v0.9.0

Twenty-four fixers to twenty-eight, plus the widest single extension this
project has made. On the shared shelf — 125 books, epubveri 0.9.14 — errors fall
13,141 → 4,027 over the 28 books epubsana touches, six books reach fully
valid, and both whole-shelf instruments report nothing introduced anywhere.

Two things in here were found by the instruments rather than by reading code, and
both are the reason those instruments exist. The <blockquote> extension came
from a shape census that showed stray text is reported in exactly two containers
on 125 books. The identity fixer's first cut cleared OPF-030 and produced three
new NCX-001s — an edit that was correct while the book got worse, which no
unit test can see.

One was found the other way, by reading epubveri's source: the playOrder
renumbering was target-blind and could have created the mismatch it now
repairs, on a book shape the corpus does not contain.

Added

• A fixer for an anchor target wrapped around a link
  (htm.epub2_dom.nested_anchor). On the shelf every case is one shape — a
  footnote reference whose outer <a> carries no href:

  <a id="bookmark1"><sup><a href="#footnote1">1</a></sup></a>
  

  That outer element is not a link but an anchor target, the legacy way of
  naming a position from before every element could hold an id. It is unwrapped
  and the id moves to its single child, so #bookmark1 still resolves, to an
  element at the same place in the same rendered line.

  Declines when the outer anchor has an href (a real link, and which of two
  nested links to keep is not ours to decide), when it carries any attribute
  besides id, when the child already has one, or when it wraps more than that
  one child. 6 findings in one book, all repaired, nothing introduced.

• A fixer for a package whose declared identifier points at nothing usable
  (opf.package.unique_identifier_unresolved + opf.package.opf_identifier_not_empty).
  Two rules, one defect at two stages, hitting disjoint sets of five books each:
  either no <dc:identifier> carries the id unique-identifier names, or the one
  that does is empty. The declared id is attached to the book's single real
  identifier, and any leftover empty element is dropped.

  It invents nothing. The value was already in the book, written by its
  producer; the id was already in the package, written in the attribute. The
  repair only attaches one to the other — the same principle as empty_title
  moving a TOC label the author wrote.

  The whole-shelf audit earned its keep here. The first cut cleared OPF-030
  and produced three new NCX-001s: making the package identifier resolvable
  is what first lets epubveri compare the NCX's dtb:uid against it, so the
  repair unmasked a pre-existing mismatch. No unit test could have caught that —
  the edit was correct and the book ended up worse. The dtb:uid now syncs in the
  same proposal, on the pattern fix.manifest_dangling_item already set.

  Measured: 10 findings across 10 books, 3 repaired and 7 declined — and the
  declines are the point. Four books carry both a UUID and an ISBN with no id on
  either, where which identity is canonical is an editorial decision; two carry no
  <dc:identifier> at all, where the repair would have to generate one. With the
  NCX syncs the run clears 8 findings and introduces nothing.

Changed

• fix.ncx_play_order now repairs all three playOrder faults, and no longer
  risks creating one. epubveri reports duplicate (different targets sharing a
  number), target_mismatch (one target reached by different numbers) and gap
  separately, and they interlock — satisfying one naively breaks another. The
  fixer now reassigns the whole NCX as the format defines: 1-based, dense, in
  document order, with elements naming the same target sharing the first number
  that target was given.

  The previous version numbered by position in the file — unique and dense, but
  target-blind, so on a book whose navigation reaches one position by two
  routes it would have created target_mismatch. No shelf book had that shape,
  so no audit could have shown it; the defect surfaced from reading epubveri's own
  rule, which skips a repeated number when all its holders name the same target.
  A corpus cannot find that class of defect; the detector's source can.

  The renumbering now parses the NCX rather than scanning it for
  playOrder=, since a target cannot be read off a string — so an NCX that will
  not parse is declined rather than rewritten, consistent with the other
  structural fixers.

  Measured: 8 target_mismatch and 1 gap cleared on top of the 14
  duplicate, nothing introduced, and one more book reaches fully valid.

• The body-level wrapper now also works inside <blockquote> — the single
  largest lever this project has found. XHTML 1.1 requires block content in
  <blockquote> exactly as it does in <body>, and the 10 EPUB 2 books added on
  2026-08-08 carry 2,508 stray-text and 3,009 incomplete-content findings there.

