mirror of
https://github.com/deusyu/translate-book.git
synced 2026-09-19 08:44:23 +08:00
Rainman
f0c3faaffa
* fix(glossary): include aliases in term_hash so alias edits trigger re-translation Aliases are injected into the per-chunk term table (别名 column), but term_hash only covered source/target/category. Adding or editing an alias on a term whose source already appears in a chunk changed the prompt contract without changing the recorded hash, so run_state.py plan reported the chunk as unchanged — contradicting the SKILL.md promise that alias edits re-translate affected chunks. The alias segment is appended to the hash payload only when aliases exist, so alias-less terms keep their historical hash and existing run_state.json records stay valid across the upgrade. Co-authored-by: Rainman <deusyu@users.noreply.github.com> * fix(merge): reject blank or unreadable translated chunks instead of silently dropping them A whitespace-only output_chunk*.md (non-zero bytes) passed the merge validation's size==0 check, produced only a 'Suspiciously short' warning, and then vanished during merge because merge_markdown_files skips content that strips to empty — the chunk's entire source text was silently missing from the final book while the build reported success. - manifest.validate_for_merge now errors on blank (whitespace-only) and undecodable outputs, via the new read_output_text helper - merge_markdown_files aborts on blank/unreadable outputs in both the manifest-ordered and legacy glob paths (read errors previously just printed and continued, which also dropped content silently) - run_state.plan marks blank/unreadable outputs for re-translation (reasons: blank_output / unreadable_output) and record refuses to certify them Co-authored-by: Rainman <deusyu@users.noreply.github.com> * fix(convert): abort when temp dir was built from different source bytes convert.py reused every cached artifact (input.html, input.md, chunks) whenever it existed, with no check that the input file itself was still the same book — replacing the source file with a same-named one silently translated the old book's chunks. A source_fingerprint.json (path/size/sha256) is now written into the temp dir on every successful conversion. On re-run: - fingerprint matches → resume as before - fingerprint differs → abort with a clear message (delete the temp dir or use a fresh --temp-root) - no fingerprint but cached artifacts exist (pre-upgrade temp dir) → adopt with a warning and fingerprint on the next successful run, so existing in-progress books stay resumable Only content identity is compared, so moving or renaming the source file does not invalidate the cache. Also removes the dead is_legacy return value from _find_existing_chunk_files — page* support was removed long ago. Co-authored-by: Rainman <deusyu@users.noreply.github.com> * chore(calibre_html_publish): remove dead logic and align style with the pipeline - drop the SIGALRM timeout layer — subprocess.run(timeout=...) already enforces the same limit, and signal-based alarms are redundant - remove the post-conversion image copy-back block: DOCX/EPUB/PDF embed images during conversion, and in pipeline use the copy target was the directory the images came from (rmtree + copytree round-trip) - book-producer metadata 'Claude Translator' → 'translate-book' - drop the unconditional '仿宋体 (FangSong)' font line printed even for non-Chinese output languages - normalize emoji log prefixes to plain text like the other scripts - delete the three skipped tests for extract_cover_from_epub — EPUB cover auto-extraction was declared out of scope (closed #3), so feature-gated tests for it are dead weight Co-authored-by: Rainman <deusyu@users.noreply.github.com> * docs+ci: document new validation guarantees; align CI Python with docs - SKILL.md: list source_fingerprint.json in the temp dir contents, explain the mismatch abort / adopt-with-warning behavior, and extend the Step 5 verification to blank (whitespace-only) outputs - README.md / README.zh-CN.md (synced): source fingerprint description in Step 1, blank-output rejection in the merge validation list, alias edits called out in the selective re-translation note, and two new troubleshooting rows - CI: Python 3.11 → 3.12, matching the documented 3.12+ environment Co-authored-by: Rainman <deusyu@users.noreply.github.com> --------- Co-authored-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rainman <deusyu@users.noreply.github.com>