v0.9.0

• Document generic JSON tool calling as an intentional fallback, including
  its prompt and model-reliability limits; runtime behavior is unchanged
  (#755).
• Load Qwen3.5-0.8B on the Web CPU (WebAssembly) backend at the default
  contextSize using smaller processing batches, while preserving
  full-context defaults for unknown models, including embedding models
  (#752).
• Report a failed Web model load on a page without cross-origin isolation as
  LlamaModelException with its real cause, not as a COOP/COEP
  worker-thread error; only a real worker-thread failure still names COOP/COEP
  (#753).
• Log a Dart warning when an explicit preferredBackend GPU module is not
  bundled and the model loads on CPU instead, as happens for cuda with the
  default Windows bundle; the native runtime docs now say when CUDA is bundled
  (#756).
• Fix the llamadart_server example exiting at startup on Windows; it stops
  on Ctrl+C there, and on SIGINT or SIGTERM elsewhere
  (#757).
• Accept MP3 and FLAC bytes, as well as WAV, for Qwen3-ASR speech to text on
  Web (#723).
• Apply presencePenalty, minP and thinkingBudget, and runtime LoRA
  adapters (setLora, removeLora, clearLoras), on WebGPU with bridge
  assets whose capability probes report them; other assets still reject them
  (#722).
• Run speculative decoding on WebGPU with bridge assets whose capability
  probe reports the strategy, and report each runtime's strategies in
  backendGenerationCapabilities.speculativeDecodingStrategies; other assets
  still reject it (#722).
• Reject a non-zero GenerationParams.minP on WebGPU when the bridge lacks
  Min-P, with LlamaUnsupportedException instead of ignoring it, and ignore a
  stop sequence equal to a preservedTokens entry there, as native llama.cpp
  does (#661).
• Add LlamaEngine.backendGenerationCapabilities, which reports whether the
  loaded runtime applies presencePenalty, minP and thinkingBudget; the
  example chat app uses it to send Min-P and enable its slider only where
  supported (#661).
• Return a DeepSeek V3 forced-open thought that never closes as reasoning, as
  llama.cpp does with the DeepSeek V3.1 template, instead of as content with
  its tool calls (#743).
• Keep escaped \n and \r in Qwen3-Coder XML reasoning, as llama.cpp does
  with the Qwen3.5 template; before, they became line breaks unless a tool
  call ended the thought
  (#743).
• Stream content and reasoning with the whitespace the non-streamed parse
  keeps, with or without tools, so streamed answers and ChatSession history
  no longer start with the blank lines after </think>. Only whitespace at
  either end, and text that may be a tag or tool-call opening, waits for more
  output, so reasoning still streams token by token. Without tools, Hermes,
  DeepSeek R1, Qwen3-Coder XML and the other formats the template engine
  guide lists also drop a start tag repeated at the start of a forced-open
  thought, as the parse does. The guide lists the exceptions
  (#754).
• Stream content that equals the non-streamed parse for Qwen3-Coder XML,
  Mistral Nemo and 15 more tool-call formats, and for output parsed with a PEG
  parser, so text before a tool call no longer carries the tool-call envelope
  into streamed content or ChatSession history. Content and reasoning are
  trimmed as for Hermes, and text after a call arrives at the end of the
  stream. See the template engine guide for the formats and exceptions
  (#732).
• Stream Hermes-format content that equals the non-streamed parse, so text
  before a tool call no longer carries the <tool_call> envelope into
  streamed content or ChatSession history; only a possible envelope opening
  and trailing whitespace wait for more output. With tools, streamed content
  is now trimmed as the parse trims it, and text the parse keeps after a tool
  call, including a malformed envelope, arrives at the end of the stream.
  Streamed reasoning, and so ChatSession thinking, is trimmed per thought as
  the parse trims it. The exception is a forced-open thought that never
  closes and starts with whitespace: the parse keeps it untrimmed, but it
  streams without its leading and trailing whitespace, so
  " \n Hello there. \n\n" streams as "Hello there.". Before, it streamed
  as the parse gives it, except for some thoughts containing a backslash,
  depending on chunking. After a forced-open
  thought, text after a tool call arrives at the end
  (#701).
• Throw LlamaModelException when a WebGPU model load fails with a bridge
