0.5.0

Breaking

• In-process image generation is gone. Both the SD-Turbo engine
  (ONNX/WebGPU, running inside the plugin) and the FLUX.2 klein 4B engine
  (via a local mflux child process) have been removed entirely, along with
  the model catalog that drove them. The plugin now requires a separately
  running local A1111-compatible image server — Draw Things,
  AUTOMATIC1111,
  Forge, or
  SD.Next — configured via a new
  Server endpoint setting. See the rewritten README for setup
  instructions.

Added

• Thin-client architecture: the plugin talks to /sdapi/v1/txt2img,
  /sdapi/v1/options, and /sdapi/v1/progress over HTTP instead of running
  a model in-process. Model choice now lives entirely in the server app; the
  plugin sends generic parameters and shows the server's active model as a
  status hint.
• Real negative prompt and CFG (guidance scale) controls — previously
  impossible, since both old models (SD-Turbo, FLUX.2 klein 4B) were
  guidance-distilled and didn't support them meaningfully. Server-hosted
  models generally do.
• Generic size (7 curated aspect ratios), steps (1–50) and CFG
  (1–15) controls replace the old per-model catalog and its fixed
  SD-Turbo resolution / FLUX.2-only size dropdown.
• Test connection button in settings, checking reachability and
  reporting the server's active model.
• Legacy-cache cleanup: upgraders from pre-0.5 versions may still have
  ~2.5 GB of old SD-Turbo model weights cached in the browser's Cache API
  from the old in-process architecture. The plugin detects this once on
  load and shows a one-time notice; a new settings button deletes the
  cache in one click.

Changed

• Generation history is preserved across the upgrade: existing entries are
  loaded with default values for the new fields (negativePrompt: "",
  cfg: 7) rather than being discarded or breaking.
• Bundle size dropped from ~34 MB (including the bundled ONNX-runtime-web
  WASM runtime) to ~39.7 KB, since no inference runtime or model code ships
  with the plugin anymore.
• Confirmation dialogs now follow the shared UI convention (cancel on the
  left, standard button container).

Removed

• SD-Turbo (ONNX/WebGPU) and FLUX.2 klein 4B (mflux) engines, the model
  catalog, and all associated settings (model download, mflux path, FLUX
  weights location).

Fixed

• Draw Things' model name is now detected. Draw Things reports the active
  model under model in /sdapi/v1/options, while A1111, Forge and SD.Next
  use sd_model_checkpoint; only the latter was read. As a result the status
  hint stayed at "(chosen in server)" and — more importantly — every generated
  note got model: unknown in its frontmatter, the very field meant to make a
  recipe reproducible.