  The two messages are one defect seen from opposite ends, which is why both
  now trigger the fixer: a <blockquote> holding only text is stray text is not
  allowed directly in "blockquote" and element "blockquote" has incomplete
  content, because its model needs at least one block child. One wrap clears
  whichever fired — verified on real books before a line was written: one file
  went {incomplete 53, stray 54, span 1} → {}, another
  {incomplete 96, span 98, img 7} → {img 7}, the inline elements clearing as
  part of the run.

  The container set is now a principle rather than a name: the containers
  whose XHTML 1.1 model requires block content and admits <div> — body and
  blockquote. <ol>/<ul> want an <li> and <head> a <title>, wrappers
  that assert what the content is, so they stay declined. On the 125-book shelf
  that rule is not a compromise but the whole population: stray text is reported
  in exactly those two containers and nowhere else.

  Measured: findings cleared through schema_violation go 1,273 →
  8,624, total cleared 1,745 → 9,096, books touched 25 → 26, and both
  whole-shelf instruments still report nothing introduced anywhere.

  Honest ceiling: this does not clear all 3,009 incomplete-content findings,
  any more than it clears all 403 on <body> (it clears 2 of those). The rest
  are containers whose only children are <figure>/<section> — elements XHTML
  1.1 does not have — where the repair would be renaming, which epubsana
  deliberately does not do.

• Tracks epubveri 0.9.14. Zero source change; styloria 0.8 → 0.9 comes
  along transitively. Upstream's own figure verified from here: the only
  (id, rule) pair that moved is CSS-008 / css.stylesheet.invalid_selector,
  0 → 21 in one book, and the shelf run confirms it passes straight through —
  errors before 13087 → 13108 and after 3991 → 4012, the same +21 on both
  sides, because no fixer of ours consumes it. Everything else identical, zero
  regressions.

  Two upstream items worth recording even though they move nothing here:

  • RSC-010 gained its fallback clause, the defect diagnosed from this side
    on 2026-08-08 (a nav/NCX link to a non-Content-Document is legal when the
    manifest declares a fallback chain reaching one). No shelf book has the
    shape; the IDPF haruko-jpeg sample that prompted it loses its three errors.
  • epubcheck's JSON and XML reports cap identical messages at 25
    (CheckMessage.java, MAX_LOCATIONS), with the "N additional locations"
    line commented out — so a consumer sees a truncated list with no indication
    anything was dropped. epubsana runs no second oracle, so nothing here is
    affected, but any cross-tool count taken from epubcheck's JSON is
    incomparable above 25 per message. Recorded so it is never quoted by
    accident.

• Tracks epubveri 0.9.12 (from 0.9.9). Zero source change again; the shelf
  grew to 125 books the same day. Re-measured: errors 6376 → 4631 over
  the 25 books epubsana touches, 5 books reach fully valid, nothing
  introduced by either whole-shelf instrument. Every fixer's proposal count is
  unchanged, which is the answer to upstream's heads-up: 0.9.12 adds 171
  element "img" is missing a required attribute findings across 10 books, and
  none of them reaches a fixer of ours.

• handled_rules() now lists ncx.uid.package_identifier_mismatch and
  ocf.mimetype.not_first_entry. Both sites were rule-less when their fixers
  were written, so those fixers dispatch on the bare NCX-001 / PKG-006 id;
  epubveri 0.9.11 named them. The dispatch is unchanged — our floor is 0.9.7,
  where the slugs do not exist — but a census reading this list would otherwise
  have filed two rules we do fix under "no fixer at all", which is exactly the
  mislabelling the list exists to prevent.

Not built, deliberately

• <img> with no alt (171 findings / 10 books, the widest-spread shape on
  the shelf) is a decline, and upstream predicted it would be. XHTML 1.1 makes
  alt required where HTML5 does not, so this is one of the few places EPUB 2 is
  the stricter version. There is no determinate repair: alt="" is correct for a
  decorative image and wrong for a meaningful one, and nothing in the finding —
  or in the book — says which. Supplying either would be inventing content, which
  is the line empty_title already draws. Recorded in docs/COVERAGE.md rather
  than left to be re-derived.

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