  error that has no specific mapping, and LlamaInferenceException or
  LlamaStateException for such Web embedding, next-token scoring and state
  errors, with URL credentials and signed query values redacted from the
  details and the load-failure console log
  (#704).
• Keep URL credentials, signed query values and fragments out of
  LlamaEngine model and projector load errors and logs and the model field
  of completion chunks for every URL form, including scheme-relative
  //user:pass@host/... URLs and relative paths with a query, and out of
  native model download errors. A projector load error that is not a
  LlamaException now throws LlamaModelException. The details of a
  model or projector load failure is now a {type, message} map instead of
  the original error, and a native download that fails with a network error
  carries the error text as a String in details
  (#704).
• Keep the text after a U+0000 in native llama.cpp tokenization, embeddings
  and generation prompts instead of dropping it
  (#608).
• Make DecisionEngine.load throw LlamaStateException when another model is
  loaded while it runs, even under the same backend handle
  (#626).
• Bound speech validation pack memory by a footprint counter instead of the
  resident set: phys_footprint on macOS and iOS, RssAnon plus RssShmem
  plus VmSwap on Linux and Android, read after malloc_trim(0) where the C
  library provides it (glibc, not Android), and PrivateUsage plus
  SharedCommitUsage on Windows (PrivateUsage alone on builds without it).
  Evicting file-backed pages, such as the
  mmapped weights, compressing memory under pressure, or glibc keeping freed
  memory across reloads no longer fails peak_memory_bound without memory
  growth, and each report names its counter
  (#633,
  #762).
• Fail speech validation leak_slope_bound when the least-squares footprint
  slope over cleanup cycles 1-8 exceeds 7 MiB per cycle; it failed only when
  every cycle grew by more than 7 MiB, and passed leaks of 16 MiB per reload
  (#762).
• Detect chat template capabilities with llama.cpp's probes, and give
  templates that read only typed content text parts, as llama.cpp does:
  SmolVLM prompts keep the message text, Ministral 3 renders an image
  followed by a reasoning-only turn as llama-server does, TranslateGemma 2B
  keeps the text next to an image, and Kimi-K2 tool results after an image
  are plain text
  (#720).
• Use the LFM2 format for LFM2.5 templates that list tools without
  <|tool_list_start|>, as llama.cpp does, so LFM2.5-1.2B-Instruct and
  LFM2.5-1.2B-Thinking tool prompts drop the stray "Respond in JSON format"
  instruction (#716).
• Report per-request usage on WebGPU with the newly pinned bridge assets
  v0.1.54, on the final create chunk and to observers
  (#696).
• Give assistant turns that hold only tool calls or only reasoning empty
  content instead of null in chat templates, as llama.cpp does: QwQ-32B
  renders them instead of throwing, and LFM2 and Devstral prompts drop a stray
  null or <function text>
  (#715).
• Pass Map and List tool results as compact JSON text to LFM2, gpt-oss, Solar
  Open, Ministral, DeepSeek V3 and TranslateGemma templates too, instead of
  Python-style or spaced text
  (#717).
• Pass earlier tool-call arguments as JSON objects to templates that read them
  as objects, as llama.cpp does, so Qwen3, Ministral, Devstral, gpt-oss and
  similar prompts format them with the template's own JSON spacing
  (#702).
• Require dinja 1.2.0, so more chat prompts match llama.cpp: tojson
  output such as tool declarations uses llama.cpp's spacing, number format
  and non-ASCII text; Qwen3-Coder, GLM-4.6, GLM-4.7-Flash, MiniMax-M2,
  Nemotron-3-Nano, Command R7B, Cohere2 MoE and Ling 3.0 prompts lose stray
  indentation; Functionary v3.1 adds no tool instructions without tools;
  Granite 3.3 spells out the month in its date; Hunyuan Hy3 keeps the system
  prompt first instead of merging it into the user turn; and Bielik 11B v3
  tool-call turns without text render instead of throwing.
• Count generated tokens with an empty text piece in llama.cpp
  getPerformanceContext() evalTokens and sampleCount without speculative
  decoding, as the speculative path already did
  (#706).
• Report per-request token usage and timings on the final create chunk as
  LlamaCompletionChunk.usage on native llama.cpp
  (#696).
• Add LlamaEngine(observers: ...), which reports chat and text completions,
  embeddings and model loads, with their usage, to tracing and metrics code
  (#696).

