v1.6.0

Added

• Shared panel cache (Settings → Library, on by default): panel
  layouts are looked up from a community cache before detecting — so a page
  detected or hand-corrected on any server gives you instant guided view —
  and your own detections and corrections are contributed back (only
  ML-detected layouts and hand-corrections are shared — the built-in
  detector's output stays local). Only panel
  rectangles and a hash of the page's image bytes are sent; never image data
  or filenames. Hand-corrections outrank model output everywhere; model
  layouts are only shared once two servers independently corroborate them.
  Turn the toggle off to opt out entirely (detection then runs locally only).
  Confirming a page in the studio's review mode also counts as a vote of
  confidence on the shared layout (retracting the review retracts the vote);
  community layouts that collect enough rejections stop being served.
• Panel studio: users with the new reader.panels.edit permission
  (admin tier by default) get an "Edit panel layout" action on issue rows
  that opens a full-issue editor — a page-thumbnail rail beside an editable
  layout view. Drag corners (slanted panels supported) or whole panels,
  add/delete panels, reorder the reading sequence, force a page to read
  whole, or revert pages to automatic detection; staged edits across any
  number of pages commit with one Save. Corrections are saved per file,
  apply to everyone on the server, beat any detector's output, and survive
  model upgrades.
• Guided panel view: read a page panel by panel — the reader detects the
  panel layout server-side (computed once per issue, cached) and the g key
  or the new toolbar button steps through panels with an animated camera,
  and a spotlight dims everything outside the current panel. While an issue's
  layout is being detected the reader shows live page-by-page progress.
  Right-to-left series tour panels in manga order. Pages without a confident
  layout stay whole pages.
• ML panel detection downloads itself: when the ML toggle is on and no
  model is installed, the reader fetches the current panel model from the
  BackIssue CDN on startup (verified by checksum before it's ever loaded)
  and activates it without a restart. A model you placed by hand is never
  replaced; a custom readerPanelModel path disables auto-management.
• ML panel detection (optional): drop a panel-detection model at
  <data dir>/models/panels.onnx (or set readerPanelModel to its path) and
  panel layouts come from a neural detector that handles black gutters,
  borderless cartoon panels, and low-contrast layouts the built-in detector
  can't — in testing it produced guided layouts on 91% of pages vs 34% for
  the classical detector. Without the model (or the optional
  onnxruntime-node dependency) the classical detector keeps working
  unchanged, and a Settings → Library toggle switches between the ML and
  built-in detectors at any time. Turning the model on or off recomputes
  each issue's layout once.