vavallee/bindery v1.29.0

on GitHub

A release about failures that never said anything. Almost every fix here was a
bug you could only find by noticing an absence: an author's catalogue arriving
50 books short with zero skip counters and no log line, ON DELETE CASCADE
quietly switching itself off partway through a process's life, a language
filter dropping books at a log level nobody runs, dismissed recommendations
returning because two code paths were writing to two different users. None of
them errored. Several had been reported as "it just doesn't work" precisely
because there was nothing to quote.

Most of this came from the Discord bug reports, several with source-level
diagnoses attached. Credit is on the individual entries.

Added

• User guide (#1721) — a concepts guide for new users covering the five
  things that account for most support traffic: catalogue first and files
  second, status versus monitored, how monitor modes decide what gets grabbed,
  what the importer will and won't do with an existing folder, and where
  metadata actually comes from. Written against a sweep of real issues and
  Discord threads rather than from the code.

Fixed

• Audible catalogue no longer truncates prolific authors (#1751) — the
  Audible author lookup issued a single request capped at 50 products and never
  paged, so any author with a larger audiobook catalogue silently lost the
  remainder. There was nothing to notice: no error, no log line, and the sync's
  skipped_language / skipped_junk / skipped_media_type counters all read
  zero, because the missing books were never enumerated in the first place. One
  reported author with 105 titles imported 56. The reason it went unspotted for
  so long is that the endpoint's page parameter is 0-indexed, so an unpaged
  request and page=1 return disjoint windows rather than the same one — an
  entire 50-item window was going unrequested, which looks nothing like a
  normal off-by-one. Pages are now walked to completion and deduplicated by
  ASIN. Reported by .v.e.g.a.
• Foreign-key cascades no longer stop working on long-running instances
  (#1727, #1728) — foreign_keys and busy_timeout were applied once at
  startup, but they are per-connection settings in SQLite. Whenever the pool
  replaced its connection, the replacement came up with foreign keys switched
  off, silently disabling every ON DELETE CASCADE in the schema for the rest
  of the process. They now travel in the connection string, so each new
  connection gets them. The visible symptom this fixes: deleting an author left
  its identifier rows behind, and re-adding that author then failed forever with
  "author already exists" even though the author was gone. Author deletes now
  remove those rows explicitly, so correctness no longer depends on pragma state
  at all, and an upgrade sweeps away any orphans an affected install is already
  carrying.
• Adding a specific book to an author you already track no longer fails
  forever (#1612, #1735) — the add only polled for a row the catalogue sync
  had already refused to create, so it returned "book not found after author
  sync — try again shortly" on every retry and no amount of waiting helped. The
  requested book is now fetched and saved directly, and if you already own it
  under another source (a Calibre import, or one half of a split
  ebook/audiobook record) the existing entry is reused instead of a duplicate
  being created. Diagnosed and originally fixed by helios57.
• Dismissed recommendations no longer come back (#1725) — requests that
  authenticate as the install itself, from a trusted local network or with an
  API key, were treated as an admin but carried no user identity, so a refresh
  triggered that way saved its recommendations against a user that does not
  exist while the nightly job saved them against a real one. Dismissals
  recorded on one batch never applied to the other. Both now resolve to the
  same operator account, and an upgrade hands any stranded recommendations and
  dismissals back to it.
• Discover no longer recommends books you already have (#1726) — ownership
  was matched on provider-specific IDs, so a book owned via OpenLibrary or
  Audiobookshelf still came back as a Hardcover recommendation (one reported
  batch was 15 of 22 owned books). The recommender now also matches by work
  identity (canonical title key plus author), and books on your wanted list
  count as owned too instead of being recommended back.
• A profile allowing "eng" no longer rejects books that arrive as "en" or
  "en-US" (#1729) — the allowed-languages filter compared provider codes raw,
  but providers speak different vocabularies: Google Books returns ISO 639-1
  (en, en-US) while profiles are written in ISO 639-2/B (eng), so those