• Updated the default llama.cpp native runtime pin to
  leehack/llamadart-native@v0.5.0 (llama.cpp v0.5.0) with Apple companion
  0.0.20, regenerated matching Dart FFI bindings, refreshed the llamadart_llama_cpp_flutter Apple SwiftPM
  checksum, and aligned current README/website native override docs.

• Add LlamaEngine.scoreNextToken(...) for next-token log-probabilities on
  native llama.cpp and WebGPU bridge assets v0.1.52+, matching llama-server
  n_probs; check
  supportsNextTokenScoring first
  (#694).

• Add example/laya_command_bar, a Flutter text field that reshapes into a
  reminder, message, calculation or other command as you type, read by a
  Laya decision model, by EmbeddingGemma and labelled examples, or by small
  LLMs' next-token scores, including the decision model decider-2b.

• Throw LlamaModelException when native llama.cpp cannot find or load a
  multimodal projector, and LlamaUnsupportedException when the runtime lacks
  the mtmd functions; LlamaEngine.supportsAudio also throws the latter.
  Speech-to-text capabilities now say when no projector is loaded, using the
  new LlamaEngine.hasMultimodalProjector
  (#325).

• Honour LlamaEngine.cancelGeneration() issued right after listening to a
  create, generate or ChatSession.create stream, before it reaches the
  backend, instead of running the whole generation
  (#602).

• Cancel an active text-to-speech synthesis on LlamaEngine.unloadModel() and
  dispose() instead of waiting for it to finish
  (#628).

• Cancel an active Qwen3-ASR transcription on LlamaEngine.unloadModel() and
  dispose() instead of completing it with the transcript cut at the unload
  (#670).

• Send LiteRT-LM tool calls and tool results in the runtime's own message
  format, so Gemma 4 reads tool output and Qwen3 tool histories no longer
  fail
  (#681).

• Start a native llama.cpp generation requested right after a cancel once the
  cancelled run stops, instead of failing with generation is already in progress. An overlap with a running generation that was not cancelled now
  throws LlamaStateException
  (#655).

• Render Qwen3 prompts as llama.cpp does: an earlier assistant tool-call
  turn without reasoning no longer gets an empty <think> block
  (#691).

• Require dinja 1.1.0. Its Jinja string comparison makes three more chat
  templates render as llama.cpp does: MiniMax-M1 adds no empty
  system block for an empty or whitespace-only system message; NVIDIA
  Nemotron Nano v2 drops the blank line before a tool call, the blank lines
  before its tool instructions when tools come with an empty or
  whitespace-only system message, and an empty final assistant turn without
  a generation prompt; and Functionary v3.2 tool declarations drop stray
  // Format=<|NONE|> lines and spell out nested object parameters
  (#351).

• Cancel a generation's backend run as soon as its stream subscription is
  cancelled, instead of at its next token, which during prompt evaluation
  meant after the whole prompt. A native llama.cpp generation requested right
  after such a cancel now waits for it instead of throwing
  LlamaStateException, and native llama.cpp sees a cancel between text
  prompt micro-batches (ModelParams.microBatchSize, 512 tokens by default)
  or, with speculative decoding, between batches (ModelParams.batchSize)
  (#663,
  #660).

• Render every result of a tool message holding several
  LlamaToolResultContent parts, as one tool message per result like
  llama.cpp, instead of only the first; LlamaChatMessage.toJson lists them
  all (#683).

• Render Gemma 4 tool calls and tool results as llama.cpp does, so Gemma 4
  GGUF models can read tool output
  (#669).

• Stop a Qwen3-TTS audio decode at its next chunk boundary when native
  text-to-speech is cancelled, instead of finishing the native step in
  progress first. This needs llamadart-native v0.4.1-1 or later; older
  runtimes keep the previous behaviour
  (llamadart-native#86,
  #322).

• Add an experimental DecisionEngine for Laya-style decision models (a
  ModernBERT encoder GGUF plus a safetensors head) on native llama.cpp, with
  typed ChoiceKey, ScoreKey and NoulKey questions
  (#604).

• Add example/basic_app/bin/llamadart_decision_example.dart, a console demo
  that triages a support ticket with DecisionEngine
  (#604).

• Add example/laya_tetris, a Flutter app in which a Laya decision model
  plays real-time Tetris through DecisionEngine
  (#604).

• Run example/laya_tetris on Web through the WebGPU bridge, with a live demo
  at https://leehack-flutter-laya-tetris.static.hf.space.

• Add a notebook in example/laya_tetris/training/ that fine-tunes a Laya
  decision head for the Tetris example and exports it for DecisionEngine
  (#604).

