v0.8.24

• Align native leehack/llamadart-native@v0.4.1 on upstream
  b29c606e28a01b1bc8c1351026a0fa6e616bf6c4, with matching Dart bindings
  and Apple companion 0.0.19. This resolves the 0.8.23 grammar limitation:
  native {2000} repetitions are accepted again
  (llamadart-native#76).

• Aligned the default WebGPU bridge assets to v0.1.44 for matching
  Web/native v0.4.1@b29c606e28a01b1bc8c1351026a0fa6e616bf6c4 parity.
  Immutable Web asset manifest:
  8d61f453753ac7a7d839ac12318b70986a814748d86029993118c19454293aa9.

• Update native LiteRT-LM to v0.17.0-6 with Apple companion 0.0.11,
  Qwen3 tokenizer compatibility, corrected Linux loading, and explicit Linux
  and Windows GPU selection while retaining CPU defaults. The Windows x64
  runtime bundles dxil.dll and dxcompiler.dll, which D3D12 GPU engine
  creation requires
  (litert-lm-native#47).
  Web LiteRT-LM stays at @litert-lm/core@0.15.0.
• Preserve required iOS LiteRT-LM provider and Metal plugins, handle dependency
  ordering in companion libraries, and exclude metadata/import archives from
  runtime library inventories.
• Respect greedy sampling for zero-temperature native LiteRT-LM generation.
• Restore native Qwen3 chat text when thinking is disabled and preserve plain
  system instructions when seeding LiteRT-LM conversation history.
• Settle pending LiteRT-LM requests when a worker stops, close response ports,
  and report unverified native cleanup as an error.
• Preserve Unicode when detokenizing native GGUF tokens and suppress caller
  stop markers across chunk boundaries and speculative decoding.
• Restore native Qwen3-ASR file/encoded-byte transcription parity by keeping
  encoded audio out of string chat-template prompts.
• Render typed tool results as JSON text while preserving string results,
  validate Qwen XML argument types against schemas, and reject malformed or
  undeclared tool calls without exposing executable tool deltas.
• Prevent split MiniMax M3 thinking delimiters from leaking into reasoning.
• Discover Windows backend libraries in compiled CLI bundles and provide
  bounded diagnostics for unavailable native thinking-budget helpers.
• Fix fresh macOS Flutter dependency scanning while retaining Apple ABI and
  local-override guards.
• Add locked Gemma 4/Qwen3.5 validation profiles, explicit NPU coverage, and
  opt-in speech and voice-pipeline diagnostics.
• Preserve physical iOS speech-test failure-phase diagnostics and keep
  repository writer checks independent of generated website output.
• Retain open LiteRT-LM qualification gaps: macOS Qwen3.5 GPU reload latency
  (#521), Gemma exact-history
  behavior (#513), and
  Qwen3-0.6B arithmetic on Android CPU and iOS CPU/GPU
  (#509). The affected cases
  remain unqualified; these changes do not resolve the failures or establish
  their remaining owning layer.