v0.1.0

First release.

Conversion

• Encoding repair — candidate-scoring detector covering UTF-8/16, GB18030,
  Big5, Shift_JIS, EUC-KR and windows-1252. Scores the decoded result rather
  than guessing from raw bytes, so mojibake is caught even on short files.
• Hard-wrap re-joining — detects text folded at a fixed column by measuring
  how tightly line widths cluster under a ceiling, then rebuilds the paragraphs
  without inserting spaces into CJK or splitting numbers.
• Advertising removal — a pattern library (site watermarks, download URLs,
  求推荐票, QQ groups, 本章完) plus a frequency heuristic that catches anything
  short repeating once per chapter. Every deletion is reported; none is silent.
  Rules are deliberately conservative — see Design notes below.
• Chapter detection — a library of heading patterns scored on the result set
  they produce (ordinal continuity, chapter evenness, coverage, standalone
  lines), so 章/回/话/节, Chapter N, bracketed and bare-number headings all
  work, and 第三章讲的是… inside a paragraph does not become a heading.
• Two-level structure — 卷/部/篇 detection with 楔子/番外/后记 kept at book
  level rather than buried inside a volume.
• Metadata inference — from 《书名》作者.txt-style filenames, scraper
  header blocks, config files and CLI flags, with the source of each field shown
  in the report. Language is classified as zh-Hans/zh-Hant/en.
• Generated covers — deterministic SVG covers rasterised to PNG via resvg,
  falling back to embedded SVG when no CJK font is available.
• EPUB 3 output — per-chapter XHTML, nav.xhtml plus toc.ncx, Dublin Core
  metadata with file-as sort keys, EPUB 3 collections and calibre series tags,
  landmarks, and a stylesheet that sets structure without pinning font family,
  size or colour.
• Built-in validation — container invariants, manifest/spine consistency,
  XML well-formedness and unescaped-& detection on every document, with
  optional hand-off to epubcheck when it is installed.
• inspect subcommand — full diagnostics without writing a file: encoding
  candidates, pattern scores, deleted lines and a structure preview.
• Batch conversion — directories, recursion, and parallel processing.
• Profiles — webnovel, published, lightnovel, raw.

Online metadata lookup

• --scrape searches by title and author — either inferred or supplied with
  --title/--author — scores the candidates, and fills in the author, blurb,
  genre tags, completion status, update date and cover. --scrape-url takes a
  book page directly and skips the search.
• Scraped values are written verbatim and only ever fill gaps; anything set
  in a config file or on the command line wins.
• Pages are read through the og:novel:* meta convention, which most Chinese
  fiction sites implement, so --scrape-url is not tied to one provider.
• Off at runtime unless asked for, and a compile-time feature that can be
  removed entirely with --no-default-features.

Distribution

• Static binaries for Linux (gnu + musl, x86_64/aarch64), macOS (Intel + Apple
  Silicon) and Windows (x86_64/aarch64), plus a multi-arch Docker image.

Design notes

Two capabilities were built, tested against a real 13 MB / 6.4M-character novel,
and then removed because they damaged data in ways their test fixtures did
not reveal. Both are recorded here so they are not reinvented:

• Release-credit ad rule. A pattern matching lines that ended near 制作 /
  整理 / 校对 deleted roughly 10,000 characters of story from that one book —
  those are ordinary verbs, and 制作 appears on nearly every page of 修仙
  fiction. No rule for bare credit lines remains; only the anchored 本书由…整理
  form. Author asides (求月票 and friends) additionally apply to short lines
  only, so a paragraph that merely starts with one survives intact.
• Official chapter titles. Adopting titles from a site's catalogue worked on
  the fixture and matched 0 of 1198 chapters on the real book, because real
  catalogues renumber across volumes, duplicate ordinals and skip them. Blurb
  rewriting went the same way: trimming marketing copy meant hand-tuning a
  keyword list against two examples.

The governing rule: a surviving blemish is better than deleted prose. Where
a heuristic cannot be made reliable on unseen input, it is not shipped.

Scraping requests metadata only — not chapter text, and not the chapter list.
It is rate limited to one request per 1.5 s by default, runs sequentially, and
falls back to local inference on any failure, including a site returning an
access challenge, which is never worked around.