• Run DecisionEngine on WebGPU through the bridge decision API
  (apiVersion 1), which bridge assets v0.1.47+ include
  (#604).

• Stop native image and audio requests from seeding the repeat penalty with
  leftover memory, which made output depend on the previous request
  (#603).

• After a failed native prompt decode, the next reusePromptPrefix request no
  longer runs on the wrong KV cache or keeps failing
  (#601).

• Reject embed() and embedBatch() on rank-pooled reranker GGUFs with
  LlamaUnsupportedException on native, instead of returning memory read
  past llama.cpp's classifier-score buffer
  (#583).

• Throw LlamaInferenceException from native embed() and embedBatch()
  when input to an encoder-only model or a model without a KV cache (such as
  BERT-family and ModernBERT GGUFs) does not fit one microBatchSize pass,
  instead of aborting the process or embedding only the last chunk
  (#607).

• Aligned the default WebGPU bridge assets to v0.1.54 for the decision API,
  next-token scoring, presence penalty, Min-P, thinking budgets, runtime LoRA
  adapters, speculative decoding and the Web runtime fixes below; the bridge
  also adds its supportsCompletionUsage flag, which llamadart does not use yet
  (#729). The assets 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.

• On Web, an invalid GBNF grammar now fails generation with a
  LlamaInferenceException whose details contain (invalid grammar), and the
  loaded model stays usable, instead of aborting the WebGPU bridge runtime
  (llama-web-bridge#125).
• On Web, when a bridge worker fails and the main-thread reload fails too, or
  a replacement worker cannot start, the bridge now forgets the model instead
  of being left broken with a TypeError. Later calls fail with
  No model loaded. Call loadModelFromUrl first.; call unloadModel(), then
  loadModel() and any projector again to recover
  (llama-web-bridge#123,
  llama-web-bridge#127).
• On Web, a replacement bridge worker reloads the current model before its next
  request
  (llama-web-bridge#126).
• On Web, grammar-constrained generation no longer aborts the bridge runtime
  when top-k or top-p keeps only tokens the grammar rejects; it resamples as
  llama.cpp does, and fails with Grammar rejected every candidate token only
  when the grammar cannot continue
  (llama-web-bridge#118).
• On Web, an ordinary error from a healthy bridge worker, such as a prompt that
  overflows the context or empty embedding input, is now rethrown with the
  worker kept, instead of moving the session to the main thread for good and
  re-running the request
  (llama-web-bridge#119,
  llama-web-bridge#120).
• Extend the GGUF speech-to-text validation pack with four synthetic edge
  fixtures built in-process, so no extra audio is stored: generated digital
  silence, plus a truncated RIFF, a stereo 44.1 kHz re-encode and a 33-second
  concatenation, the last three derived from the locked jfk.wav. Every GGUF
  STT pack run executes the four checks and each one gates functional_pass;
  the LiteRT-ASR and TTS packs pass no edge fixtures
  (#325).
• Raise every stt, tts and litert-asr speech validation pack run from 8
  to 15 lifecycle checks: an immediate cancel, three
  cancel/dispose/load/generate cycles, and bounds that fail the run when a
  cancellation takes over 500 ms to end its task or the peak resident set
  exceeds 1.10x the one sampled after the first generation
  (#594).
• Add cross-platform validation cases for a cancel issued right after
  listening, a generation requested right after a cancel, an overlapping
  generation, an invalid GBNF grammar and ToolChoice.auto on a prompt that
  needs no tool, and run the tool cases on the GGUF chat profiles
  (#602,
  #655,
  #654).
• Add decision-gguf-{cpu,metal,vulkan,cuda,webgpu} validation profiles that
  check DecisionEngine token ids, raw logits and answers against the Laya
  0.3.5 reference, plus batching, reload and typed rejections, on desktop,
  mobile, Web WebGPU and GCE CUDA
  (#604).
• Add a Web-only chat-gguf-webgpu validation profile, hash Web validation
  models while they stream so GGUFs over 2 GiB pass preparation, list every
  bundled profile in the validation app, and verify iOS GGUF GPU placement
  from the XCTest console log.
• Run eight cleanup cycles instead of three in every speech validation pack,
  and fail a run whose resident set grows by more than 7 MiB in each of seven
  warm cycles. The 1.10x peak ratio no longer applies on Linux CUDA, where reload
  overhead that levels off failed it without a leak
  (#686).
• Add GGUF speech validation pack checks: tts unloads and disposes the engine
