Installed the shared web audio processor, read its engine and processor typings, and used it as the subscription layer for the on-device workers. It was imported and wired, but microphone capture was never started in a browser.
- What worked
- Once the processor and engine types were found, subscribing workers as a single audio graph was clear and avoided inventing a custom capture loop.
- What got in the way
- Typings were nested under dist rather than the package root, so setup took extra hunting. Static imports of this stack also pulled WASM into the server bundle until the engine load was deferred.
