[PoonHarpoon]

This PR ports over several QoL Kotatsu-Redo features and fixes that were also missing in Futon. Below is a summary of what is contained; most features and fixes can be stripped out.

Downloads, backups and local storage

Five small, independent fixes for downloads, backups and local storage. The temp-file, decimal-settings and local-index fixes are ported from Kotatsu-Redo.

Downloads: queued chapters were silently dropped. Each manga's download is enqueued as unique work with KEEP, and the scheduler only waited while that work was RUNNING. If an earlier download of the same manga was still ENQUEUED or BLOCKED (for example, waiting for charging or an unmetered network), the next queue entry was discarded by KEEP and then removed from the queue, so its chapters were lost. Such entries now stay queued until the earlier download finishes, and an entry is removed only once its work is stored.

Downloads: concurrent downloads deleted each other's temp files. Two downloads to the same storage directory cleaned up each other's temp files. Each download now only removes its own.

Backup: restoring sources added to the current selection instead of replacing it. Keeps the welcome logic from interfering with restoring sources. Covered by a new case in AppBackupAgentTest.

Backup: decimal settings failed to load after restoring a JSON backup. JSON backups bring decimal values back as Double, which putAll skipped, and a value stored with the wrong type made getFloat throw. putAll now stores doubles as floats, and the decimal reader settings (double-page sensitivity, brightness/contrast, autoscroll speed and others) read through a helper that recovers a wrongly-typed value.

Local storage: the index rebuild blocked the database on cold start. The storage scan ran inside the index rebuild transaction and held Room's write connection the whole time. The scan now runs first, and the transaction only writes the results.

Suggestions and related manga

Ported from Kotatsu-Redo, local parts only, with a few changes noted below.

Fixes

• The suggestions worker kept the least relevant suggestions. It sorted ascending by relevance, so it kept the wrong end. The notification picker also crashed with fewer than 3 suggestions (Random.nextInt(0, 0)).
• Related manga ignored "disable NSFW". NSFW titles now stay hidden when NSFW content is disabled.

Smarter suggestions

• Seed tags are now the reader's most telling tags, weighted by how rare each tag is across the local library (TF-IDF), rather than simply the most frequent ones. Blacklisted tags are never used as seeds.
• Dropped manga (barely read, untouched for 3 weeks, not a favourite) no longer feed suggestions.
• A source is never searched without a tag (that would return its whole catalogue). Two matched tags are searched together where the source supports it, falling back to one.
• Rating sort is preferred over popularity.
• Unlike Redo, suggestions stay off by default, so existing users who never opted in don't get a new background worker hitting every enabled source.

Related manga fallback
When a source links fewer than four related titles, the row is filled from the reader's stored suggestions that share this manga's tags, weighted by rarity. If that's still short, it's filled from one search of the manga's own source for its rarest tags. NSFW and the tag blacklist are respected. Extension sources skip the source search, because their shared repository has stateful paging. The related-list screen uses the same logic. Redo's community-server part is left out.

Covered by the new SuggestionRankingTest.

Webtoon reader

Three webtoon reader fixes, ported from Kotatsu-Redo.

Broken page sizing. Pages could collapse to zero height, render hugely zoomed in (infinite scale), or get cut off.

Pages cut off at page joins. A scroll now walks every page it reaches instead of only the first two, so a fast fling over short pages can no longer carry a tall page past the edge of the screen with its image unscrolled. Each page's inner scroll is re-pinned after every layout (decode, zoom, rotation). A page that finishes decoding after it has already scrolled partly off screen is placed without jumping to its end or moving what's on screen.

Pull-to-switch-chapter gesture rewritten. The EdgeEffect-based pull is replaced by a touch tracker that only starts a pull when both the list and the edge page's image are really at their first or last edge. The edge glow is turned off while the gesture is enabled. The feedback chip fades in from 0.15 to full opacity, switches to an outlined background once the threshold is reached, and hides as soon as you let go.

Tracker

Several fixes to how chapter progress is saved, to protect progress from being lost, and to how new chapters are fetched.

Counting fixes

• Tracks reset or stuck. Tracks shared by history and favourites no longer have their baseline reset on every run, and sources with update checks disabled no longer hold batch slots forever.
• New chapters not credited. Opening the reader on an older chapter moved the tracked chapter to the newest one without counting what had been released since. Those chapters now show up in Updates.
• Sources that change chapter URLs. When the tracked chapter is gone from the fresh list, the comparison falls back to its upload date, then to the history chapter, before re-baselining. A track's first check anchors on the history chapter.
• Read and old chapters reported as new. Chapters at or before the reading position (including the same title read on another source), or uploaded more than 90 days ago, are no longer reported.
• Long-inactive tracks. Tracks with no new chapter, read or favourite in 90 days drop out of regular checks but are still checked weekly. Stale unread counters are cleared.
• Clearing updates now resets each track to the history chapter, so unread chapters are recomputed from the reading position.
• Manual checks check every track even when the system refuses a foreground service.
• One-time recovery for tracks that were already advanced past unread chapters. It runs on first launch with network, needs no schema change, and is gated by two preference flags.