  books were silently skipped at the default log level. The filter now runs both
  the incoming code and the profile's list through the same canonicaliser the
  EPUB importer already used, so any spelling of a language matches any other.
  Books previously skipped will appear on the next metadata refresh.
• "Add all" on a light novel or manga series now adds every volume (#1682)
  — filling a Hardcover-linked series created only the first volume and silently
  linked that same book to every other position, so a 5-volume series produced
  one book and a 13-volume series produced one book. The duplicate check
  compared titles by fuzzy similarity, and volume titles differ by a single
  number: Trapped in a Dating Sim Vol. 1 scores 100 against Vol. 13,
  The Mimosa Confessions Vol. 1 scores 96 against Vol. 2. An explicit volume
  number now vetoes a title match, so those are recognised as different books.
  Titles with no volume marker are unaffected, and a bare number in a title
  (Fahrenheit 451) is still not treated as one. The same cause made an author
  page show a single volume until you refreshed their metadata, which was the
  workaround people found.
• Author metadata candidates are now ranked, and you can check one before
  linking (#1754) — "Link metadata" listed candidates in whatever order the
  provider returned them. OpenLibrary emits a composite author record per
  anthology, one row naming every contributor and holding a single work, so
  those led the list while the record carrying the author's actual catalogue sat
  below them and the picker looked like it couldn't find the author at all.
  Candidates are now ranked by name match then catalogue size, reusing the
  ranking the normal author search already had. Records from different providers
  are still never merged, so picking one specific provider's record still works.
  Each candidate also gained a View on OpenLibrary link, since several can
  share a name and there was previously no way to tell them apart without
  leaving the app; it is omitted for providers whose IDs have no public page.
  Reported by nzvengeance.
• Ebook-pinned Hardcover lists no longer create books as "both" (#1732) —
  edition hydration ran right after a list-synced book was created and widened a
  pinned ebook media type to both whenever the work had any audio-shaped
  edition on Hardcover (true for most popular titles). Hydration now knows when
  the media type was deliberately pinned by the list and leaves it alone; books
  with no media type at all still promote to audiobook when an audio edition
  exists.
• A series genre override can now be seen, edited, and removed (#1709) — the
  override applied to books added later was stored but never sent to the
  browser, so the "Set genre" prompt always opened blank and there was no way to
  tell an override existed or to undo one. The series list now carries it, the
  prompt opens pre-filled, the button reads Genre ✓ while an override is
  active, and submitting an empty box removes it. Clearing drops the rule for
  future books and leaves the genres already written onto existing books locked,
  since those were a deliberate edit. The override can also now be set before
  adding a series' books, and applies to the books created by that add.
• Expanding one series card no longer stretches its neighbours (#1682) — the
  series grid used the CSS default of stretching every card in a row to the
  tallest, so opening one made all three grow and it was hard to see which was
  actually open.
• Stopped the qBittorrent poll logging a Debug line for every already-imported
  download (#1730) — the 15-second poll walked all download rows and logged
  "download not found in torrent list" for long-imported downloads whose torrent
  had been removed after import, one line per download per poll forever (96% of
  debug output on a 53-download library). Terminal downloads are now skipped
  before the log, matching the no-hash branch; failed-import rows still log and
  still feed stale-source detection unchanged.

Upgrade notes

• Two data migrations run on first start. 068 removes author-identifier rows
  orphaned by the cascade bug above. 069 re-parents recommendations and
  dismissals stranded under a non-existent user to the operator account. Both
  are one-way; the affected rows were unreachable before, so nothing you could
  see is changed.
• Audible author lookups now make one request per 50 titles instead of always
  exactly one. For a large library the first refresh after upgrading will do
  noticeably more outbound requests, and will find books it previously missed.
  A lookup that fails partway now reports an error rather than returning a short
  list, so an Audible outage degrades to "no audiobook supplement this run"
  instead of a silently truncated catalogue.