[evancohen]

Keep native hotword detection in the snowboy npm package so Sonus does not need to embed it. Companion Sonus PR: evancohen/sonus#109.

Changes:

• Prepare unpublished version 1.4.0-rc.0.
• Update node-pre-gyp/node-gyp/NAN dependencies, C++17 build settings, and TypeScript configuration.
• Compile JavaScript and declarations before npm packing; declare native build tools as runtime install dependencies so normal npm hoisting works.
• Preserve engine libraries and models. Adapt Sam's small wrapper fixes for exception handling, TypeScript indexing, and applyFrontend: false.
• Build the native addon from source rather than depending on historical S3 binary hosting. This requires Python, compiler/make and CBLAS/ATLAS at installation time.
• Explicitly package Node assets and omit the Android demo APK, iOS library, and unrelated examples. The candidate tarball is approximately 15 MB.

Validation: all Linux x64 package jobs passed on Node 14.21.3, 18, 20, and 22: https://github.com/evancohen/snowboy/actions/runs/34547751252 . Each job builds and packs the package, performs a normal consumer install from the tarball, initializes the native detector, and processes silent audio. Package contents were checked for the runtime JS, type declarations, header/source, engine library and models.

The tested tarball is available as the snowboy-1.4.0-rc.0 workflow artifact on that run. CI never publishes. npm release requires approval; prerelease should use a non-latest tag. macOS and Raspberry Pi hardware remain unverified, and real microphone/recognition-quality testing is outside this CI coverage.

Install verification is inline in the GitHub Actions workflow. No test directory or test scripts are included in the npm package.