  during a synthesis, cancels one during its audio decode and bounds the
  resident set those checks add, and stt must fail with
  LlamaSpeechTranscriptTruncatedException at maxOutputTokens and at the
  context size. stt runs now execute 28 checks and tts runs 25
  (#628,
  #636,
  #322).
• Force greedy topK: 1 for zero-temperature LiteRT-LM Web generation, matching
  the native clamp
  (#548).
• Log Model … loaded from …; native engine creation is deferred until the first generation or tokenizer call instead of loaded successfully when
  the native LiteRT-LM backend finishes loadModel, since it creates the
  engine lazily
  (#569).
• Require dxcompiler.dll and dxil.dll in the Windows x64 LiteRT-LM runtime
  cache and desktop validation bundle checks, matching the hook's v0.17.0-6
  inventory. The runtime does not preload them: Dawn's D3D12 backend loads the
  pair at GPU engine creation
  (#570).
• Document that Linux llama.cpp loads need the OpenMP runtime
  (libgomp.so.1; libgomp1 on Ubuntu/Debian, libgomp on Fedora and Arch)
  and that Linux LiteRT-LM GPU needs a hardware Vulkan ICD: with only Mesa
  llvmpipe the runtime segfaults after model load instead of failing cleanly
  (llamadart-native#82,
  #572).
• Record one startup diagnostic when every candidate of a native backend
  module family fails to load, or when a ggml/wrapper symbol is missing from
  both the primary FFI asset and every fallback library. Candidates are named
  by asset URI or file name only and loader errors are classified, never
  quoted, so no directory or loader search path reaches the diagnostic
  (#416).
• Forward llama.cpp and LiteRT-LM worker-isolate log records to the
  LlamaEngine.configureLogging handler. A worker takes the Dart logger level
  when it starts and LlamaEngine.setDartLogLevel/setLogLevel update a
  running worker; the default none sends nothing and debug records are
  capped at 1000 per worker. The LiteRT-LM program-cache pruning warnings are
  now ordinary warn records gated by that level instead of the native log
  level. Adds LlamaLogger.level and the BackendDartLogLevel capability
  (#567).
• Replace the token in Bearer <token> and the value in token=, key=,
  secret=, password=, api_key= and apikey=<value> outside HTTP URLs in
  native startup diagnostics with <redacted-secret>; URL and
  control-character handling is unchanged
  (#551).
• Move native release pins (llama.cpp tag, LiteRT-LM tag and per-bundle
  checksums) from hook/build.dart into
  lib/src/hook/native_release_pins.dart, the only file the pin sync now
  rewrites.
• Add ModelParams.liteRtLmCacheDir to choose the native LiteRT-LM runtime
  cache directory and opt-in ModelParams.liteRtLmMaxProgramCacheBytes, which
  deletes *_mldrift_program_cache.bin files above the cap before each engine
  create and logs a warning per deleted file. Defaults are unchanged: the same
  per-platform directory and no pruning
  (#552).
• Force greedy topK: 1 for zero-temperature
  LiteRtLmRuntimeClient.createConversation calls, which returned incoherent
  text on the LiteRT WebGPU sampler with the default top-k.
• Keep root-cause native startup diagnostics when the buffer or the rendered
  startupDiagnostics=[...] suffix overflows: teardown entries, now prefixed
  teardown: , are dropped first, duplicates are recorded once, entries are
  capped at 2048 characters, and each omitted run renders as ...
  (#415).
• Skip the Windows altered-search-path preload for wrapper library candidates
  whose absolute path does not exist, so lazy wrapper API lookups no longer
  record a Failed to preload Windows backend module startup diagnostic per
  missing candidate
  (#550).
• Accept promptTemplate on the non-native
  LiteRtLmRuntimeClient.createConversation placeholder, so callers passing
  it compile for Web as they do on native
  (#549).
• Cache TemplateCaps.detect results in a per-isolate LRU keyed by exact
  template source and bounded at 16 entries, so repeated chat-template renders
  skip both Jinja parses and all four capability probes. Detections in which
  any analysis step failed are not cached and keep logging on every call
  (#448).
• Detect supportsTools and supportsToolCalls for chat templates that
  reject two tool calls in one assistant message (Llama 3.2) or a user turn
  directly after a tool call (Ministral 3). The tools capability probe now
  renders a single tool call, and a separate parallel probe clears only
  supportsParallelToolCalls when it throws
  (#557).
• Limit the Ministral tool-call grammar to a single [TOOL_CALLS] block unless
