v2.6.0

PDF Craft v2.6.0

PDF Craft 2.6.0 introduces shared retry and error-recovery policies, explicit
missing translations, and task-wide protection against reporting success when
all primary work fell back. It also corrects PDF preview font-size units.

Upgrade note: this release removes public arguments and changes default error
handling. Update callers using the English migration guide
or 中文迁移指南.

Retry and recovery policies

• RetryPolicy(translation=3, analysis=3, fill=3) supplies business retry budgets.
  Values mean additional attempts after the first; 0 allows one attempt. Analysis
  covers TOC/title inference, metadata JSON, footnote repair and JEV requests.
  Fill covers XML structure repair. Translation is reserved; no new semantic
  translation-quality loop is introduced. LLM transport retries remain separate.
• ErrorPolicy groups analysis, translation, XML fill, OCR, PDF page-rendering and
  PDF writeback failures. Each field accepts a bool or a synchronous/asynchronous
  decision callback. The execution's error_policy also accepts a bool, callback
  or None directly. Defaults are strict: final failures raise unless recovery
  is explicitly allowed.
• Both Craft façades and composable execution APIs accept policies. Per-call
  overrides do not mutate shared translators. Omitted error policies inherit the
  instance default; explicit None or False restores strict behavior.
• Removed arguments include component max_retries, JEV.retry_times, extraction
  ignore flags, PDF ignore_errors, ignore_translated_error, and
  on_fill_failed/FillFailedEvent. The migration guide explains replacements,
  including the old XML total-attempt versus new additional-retry distinction.

Explicit recovery and source locations

• Analysis can retain statistical TOC results, native PDF metadata, or the
  pre-refinement algorithmic result when permitted. Failed JEV evaluations are
  not low scores and do not automatically send every page to LLM repair.
• Translation or XML-fill failure leaves no accepted translation for the affected
  unit. PCEX coverage distinguishes missing translations from accepted ones;
  replacement rendering uses the source, and bilingual rendering shows it once.
  EPUB keeps failed source fragments while committing accepted fragments.
• Decision callbacks receive detached FailureContext values: structured error
  details, failed targets, proposed fallback and affected locations. Paragraphs
  retain all cross-page regions. EPUB retains archive/fragment identity and
  reliable CFI without requiring on_body; unavailable CFI remains None.
• Page furniture failures retain page indexes, source boxes, known dimensions and
  DPI, including streaming PDF translation and reusable templates. Invalid folio
  markers, identities, section counts and empty translated fields now obey the
  same error policy instead of silently recovering.

Task completion and PDF previews

• A task whose primary targets all failed and recovered raises AllFallbackError
  even with error_policy=True. Extraction covers local and vendor OCR; PDF
  writeback retains its page-level specialization. Successful metadata or EPUB
  headings cannot conceal a wholly untranslated body. Partial success, cache
  hits, valid unchanged translations and genuinely absent targets remain distinct.
• Cancellation, callback errors, fatal service failures and failed recovery still
  propagate. Staged publication protects existing final outputs on failure.
• PDF preview font sizes now use Qt logical pixels before conversion to source
  OCR coordinates, matching the exported baselines and line positions.

PCEX remains format v4. Content callback payloads and archive compatibility are
unchanged. Related work: #521,
#522, and
#523.

Full Changelog: v2.5.1...v2.6.0