Throughput and save fixes (Not a port: New feat/fix)

• Per-source limit. At most 2 checks run per source, within the existing 6. The source slot is taken before the global one, so a busy source can't hold global slots while it waits.
• Blocked sources. After a source's first Cloudflare or rate-limit (429) failure, its remaining manga in the run aren't requested. They're stamped as checked in one update, so a blocked source can't fill the next batches either.
• No network inside the save transaction. Progress refresh could fetch pages while holding the save transaction open. It now runs after the save, and a failed refresh no longer fails a check that was already saved.
• The last-chapter date now comes from the tracked branch rather than from all branches.

Re-check after a captcha solve (Not a port: New feat/fix)

• Captcha failures are stored as a new result, RESULT_CAPTCHA (5). last_result is a plain int, so no migration is needed.
• Each successful solve queues one re-check of that source's captcha-blocked tracks. Solves come from the captcha notification, or from the automatic solver if that lands. It uses the existing CaptchaHandler.discard(), so it doesn't depend on any other PR.
• It can't loop: a re-check that hits the captcha again only shows the notification.
• New setting: "Re-check after solving a CAPTCHA", on by default.

New Features Ported

• Swipe to remove a feed item, with undo.
• New "Recently read manga" download strategy: new chapters of manga read in the last 30 days are downloaded, up to 5 per manga per run. Nothing is queued with less than 2 GB free or when the download quota is reached. Tracker downloads now go through the download queue, so the Wi-Fi, charging and off-peak settings apply to them. (Modified: in Redo the 5-chapter cap also hit the existing "Downloaded manga" strategy, so chapters over the cap were never downloaded; here only the new strategy is capped.)

Cloudflare

Fixes to how the app gets through Cloudflare, plus an experimental opt-in automatic captcha solver. Most of this is ported from Kotatsu-Redo.

Fixes

• Request interception broke after a Cloudflare interstitial. The page script was injected only once, so captures timed out when a challenge page came before the real one. It's now re-injected on every navigation and polled while capturing. Captures are serialised and reuse a single WebView, and evaluation WebViews are now destroyed after use, so they no longer leak.
• The clearance cookie didn't match the User-Agent. If the solver WebView had a non-Android User-Agent, Turnstile could fail. The UA is now realigned with the engine's default, and the one that earned clearance is stored per source, so OkHttp sends the agent the cf_clearance cookie is bound to.
• The user's own User-Agent setting wins (Modified). If the user changes the source's User-Agent setting, the stored clearance agent is ignored. In Redo the stored agent was never cleared and silently overrode the user's setting.

Automatic captcha solving (experimental, off by default)

• New global switch under Storage and network: "Solve CAPTCHAs automatically (experimental)". It's off by default, and there's also a per-source opt-out.
• When it's on, opening a catalog, a manga or a chapter first tries to pass the challenge silently, in a translucent WebView activity. A read-only page-state probe decides when the real page has loaded.
• Only one session runs at a time. Cooldowns are 30 s after a success and 3 min after a failure. It only starts while the app is in the foreground and is cancelled when it goes to the background. Background workers, prefetch and image loading never start a solve.
• The headless WebViewExecutor.tryResolveCaptcha path and its two WebView clients are removed, because CaptchaHandler was their only user. AdBlock is kept.
• It's marked experimental because it can't solve interactive challenges, and every challenge I got while testing was interactive. It should pass non-interactive ones, but I haven't been able to confirm that.

Library

• Progress badges no longer show a leading "-" on "percent left" and "chapters left".
• Hide NSFW from launcher shortcuts: a new setting, on by default. NSFW titles are also left out of shortcuts whenever NSFW content is disabled.
• Pin favourites. Pin or unpin from the favourites selection mode. Pinned entries sort first under every sort order and show a pin badge in grid and detailed list modes. A manga pinned in any list counts as pinned in the all-favourites view, and pinning from that view pins it in every list it's in. It uses the existing favourites.pinned column, so no migration is needed.
• Quick filters for favourites: Manga / Manhwa / Manhua chips, plus the three most common tags of the current list. Tags are matched by title, so the same genre from different sources is one chip, and blacklisted tags aren't offered. Mihon sources report everything as plain manga, so their content type comes from a manga/manhwa/manhua genre tag. A Mihon title without one is only listed when no content-type chip is selected, instead of being misfiled under "Manga".
• Search within History and Favourites. On the History and Favourites screens, the search bar can now search just that screen. An option in the search bar switches between normal search and "History" or "Favourites". In screen mode, typing filters the list you're looking at instead of opening a new search. It matches the manga's title (including alternative titles), its source, and for favourites, the list it's in. While the filter is active it stays visible in the search bar with a button to clear it, and it's cleared automatically when you leave the screen.
  • Changes from Redo: it also works with Mihon sources, the search text is passed to the database safely rather than pasted into the query, and the Downloaded view is filtered the same way.

