v0.15.4

• Keep GPT-OSS reasoning and tool frames out of answer text, including native header continuations.
• Keep GPT-OSS automatic tool choice optional for ordinary text replies.
• Preserve Gemma 4 answer text after tool calls across split token delivery.

• Aligned the default WebGPU bridge assets to v0.1.66 (unchanged from 0.15.3).
  Runtime pins are unchanged: native v0.7.0-1 uses upstream
  v0.7.0@620d2225cc436ab0b001e3200e590e3b286ef9ab; Web LiteRT-LM stays
  @litert-lm/core@0.18.0.
  Manifest SHA-256:
  6a82e16cd850c79ebb8d5f5613dccc9adc42030bf5788bc3cffcda21f19f2f1d.