOS hosts load a .vst3 via CFBundle and call bundleEntry first; without it the module is discarded before GetPluginFactory is ever looked up. Fixing only the factory would still have produced "Num plugins found: 0".

Before / after

before:  Num plugins found: 0
         !!! Test 1 failed: No types found.
         FAILED!!  1 test failed, out of a total of 1

after:   Num plugins found: 1
         Testing plugin: VST3-Danzig Gain-43e2804f-c026063c
         Superelectric: Danzig Gain v0.1.0
         SUCCESS

Passes at --strictness-level 5 and 10, in Debug and ReleaseFast. No strictness was lowered.

Also corrected in src/vst3.zig

Several constants were wrong: kResultFalse was 0 (it is 1), kBypass was 1 << 5 (kIsBypass is 1 << 16), AudioBusBuffers.channelBuffers32 had = null on a non-optional pointer, and IComponent.getBusCount was missing its MediaType argument. Comptime assertions now pin the struct sizes a host reads.

The test that was lying

The harness asserted getClassInfo(0) == 0, which passes against the stub because the stub returns 0. Every buffer is now zeroed before the call and the content asserted. Reverting the factory to its stub bodies fails 9 checks:

  FAIL  getClassInfo(0) writes the "Audio Module Class" category
  FAIL  getClassInfo(0) writes a non-empty class name
  FAIL  createInstance writes a non-null object
  ...

getClassInfo(0) returns kResultOk still passes against the stub, which is exactly the point: the result code alone proves nothing.

The harness went from 12 to 49 checks and now drives the object as a host does, through initialize, bus info, parameter info, setupProcessing, and real audio: DC at 0 dB comes out 1.0, at +6 dB 1.995, bypassed exactly 1.0.

Not implemented

IConnectionPoint, IUnitInfo, IMidiMapping, IPluginFactory3, 64-bit processing, a custom editor, programs/presets. All refuse cleanly with kNotImplemented rather than a false kResultOk. The plugin does not call restartComponent(kParamValuesChanged) after a state load; pluginval passes without it, but a host caching parameter values aggressively could show stale numbers after loading a session.

Verification

35/35 unit tests and 49/49 integration checks on both 0.14.1 and 0.15.2 from a clean cache. lipo -info reports x86_64 arm64.

🤖 Generated with Claude Code