Rescripto v1.3.0

Changed

• Rewrites read more like a person wrote them, in every tone and in Local,
  Cloud and Hybrid mode. All tones now share one set of writing rules: no
  added greetings, sign-offs, filler phrases, hype or em dashes, and no new
  facts or apologies the draft didn't have. Tones still sound different from
  each other.
• Rewrote the 14 built-in tone instructions. Persuasive and Marketing no
  longer invent claims or reach for words like "seamless". Academic keeps
  claims as certain as you wrote them. Confident drops "I think" and "maybe".
  Longer means fuller wording, not made-up detail. Tones you edited yourself
  are not changed.
• Small on-device models turn plain statements into "Could you…" questions
  much less often, and extra versions that a model labels "Variant 1:" are
  split properly.
• If a model hands your draft back unchanged, Rescripto asks it once more
  before showing the result. Light polish is left alone, since correct text
  can fairly come back the same.
• The on-device engine is now Rescripto's own llama.cpp plugin. It runs on the
  same llama.cpp release and CPU backends as before. The previous plugin was
  under a non-commercial license that ruled out F-Droid and other stores; the
  whole app is now Apache 2.0.
• Clearer wording across the app: errors say where to fix the problem, dates
  in backup and sync read normally, and several descriptions were corrected.

Added

• Ollama and custom endpoint providers have a Models field. Before, they could
  be saved but never used, because there was nowhere to enter a model name.
  Other providers can list extra models there too.
• Release builds without signing keys now produce an unsigned APK, and APKs
  no longer carry Google's encrypted dependency block. Both are needed for
  F-Droid.

Fixed

• WebDAV server passwords shorter than 8 characters were rejected, so those
  servers couldn't sync at all.
• Exporting or restoring a backup, and syncing, could fail without any
  message.
• Stopping a workflow showed "Cancelled." in red as if it had failed, and a
  failed workflow showed its error twice.
• Stopping a rewrite while the on-device model was still loading now stops it
  before generation starts.
• Removed the GPU acceleration switch. The engine has been CPU-only for a
  while, so it changed nothing.
• Removed the "Update checks" privacy switch for a feature the app doesn't
  have.
• "Get an Groq API key" and similar button labels, and a workflow example
  that mentioned a Proofread tone that doesn't exist.