v0.7.2

What's Changed

Image generation and local image editing, on top of the text chat. Needs a
Pictor ≥ 0.3.0 server for the image features; text-LLM servers are
unaffected. Existing settings and chat history are read as-is.

Added

• Image generation with a Pictor server. Pick an image model in the
  model list (they are marked with an icon) and the settings panel swaps the
  text-LLM sections for an Image section: mode (auto / generate / edit /
  chat), aspect ratio, size, steps, seed (with a dice for a random one),
  guidance, negative prompt, transparent background — each greyed out when
  the backend cannot do it, each overridable per conversation with a reset.
  The bubble shows the server's progress bar ("generating 12/40") instead of
  the typing dots; the result is stored like any attachment and displayed
  inline. A generated image is re-sent on the next turn, so "now make it
  blue" edits it.
• Images go into messages. Attach with the paperclip (native file
  dialog), drop files anywhere on the chat panel, or Cmd/Ctrl+V an image
  from the clipboard. PNG, JPEG, WebP and GIF are recognised by content, not
  extension; anything larger than 2048 px on a side is downscaled first; up
  to 10 images per message, shown as a thumbnail strip before sending. A
  message may be an image with no text.
• An annotation editor for local edits. Open any pending image (or hit
  "Annotate & reuse" on an image already in the conversation) and draw the
  region to change: pen, ellipse or rectangle in one of five colours and
  three widths, a mask brush, an eraser, undo / redo (⌘Z / ⇧⌘Z) and a mask
  preview. On send the annotated copy goes to the model, optionally with a
  black-and-white mask and the untouched original; an empty composer is
  pre-filled with "In the red area: …" naming the colours used. The
  original is never modified — the editor reopens on it.
• Images in the conversation gain "Save as…" and "Copy" on hover, next
  to the new annotate action.
• The app remembers which models are image models between launches, so
  the Image section is right at startup, before the model list has been
  fetched.

Changed

• Protocol 0.3.0. The image extension object is named generation in
  requests and in the model list (it was specterforge); a server that
  still tags its models with the old key is recognised, but image options
  are only honoured by Pictor ≥ 0.3.0.
• Fewer round-trips to the server. The model list is fetched at
  startup, when the base URL or API key change (once you stop typing) and
  on refresh — no longer every time a model is picked or a slider moves.
  Sampling and image options now travel with each request instead of
  rebuilding the HTTP client on every change.
• Composer, editor and model catalogue logic moved out of the widgets
  into testable units (ComposerNotifier, DrawingSession,
  ModelCatalogNotifier), behind domain contracts for image
  normalisation, rendering, file dialogs and the clipboard; widgets no
  longer touch platform plugins directly. Enforced by the architecture
  test. Test suite grew from 254 to 406.
• Dependencies brought up to date. Markdown is rendered by
  flutter_markdown_plus, the maintained successor of the discontinued
  flutter_markdown (same rendering); flutter_highlight, highlight and
  collection were declared but never used and are gone. Every other
  dependency is at its latest version.

Fixed

• A server error in the middle of a stream is reported. Some servers
  send the OpenAI error envelope as a stream event rather than an HTTP
  status; it was skipped as an unknown chunk and the turn hung until the
  connection closed. It now ends the turn with the error banner.
• Multi-byte characters split across two network packets (accents,
  CJK, emoji) were decoded as replacement glyphs; the stream is now decoded
  as a whole.
• A multi-megabyte stream line (a base64 image) no longer gets re-split
  on every packet, which made large images arrive slowly.

Full Changelog: v0.5.2...v0.7.2