v0.15.0

Added

• fix.fragment_wrong_path — a link whose #fragment is real but has moved
  to another document (RSC-012 / opf.content_document.dangling_fragment,
  ConfirmNeeded). If exactly one document in the book defines that anchor, the
  path is written in front of the fragment, relative to the referring document;
  where the anchor turns out to be in the referring document itself, the link
  becomes a plain same-document #fragment.

  On real books this is an editor splitting a file at a page break and leaving a
  footnote link and its return link as bare fragments pointing into the wrong
  half. Third member of the family beside fix.reference_wrong_path and
  fix.ncx_src_wrong_path, sharing their machinery and their discipline:
  exactly one answer, or decline.

  It declines an anchor defined in several documents (which one the link meant
  is a guess), an anchor defined nowhere (only "drop the fragment" is available,
  and that loses the author's target), a target outside the reading order, and a
  reference not visible as a whole quoted attribute value.

• --format json's summary reports every counter it has a concept of.
  Seven fields appear: proposed, and warnings_before/warnings_after,
  infos_before/infos_after, usages_before/usages_after.

  Under --dry-run the summary used to say applied: 0, skipped: 0 while every
  item carried "outcome": "proposed" — so the summary and the items disagreed
  about the size of the run, and a consumer asking the summary how many fixes
  there were got zero while two sat in items. Every number was true and the
  document misled. The identity a consumer may now rely on is
  applied + skipped + proposed == items.len().

  The severity counters are the same defect one field over. Three fixers here
  dispatch on usage or warning findings, so a run can clear real work while
  errors_before/errors_after do not move at all — and the summary described
  that as no change whatsoever. reverted stays absent rather than zero: a build
  that cannot revert has no concept of the value, and 0 would claim no revert
  happened where the truth is that none could.

  The wasm binding's Summary gains the same seven fields, its skipped
  remaining zero because the browser never declines.

Changed

• The envelope claims convention 0.5. The stability key is an assertion
  about this crate, not about its detector: it moves when epubsana implements a
  convention release, and the seven counters above are that release.

• fix.non_preferred_media_type takes the replacement from the finding
  instead of from a table of its own. epubveri has named the preferred media
  type in params[1] since 0.12.3; this crate went on consulting a five-row
  table written before that existed, which made it a second answer to a question
  the detector was already answering.

  It closes a hole the table could not see. Where epubveri declines to name a
  replacement — because the resource's own signature rules the candidate out, an
  OTTO (CFF) font declared application/x-font-ttf being the shape — a
  table-driven fixer renamed it anyway. No book on the 474-book shelf carries
  that shape (121 application/vnd.ms-opentype + 285 application/x-font-ttf
  declarations, none contradicted by its own bytes), so this closes a hole
  rather than repairing damage.

  Two declines are new and deliberate: a finding that names no replacement, and
  two findings about one declared type that name different replacements —
  which epubveri 0.14.1 can now legitimately produce for
  application/font-sfnt, because it answers that row from each file's own
  signature. The fix renames per declared type across a manifest, so it cannot
  honour both and proposes neither.

• application/font-sfnt is now repaired where the detector decides it. It
  was declined outright because SFNT is the container TrueType and OpenType
  share and EPUB 3.3 lists the spelling on both rows — the name cannot say
  which the file is. epubveri 0.14.1 answers it from the font's own signature
  (only font/otf admits OTTO), on its side of the boundary, where deciding
  what the defect is belongs. This crate still never opens a font. Zero books on
  the shelf carry the spelling, so nothing here moves today.

• The envelope states epubsana's own convention version. epubveri 0.13.x
  hard-coded its own key into every envelope built through
  Envelope::for_tool, so this crate's --format json was claiming epubveri's
  convention version; it happened to be right, both being v0.4. 0.14.0 made
  the key a required parameter. epubsana claims "0.4" and will move it
  when it implements v0.5.0, not when its dependency does. No output changes.

Dependencies

• epubveri floor raised to 0.14 (from 0.13.7) — a capability floor: the
  params contract the OPF-090 fixer now reads (an absent params[1] meaning
  decline, and only a real media type ever appearing in it) does not exist
  below it. Measured across the boundary on 474 books: the whole-shelf plan
  digest and the entire regression audit are byte-identical — 194 books planned,
  2,337 proposals, errors 40,584 → 15,719, 73 books fully valid, zero
  regressions at any severity.

Documentation

• fix.navdoc_empty_nav's safety argument was overstated in 0.14.0 and is
  corrected. That entry says the finding cannot say which <ol> it means. Only
  params cannot: navdoc.ol.empty is emitted with an element_path, which on
  every shelf book names the exact node. The fixer does not read it — a choice,
  not an absence.

  What actually makes its file dispatch safe is a subset property: for a
  typed <nav> whose single element child is an <ol>, epubveri always reaches
  the check that reports an empty list, so every node this fixer selects is one
  epubveri reported. No code changed and no behaviour was wrong; the reasoning
  was, and it is the reasoning a future session would rely on.

Full notes: CHANGELOG.md. Also on crates.io.