Many modern crawlers (AI scrapers, headless Chrome, link-preview fetchers) run JS and
fire the visit/summary_viewed beacon, inflating "visitors" even though there's no
human discovery channel. Apply queries.is_bot_ua() at /api/events — the same filter
the load-error beacon uses — so honest bot UAs (GPTBot, AhrefsBot, headless Chrome,
python/curl, …) are dropped before recording. Response is identical so a bot can't
detect it. Counts read lower but truer going forward (past rows unchanged). Won't catch
UA-spoofing bots; that needs a heavier heuristic. Tests: bot UAs dropped, real browser
counted; existing event tests send a real UA (default client UA contains "python").
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Path-only @junk matcher on upbeatbytes.com (*.php, /wp-*, /.env, /.git, /phpmyadmin,
/vendor, etc.) returns 403 instead of falling through try_files to a 200 SPA shell.
Never matches by User-Agent, so real users + Googlebot/Bing are untouched. Applied to
the live Caddyfile (validated + reloaded) and mirrored into the repo snapshot.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Pages tall enough to scroll showed a ~15px scrollbar; short pages didn't — so the
centered top bar jumped left/right as you navigated. scrollbar-gutter: stable on html
(SPA app.css + the server-rendered share pages) keeps the layout width constant. No-op
on overlay-scrollbar platforms (mobile), which never shifted.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The server-rendered /a/<id> and digest pages predated "HubBar everywhere" and showed
a stripped bar (logo + a bespoke Back pill). They can't run the Svelte component, so
add a hand-kept static replica of HubBar (logo + News/Play/Art nav + account glyph +
mobile burger/drop-panel) plus HubShell's borderless ← Back. A signed-in reader's
avatar paints from the same localStorage cache HubBar uses. /a/<id> now looks like any
detail page (/art, /word). Reusable _top_bar_html/_TOP_BAR_CSS/_TOP_BAR_JS/_back_link
helpers; applied to both share pages. Kept in sync with HubBar.svelte by hand (noted).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The on-demand /api/img path runs in the container (only fastapi+uvicorn), so without
Pillow it fell back to caching the original full-size bytes instead of a downscaled
WebP. Add Pillow>=10 to the web extra. The host cycle already had it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
thejayman77