All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

• (config): ConcurrencyConfig::max_concurrent_ocr and the --max-concurrent-ocr CLI flag set concurrent Tesseract recognition sessions on their own. Use it when the host has cores to spare but not the memory to run a recognition session on each of them. The value is applied as given: neither the thread budget nor the host's free memory reduces it, since both of those bound only the automatic limit. The first extraction in a process fixes the session count for the rest of that process, because the admission semaphore and the Tesseract handle pool that enforce it are built once and the pool's capacity is fixed when it is constructed; a later extraction that names a different value keeps the first one and logs a one-time WARN naming both numbers. Set the value on the first extraction, or run one process per value. ConcurrencyConfig is not #[non_exhaustive], so the added field breaks any Rust caller that builds the struct by literal without ..Default::default(); such a caller must add the field or the rest pattern. Callers on every other binding are unaffected. (GH#1727)

• (pdf): scan detection and fabricated-mapping OCR routing now use the thread budget. Native PDF extraction graded every page for raster scan evidence, then read every page's text again to check its glyph-to-Unicode mapping provenance, and ran both passes one page at a time. The two passes held a single core for the whole document whatever ConcurrencyConfig::max_threads was set to, so no thread budget could shorten them. Both now run across the thread budget, sequentially on wasm32 which has no thread pool, and report the same per-page confidences and the same page lists in the same page order. Measured on a 32-core machine at max_threads = 32, with the extracted output identical in every run: a 4778-page manual from 66.9 s to 63.3 s, and a 676-page book from 24.1 s to 21.9 s. The two passes are a small share of a native extraction, so the wall-clock gain is bounded by that share. (GH#1723)
• (ocr): concurrent Tesseract recognition follows the thread budget instead of a fixed four. The limit was a compile-time constant that no configuration reached, so raising max_threads could not raise recognition throughput, and recognition is most of the run on a scanned document. The default is now the thread budget, reduced to the number of sessions the host's free memory holds. Set max_concurrent_ocr to 4 to keep the previous behaviour. (GH#1727)
• (pdf/ocr): the per-page OCR route sizes its batch by what a page costs to hold, not by the limit on returned content size. max_content_size bounds the text an extraction returns and says nothing about the rasters a batch holds while it works, so on an ordinary 150 DPI document the arithmetic pinned the batch at four pages and no stage of the route could use more than four threads whatever the thread budget said. The width now comes from the page box and the render resolution, measured against free memory less the document and a reserve. Each page's own peak is still checked against max_content_size, one page at a time. Peak resident memory rises with the wider batch -- 2,234 MB to 3,583 MB on a 731-page document at a 32-thread budget -- because more rasters are in flight. (GH#1724)
• (pdf): the embedded-image pass now runs across the thread budget. Configuring OCR switches whole-document embedded-image extraction on, and that pass walked the document one page at a time, so no thread budget shortened it. Pages now decode in parallel (sequentially on wasm32, which has no thread pool), in the same order, and the returned bytes are unchanged. Measured on a 731-page document holding 2,881 images at a budget of 32: the pass falls from 20.1 s to 2.2 s, and the whole extraction from 57.3 s to 41.3 s. Peak resident memory rises with the budget, because each pool thread holds one page's raw pixel buffers and their PNG re-encodes at once, where a single page's were live before: over the pass itself, 1,133 MB before against 1,463 MB at a budget of 8 and 2,186 MB at 32. The pass carries no memory budget of its own, so the ceiling is the thread budget, which is min(cpu_cores, 8) until ConcurrencyConfig::max_threads raises it. The pass also no longer decodes and PNG re-encodes a full-page image on a page that already carries native text when OCR is the only reason the pass is running at all: should_skip_pdf_image_ocr excludes exactly that image from OCR and drop_ocr_only_images then drops it from the result unread, so the decode was pure waste. That image still appears in the result with its dimensions, bounding box, page number and alt text intact and an empty data; a document where any other consumer -- image extraction, captioning, QR codes, inline-image OCR, page rasters, or styled HTML rendering -- would actually read the bytes is unaffected. (GH#1732)
