Repository navigation
PDF Craft v2.6.7
Restore incremental PDF extraction on_body / on_footnote delivery after the
2.6.6 regression. Closed page-processed paragraphs and footnotes arrive while OCR and
later analysis are still running, without a duplicate replay at completion.
Page refinement now runs as a bounded pipeline: JEV starts when its initial
neighbor inputs are ready; the target LLM follows only that page's score. Neighbor
scores and refined neighbor content never enter its prompt. Ordered assembly
emits closed processed paragraphs/footnotes without waiting for whole-book OCR,
other JEV jobs or final TOC. This supersedes the temporary pre-refinement preview.
The initial paragraph boundary survives intervening image/table-only pages without
retaining those pages in memory. Cancellation also interrupts a full refinement
window, then settles its page jobs before releasing providers and workers.
OCR completions now enter refinement by neighborhood readiness rather than a
contiguous page prefix. An unrelated earlier slow OCR page no longer blocks a
ready JEV/LLM job. Cross-window identities remain resolved at ordered assembly;
finished later jobs free request slots while their results wait on disk.
Callbacks describe page-processed content; final TOC/heading levels and directory
exclusion may organize the PCEX differently. Final TOC, chapter and furniture consistency checks remain in place. See streaming callbacks
and 中文说明.
Ordered delivery, source identities, disk buffering, callback errors and cancellation
remain supported. Barrier regressions guard timely delivery independently of unrelated pages,
whole-book scoring and final packaging. The progress event schema and PCEX format are unchanged; PAGE completion now reports local page-work completion rather than waiting for publication.
EPUB accepted atomic fragments now notify before unrelated chapter groups finish.
Word progress and content delivery remain independent. PDF layout no longer waits
for document metadata or all furniture: shared templates and page fragments feed
individual page dependencies; body/note previews keep separate ordered channels.
Final metadata and PDF page ordering are joined at publication. Necessary whole
paragraph and page-level typography constraints remain unchanged.
Independent PDF pages isolate preceding-body font state: heading sizes and preview geometry no longer depend on translation completion order. Pages without a body reference use the configured headline fallback.
See the streaming design rationale (Chinese)
for the product latency contract, immutable inputs, backpressure and the deliberate
tradeoff of omitting neighbor JEV scores.
PDF layout now fits each ready paragraph once and finalizes pages from their direct
inputs, rather than waiting for a transitive chain of cross-page paragraphs.
Cross-page previews accumulate actual final regions on disk and publish only when
the paragraph's own regions are complete. Page typography remains independent of
arrival order; paragraph fitting and page normalization retain separate boundaries.
Related changes: #564, #565, #566, #567.
Full Changelog: v2.6.6...v2.6.7