v1.5.0

Changed

• Ground every run against a written inventory of the screenshot, and retire the "Ground with
  description" toggle (along with the CLI --ground flag and the MCP ground argument). A
  head-to-head test over two screenshots, five vision models, two prompts and two variants gave
  grounding 9 of the 20 model-and-prompt matchups on content fidelity against 1, the rest tied and
  the single loss going to a model that was truncating at its token limit under both conditions. It
  kept all of a dense panel's content where ungrounded runs kept 87%, and blind judging over the
  same 39 output pairs preferred it overall, with a third as many dropped elements and half as many
  fabricated ones. The cost is one shared description call per input, so there was nothing left for
  a setting to decide.

Fixed

• Rotate to another API key when a generation fails on a dead or exhausted one, instead of failing
  the request while healthy keys sit unused. The screenshot-description and run-naming calls used to
  try a single key and silently give up, which cost a run its grounding; they now rotate like the
  main generation loop, which itself now gets one attempt per key in the pool rather than a fixed
  six. A model whose keys are all revoked is also no longer chosen to describe screenshots.
• Clear the run progress card on load when the run it was watching finished more than five minutes
  ago, so opening the app later starts on an empty card instead of replaying an old run's results.
  A run still queued or generating is always restored, however long the app was closed.
• Raise Qwen 3.5 Plus's output limit, which was half the rest of the fleet and truncated most of its
  pages mid-document.