• (ocr): candle-deepseek-ocr reads device tensors back in one transfer instead of one element at a time. The mixture-of-experts gate copied its score matrix per expert, per row, per layer, per generated token, and the SAM relative-position lookup read its index inside a nested loop per attention layer, per crop. Both are now a single bulk read. Measured on an A100 at BF16: 62.71 s to 8.09 s per page (7.8x), output byte-identical, peak memory unchanged. A 22-page scanned document that previously exceeded the 600 s extraction ceiling now completes in 322 s. (GH#1711, GH#1714)
• (ocr): candle-deepseek-ocr resizes the relative-position table and scatters the image embeddings as device operations instead of per-element loops. The resize sampled its axis one position at a time and concatenated one tensor per output position, on each of the four global-attention layers of the batched 640 px local-crop pass; it is now candle's bilinear resize over a one-row image, which is the half-pixel sampling the reference implementation uses. The scatter gathered rows one at a time and stacked one tensor per sequence position, once per page; it is now two gathers and a select. Measured on an NVIDIA L4 at BF16 over a full Letter page, five repetitions: 47.610 s to 47.546 s per page, which sits inside the spread of the unchanged arm, so this is not a speedup on that workload. Both arms decode the same 2,052 characters. Peak GPU memory falls from 8,818 MiB to 8,722 MiB. (GH#1719)
• (ocr): the per-page OCR-image render/encode size check now runs through one shared helper instead of three separate copies. GH#1724 and GH#1731 each fixed a route that summed a whole page batch against security_limits.max_content_size -- a limit its own error text names as a per-image bound -- rejecting every page in the batch once the sum crossed it. Both fixes, and the mixed native/OCR route's own single-backend path, now call one validate_png_encode_pages_individually helper that charges each page on its own, so a new call site cannot reintroduce the batch-summing shape by calling the lower-level batch-peak function directly. No behaviour change on any of the three existing routes. (GH#1748)
• (pdf): a /Font dictionary shared across pages no longer redoes TrueType cmap donation on every page that touches it. share_truetype_cmaps clones and mutates each undonated font with Arc::make_mut, and the font-set caches deliberately store each dictionary's pre-donation fonts (#1725 made that order-independent), so every cache hit re-ran that clone for the same font. Donation among fonts within one dictionary's own resolved set is now cached alongside the pre-donation set, under the same key, so it depends only on that dictionary's own resources and not on which page or Form XObject was read before it. Donation whose donor lives in a different dictionary (a Form XObject donating to its page, say) is unaffected and still runs on every load_fonts call. (GH#1746)
• (pdf): fabricated-mapping OCR routing no longer reads every page's text a second time. The main text pass and the provenance pass each read every page of a native PDF: the main pass to assemble content, the provenance pass (GH#1723) to grade each page's glyph-to-Unicode mapping. Neither carried its read into the other, so a 731-page document spent 14.2 s of a 24.6 s extraction on the two passes together. The main pass now keeps each page's fabricated-character counts from the raw spans it already reads, and the provenance pass consumes those counts instead of reading the page again; a document with default-off optional-content layers still reads separately, since the main pass reads layer-filtered spans there and the two would otherwise disagree. Retaining the counts holds page text already held for content assembly, not rasters, so peak memory is unaffected. fabricated_text_pages, scan confidences and extracted content are unchanged. (GH#1744)
• (pdf): scan detection no longer decodes an embedded image's pixels just to classify its codec. Grading a page for raster scan evidence decoded every embedded image on the page in full -- the same decode the embedded-image extraction pass repeats moments later when OCR is configured -- only to read whether the result was JPEG or CCITT and how large its bounding box was. Both are already known from the cheap Phase 1 handle enumeration that walks the content stream without decompressing anything, so scan detection now reads bounding box and filter chain from there instead. The embedded-image extraction pass, the only remaining caller that needs decoded pixels, is unaffected; classification output (image area ratio, codec class) is unchanged for every image that decodes; an image the decoder rejects now counts toward the page's image coverage where it was silently dropped before, and images under the 8 x 8 px extraction floor stay excluded. (GH#1732)

• (pdf): layout detection no longer fails partway through a long document. The layout pass charged every page raster it had already produced against security_limits.max_content_size, so the running total crossed the 100 MiB default after about twelve standard pages, whatever the page size. Layout detection then failed for the whole document while extraction carried on, and the caller got a successful result with no layout hints and a warning naming the page's pixel dimensions, which reads as a page-size limit and is not one. Each batch is now charged on its own, so the limit bounds the work in flight rather than the length of the document. The layout pass still retains every chunk's raster, and security_limits.max_content_size no longer bounds that retained set; only security_limits.max_pages does, and it is unset by default. (GH#1721)