Covered by the new ContentTypeConditionTest and LibrarySearchTest.

Sources and alternatives

• "Content unavailable" dialog. When a manga's source says it's gone (404 or removed), the app now says so and offers a shortcut to search for it on other sources. This also works with Mihon extension sources.
• Source presets. Save named sets of languages and sources, and switch between them from Explore. While a preset is active, Explore and global search use only its sources, and adding or removing sources in the catalog edits the preset rather than your enabled sources. Uninstalled sources and deleted presets are handled cleanly. The active preset survives a settings backup and restore.
  • This adds a new database table, source_presets (migration 30 → 31). Note for sync: requests will now send X-Db-Version: 31.
• Replace source (Modified). "Fix" in the favourites and history selection menus becomes "Replace source", and it's also in the manga details menu. It checks every candidate source against your library titles from the old source, lets you pick the best one, and moves the matches over automatically. Unlike Redo, it asks before replacing every matched title. Anything it can't match, you choose by hand. Each title is migrated on its own, so one failure doesn't stop the rest.
• Search options for alternatives. The alternatives screen gets options for:
  • which sources to search (enabled, disabled, or a preset)
  • same language only
  • same content type only
  • hiding results without chapters
  • four sort orders
  • an editable search query
  • a progress counter
• Covers from other sources. The manga edit screen can pick a cover from other sources. Duplicate covers are filtered out, and the chosen cover is saved locally.

I did not port Redo's community ranking system.

JPEG XL

Ported from Kotatsu-Redo.

• JPEG XL images now display everywhere: covers, thumbnails, previews and reader pages. They're detected by their content as well as by the .jxl extension.
• If the JPEG XL library fails to load on a device, only JPEG XL images fail; nothing else is affected.
• New dependency: io.github.awxkee:jxl-coder 2.6.1. It bundles native libraries for each CPU type and adds about 11 MB to the debug APK (48.7 → 59.8 MB). It's a large dependency for a niche feature, so it's up to you whether it's worth adding.

Reader and history

• Reading position kept when a source changes chapter IDs. Some sources give chapters new IDs between loads, which used to lose your place. The history now follows the same chapter in the new list (matched by URL, then volume and number). If it can't be matched, the position is estimated within the branch you were reading, not across all branches. If your history chapter is gone entirely, progress is rescaled to the new chapter count instead of being dropped; that estimate only ever lowers progress, and never marks a manga as completed.
• Default branch re-picked once the fresh chapter list loads, not only from the cached one, so a recovered position opens in the right branch.
• Smart resume fix (Not a port: New feat/fix). Continue reopened the finished chapter after a new release instead of moving on, because the progress recalculation overwrote the chapter count that smart resume relies on. The count is left to the reader again.
• Double-page rework.
  • Two chapters never share a spread.
  • New option "First page as cover (solo)" shows each chapter's first page on its own.
  • Pinch-zoom and pan now work on the whole spread instead of one page at a time.
  • Pages are centred against each other instead of the screen edges.
  • Wide images span the full spread.
  • Right-to-left mode gets identical spreads.
• E-Ink flash on page change. An optional full-screen white or black flash on page change clears ghosting on E-Ink screens. Duration and frequency are configurable. It doesn't flash when opening a manga, switching reader mode, or during webtoon autoscroll.

Overlaps with #118 in DoubleReaderFragment and ReversedDoubleReaderFragment; I'll rebase whichever lands second.

Testing

Unit tests pass (157, 0 failures) and the debug build succeeds. AppBackupAgentTest and ProgressUpdateRaceTest are instrumented tests and weren't run.

I tested all of these together and confirmed they work on my device (S26). Automatic captcha solving is the exception: it couldn't be verified, since every challenge I got was interactive.

[LandWarderer2772]

feels the aura
holy moly this is gonna be a pain to review

[LandWarderer2772]

brother u on discord?

[PoonHarpoon]

I have an old account but I do.

[PoonHarpoon]

Most of this is 99% the same code as Kotatsu-Redo but some more than others needed work to get them, well, working.

[LandWarderer2772]

Most of this is 99% the same code as Kotatsu-Redo but some more than others needed work to get them, well, working.

promoting u to maintainer cause i need help 🙏🏻

[PoonHarpoon]

These issues are probably fixed by this PR:

• Tracker worker aborts mid-batch when one tracked title is cancelled; remaining titles are never checked #125: the tracker no longer tries to solve captchas, so one blocked source doesn't stop the rest of the batch
• Memory leaks when enabling all sources #27: failed favicon loads no longer queue a captcha solve that holds the view, which is the path every LeakCanary trace in the issue goes through
• reading progress percentage calculation #102: the stored progress now refreshes after new chapters arrive, and titles already stuck at "completed" get fixed once

#124 is partly covered: JPEG XL images now load, but there's no conversion feature.

[PoonHarpoon]

Currently fixing some regressions with this

[LandWarderer2772]

Currently fixing some regressions with this

got it

fixes #125
fixes #27
fixes #102