[kokoro-js] Bump @huggingface/transformers to ^4.2.0

[shreyaskarnik]

Summary

Bumps the declared @huggingface/transformers range in kokoro.js/package.json from ^3.5.1 to ^4.2.0.

Motivation

When a downstream app depends on both kokoro-js and @huggingface/transformers directly, capping kokoro-js's range at 3.x forces npm to install two copies of @huggingface/transformers whenever the app pins ^4.x. Each copy ships its own onnxruntime-node, and loading two ORT native modules into the same Node process segfaults — ORT is not designed to be initialized twice in one process; the duplicate global state corrupts the runtime.

We hit this in production. The current workaround consumers have to use:

"overrides": {
  "kokoro-js": {
    "@huggingface/transformers": "$@huggingface/transformers"
  }
}

Bumping the declared range removes the need for any consumer to write that override.

API surface verified

kokoro-js consumes { env, StyleTextToSpeech2Model, AutoTokenizer, Tensor, RawAudio } and writes through the internal env.backends.onnx.wasm.wasmPaths path used by the env shim. All exports and the internal path are present in v4.2.0.

Verification

npm test — 276/276 passing on v4.2.0 (Node v20.20.2, darwin-arm64).

A separate fp32-CPU benchmark (cold-load, per-utterance latency / RTF, streaming TTFA, peak RSS, on onnx-community/Kokoro-82M-v1.0-ONNX, 1 warmup + 3 measured runs each):

Metric	v3.5.1	v4.2.0	Δ
Cold load	3475 ms	3455 ms	~flat
Short (37 ch) — wall / RTF	447 ms / 5.99x	442 ms / 6.05x	~flat
Medium (164 ch) — wall / RTF	1721 ms / 6.03x	1731 ms / 5.99x	~flat
Long (516 ch) — wall / RTF	4621 ms / 5.87x	4609 ms / 5.89x	~flat
Streaming TTFA	464 ms	480 ms	+3% (1 sample)
Peak RSS	1832 MB	1948 MB	+6%

Compute is flat (kokoro's hot path is one ONNX inference call dispatched by onnxruntime-node, not by transformers.js itself; v4's headline 4x BERT-MHA win doesn't apply to StyleTextToSpeech2). RSS is ~6% higher, expected from v4's larger feature set (modular models, WebGPU-in-Node, ModelRegistry). Audio output length is byte-identical.

Test plan

• cd kokoro.js && npm install && npm test — expect 276/276
• Run the README's basic and streaming examples on Node — audio.save produces a valid wav

[shreyaskarnik]

cc @xenova