  parallel tool calls are enabled; it previously always allowed repeats while
  the parser kept only the first call
  (#559).
• Limit the Nemotron v3 tool-call grammar (Qwen3-Coder XML format) to a single
  tool call unless parallel tool calls are enabled; it previously always
  allowed repeats while the parser kept only one call
  (#562).
• Pin the WebGPU model-load retry ladder with browser tests for the ladder
  advance, the wasm64 BigInt restart on wasm32, the restart without the remote
  fetch backend, and forced remote-fetch chunk halving stopping on both its
  ten-restart cap and its 4 KiB minimum chunk, then collapse the duplicated
  attempt thread-count switch into one helper
  (#361).
• Apply the JSON Schema pattern keyword when generating GBNF, for anchored
  patterns built from literals, positive character classes, (...) groups
  nested at most 32 deep, grouped alternation and */+/?/{m,n}
  repetition; any other pattern, including deeper nesting, falls back to the
  rule the schema would have produced without it, so minLength and
  maxLength still apply there. An applied pattern replaces
  minLength/maxLength as it does in llama.cpp, so a schema carrying both
  pattern and maxLength is no longer length-bounded. A schema carrying
  pattern but no explicit type now yields a string rule instead of
  throwing Unrecognized schema. Mistral Nemo and Magistral tool-call ids are
  now grammar-constrained to exactly nine alphanumerics
  (#582).
• Stop a cancelled llama.cpp image or audio prompt at the next prompt chunk,
  or between a media chunk's encode and its embedding decode, instead of after
  the whole prompt is ingested. The native call already running still
  finishes
  (#599).
• SpeechToTextEngine now fails a native Qwen3-ASR transcript that reaches
  the context size or maxOutputTokens with
  LlamaSpeechTranscriptTruncatedException instead of completing with
  truncated text, and create() reports finishReason: 'length' when native
  llama.cpp stops at either limit. The chat app caps Qwen3-ASR recordings at
  the validated 30 seconds
  (#636).
• Stop ToolChoice.auto on WebGPU from forcing a tool call: it now skips
  the lazy tool-call grammar and parses tool calls best-effort. WebGPU
  rejects GenerationParams.grammarLazy and a non-root grammarRoot with
  LlamaUnsupportedException; backends report this through the new
  BackendLazyGrammarSupport
  (#654).
• Name the CUDA 12 runtime libraries (libcudart.so.12, libcublas.so.12)
  that the Linux cuda backend needs on the default loader path; llamadart
  does not ship them, and validation bundles refuse LD_LIBRARY_PATH
  (#587).
• Remote validation runs resolve packages/llamadart_validation before
  building the report, instead of reporting FAILED with a null error on a
  fresh checkout. A failed report step is now the run's error, with its exit
  code and a redacted stderr tail
  (#688).
• Select the devices of an explicit GpuBackend.metal or GpuBackend.hip
  on llama.cpp: they looked up ggml registries named Metal and HIP, but
  ggml names them MTL and ROCm, so loading fell back to automatic device
  selection. A HIP load on a ROCm build now reports its backend as HIP
  instead of CPU
  (#611).
• Report a WebGPU model load that fails with error 138 as the documented
  cross-origin isolation (COOP/COEP) UnsupportedError, as
  thread constructor failed already was, instead of rethrowing the raw
  bridge error
  (#598).
• Throw LlamaModelException when WebGPU cannot fetch or load a multimodal
  projector, instead of the raw JavaScript error. Its details drop these
  parts of the projector URL the app passed, as written, JSON-escaped,
  percent-encoded or percent-decoded: the userinfo and password, as whole
  tokens of any length; the ?query and #fragment, where they directly
  follow a non-space character; the query and each &-separated part that
  contain =, as whole tokens; and bare query values and the fragment of 10
  or more characters, as whole tokens. A whole token has no ASCII letter or
  digit directly before or after it. Shorter bare values printed on their
  own, such as the 1 of ?v=1, stay. Other URLs in the details lose
  userinfo, query and fragment on a best-effort basis
  (#642).
• Leave no envelope text in the parsed content when Qwen2.5 wraps a Hermes
  tool call in double braces (<tool_call>{{"name": ...}}</tool_call>, with
  any number of extra closing braces) without a grammar. Calls are extracted as
  before; a double-brace call with other malformed envelope text keeps that
  text. This deliberately differs from upstream llama.cpp, which rejects that
  output and extracts no call
  (#662).