v0.14.0

• ABI break: Update the native llama.cpp runtime to v0.7.0 with matching
  bindings and Apple companions; v0.6.0 and older native overrides are
  incompatible, and old companion pins are rejected. Upgrade core and all
  companions together.
• Update the LiteRT-LM native runtime to v0.18.0-2, which reduces duplicated
  words at untimed ASR window boundaries with the default overlap; other boundary
  and recognition errors can remain.

• Aligned the default WebGPU bridge assets to v0.1.63, embedding llama.cpp
  v0.7.0 and qualified against native v0.7.0, preserving Web/native
  v0.7.0@620d2225cc436ab0b001e3200e590e3b286ef9ab parity. Web LiteRT-LM
  remains @litert-lm/core@0.18.0. Immutable Web asset manifest:
  65d146539eaba034d2c42bfc88a8ff376b937254ea089772401e07cade79f0b9.

• Behavior change: Raw responseFormat schemas with unsupported or shadowed
  assertions now throw LlamaUnsupportedException with a path before generation
  instead of being silently unconstrained; supported formats, bounds, allOf
  and tuples are now enforced.
• Bound tool-call argument nesting to report an inference error instead of
  overflowing the stack.
• Load progress callbacks now declare void Function(double progress)?;
  existing Function(double progress)? overrides still compile; prefer void.
• Hide stale performance counters after cancelling a queued llama.cpp generation.
• Free multimodal input buffers when native media initialization fails.
• Render user follow-ups after tool results on Gemma 2/3/3n and legacy Mistral
  templates.
• Expose liteRtLmQwen3ChatTemplate for stable thinking control on renamed
  native LiteRT-LM Qwen3 bundles; reject unsupported Web overrides explicitly.
• Qualify native EmbeddingGemma-2 text embeddings with 768-dimensional output.
• Document a fresh macOS SwiftPM bootstrap for generated deployment-target
  mismatches with Apple companions.