v0.11.1

• Give a chat template that reads content part lists but rejects image and
  audio parts the media marker in its text instead of failing in the
  template, and render a video part in a chatTemplate prompt as the media
  marker on any template when the messages hold no image or audio part
  (#959). Beside an image
  or audio part, a template that is given typed content parts still receives
  the video part (#963).
• Fix a possible llama.cpp crash when an image or audio part of a multimodal
  request fails to load while later parts remain, and native memory leaked
  when a backend is disposed with a context still open or an embedding fails
  because the model reports no embedding size
  (#947,
  #610).
• Fix Qwen 3 chat and tool-result history with LiteRT-LM runtimes that
  normalize message content to arrays.
• Work around wrong llama.cpp Vulkan output on some Android GPUs for prompts
  over 32 tokens (seen on a Pixel 9 Pro, Mali-G715): an Android Vulkan
  context now decodes a text prompt in micro-batches of at most 8 tokens
  unless ModelParams.microBatchSize is set
  (#948). Only text-prompt
  decoding is covered: prompts with image or audio input, embeddings, decision
  models, text-to-speech and speculative-decoding verification batches are
  not, so speculative decoding can still produce wrong output on an affected
  GPU.
• Document that a Flutter macOS app that quits without disposing its engines
  still exits cleanly, and why disposing them in onExitRequested is still
  recommended (#826).

• Aligned the default WebGPU bridge assets to v0.1.54, unchanged from 0.11.0:
  they embed llama.cpp v0.5.0, are qualified against native v0.5.0, and keep
  Web/native llama.cpp v0.5.0@7fe450e19305b828c199d602c23a8337aaa1f03b parity
  and Web @litert-lm/core@0.15.0. Immutable Web asset manifest:
  8a9f83c15035eeb034a6563e6f753382d7d7f9be81503ef76902138da7841176.