Modernize the comparison UI and add a review workflow

[andiwand]

Reworks the pages served by compare-html-server so that a regression review can be done from the compare page itself, and moves the markup out of Python f-strings into Jinja templates plus static assets (shipped via [tool.setuptools.package-data]).

Compare page

• Update reference button (and u) copies the monitored file over the reference one, then reloads both panes and the diff.
• Accepted indicator: a file whose reference was updated is marked ✓ accepted next to its live status for the rest of the session. The mark is dropped again if the monitored file changes afterwards. In-memory only, nothing is written to disk.
• Previous / next navigation with a position counter (5 / 9) and a diffs only toggle that skips files that already match.
• View modes (A / A|B / B), a collapsible and expandable diff panel, and keyboard shortcuts (←→ jk 1 2 3 d u r Esc ?) that also work while the focus is inside a pane.
• The diff strip became usable: raw difference images are almost black, so differing pixels are painted red, differing regions are drawn as overlay bars (a handful of pixels disappears when the image is scaled into a 72px column), a marker tracks the visible part of the document, clicking jumps both panes to that position, and it reports where the first difference is.
• Scroll sync is attached on every iframe load — previously the listeners were registered on the pre-navigation window and dropped again.

Overview page

• The per-status counts moved into the filter buttons (All 9 · Differences 6 · Pending 0 · Same 3 · Accepted 1) instead of a separate chip row that named the same states in a different order.
• The status cell is a plain table cell again, so the columns line up.
• Added an Accepted filter, toasts instead of alert(), and polling that patches rows in place and only reloads when the set of files actually changed.

Fixes found on the way

1. A comparison could start before submit() had stored its future, so compare() raised RuntimeError("Path not submitted for comparison") inside the worker and the file stayed on pending forever. Reproduced on every run.
2. A failing comparison (file vanished, render error) left the same permanent pending; it is now logged and reported as different.
3. Screenshots are taken in a 10000px tall window, so a typical page was ~5% content and ~95% blank and every position derived from a diff was off by that factor ("first diff at 24%" for a difference at 51%). Renders are now cropped to the content, which also fixes html-render-diff's first error at X% output.
4. Rendered diffs are cached by both inputs' mtimes; previously every request drove the browser again, and the compare page needs the image and its statistics.
5. Paths are resolved against their root before serving or copying.

Compatibility

update_ref is now POST /api/update_ref/<path>. The previous GET /update_ref/<path> and /status are kept for scripted use.

Tested manually against --driver chrome, including from a clean install of the package to confirm the templates and static files are packaged.

🤖 Generated with Claude Code