• (ocr): the auto device preference reports when it cannot reach the accelerator, instead of running the whole job on the CPU in silence. A failed CUDA or Metal init discarded its error and returned the CPU device with no log line naming the cause, so the run looked like a hang: the GPU stayed at 0%, several cores were busy in the forward pass, and the last log line was whatever ran before it. A warning now names the accelerator and the underlying error. GLM-OCR, PaddleOCR-VL and TrOCR already resolved the device once per engine, inside the engine pool's cold start; candle-deepseek-ocr now does the same, instead of resolving -- and warning -- once per page. A probe that fails is retried on the next page rather than remembered, so a transient accelerator failure does not pin the process to the CPU for the rest of its life. (GH#1722)
• (pdf): a row-padded image buffer is rejected instead of panicking. ImageBuffer::from_raw rejects a buffer that is too small and accepts one that is too large, keeping the extra bytes. A decoder that pads each scanline to an alignment boundary produces exactly that, and the mismatch surfaced inside the PNG encoder as its own size assertion, so extraction panicked instead of returning an error the caller could act on. Measured on a 229x265 RGB image whose buffer held 182,320 bytes against the 182,055 the image needs: one byte of padding per row. The exact length is now checked before construction for all three pixel formats, returning the same recoverable error an undersized buffer already returned. to_png_bytes in xberg-native-pdf already carried this guard; it was never applied to this call site. (GH#1735)
• (pdf): a page's extracted text no longer depends on the order pages are read. Reading a document's pages concurrently could return different text on different runs. Sequentially, a page whose /Font dictionary is written inline under the same resource names as an earlier page's decoded through that earlier page's fonts, so where the two fonts differed the later page read wrong. Extraction now returns the same text every time, whatever order the pages are read in, and each page decodes through the fonts its own resources name. (GH#1725)
• (ocr): OCR with layout detection no longer drops every page of a batch. The route that OCRs pages a layout pass has already rendered charged the whole batch's render-and-encode peak against security_limits.max_content_size, a ceiling that bounds one image. Batch width on that route is the resolved thread budget, so from five US Letter pages up at the default 150 dpi the sum crossed the 100 MB default and the OCR run was refused outright; the automatic OCR fallback then returned the document's native text with a warning, which reads as a successful extraction that is missing its OCR'd text. Each page is now charged on its own, as the limit's name says. The batch's own footprint is bounded by the thread budget alone and not by max_content_size, so peak live bytes on this route rise with a wider budget. (GH#1731)
• (pipeline): embedded images populated without OCR are no longer dropped from images. The GH#1703 fix (1.2.6) dropped every embedded image's bytes after extraction unless the caller had asked for image extraction, captioning, QR codes, inline-image OCR or page rasters. It ran on every extraction, including ones with no OCR configured at all, so a Markdown inline SVG data URI, a Jupyter output or attachment image, or an ODT/DOCX/PPTX embedded picture came back with images empty even though nothing had read those bytes for OCR. The drop now runs only when OCR was configured to run on embedded images, which is the case GH#1703 addressed. (GH#1703)
• (ocr): candle-glm-ocr resolves its layout model once per process, not once per page. Each page re-ran the Hugging Face path resolution and checksum verification of the 131 MB PP-DocLayout-V3 model; that is now cached by directory, the same shape the layout engine's sibling models use. Measured on 22 pages: 82.71 s to 6.24 s. A failed resolution is not cached, so the next caller retries. (GH#1718)
• (pdf): a document whose text-plausibility check could not judge any page now says so. implausible_text_pages: [] meant either that every page was checked and passed or that no page held enough prose to be checked at all (contracts, invoices, forms, agenda packets), and a caller could not tell the two apart. A document where no page could be judged now carries a processing warning naming how many pages were examined; the warning is suppressed when OCR was already forced. Extraction behaviour and OCR routing are unchanged. (GH#1709)
• (pdf): a numbered heading that wraps twice is no longer closed after its second line. The paragraph grouper's heading-wrap exemption only recognised a heading that had absorbed exactly one wrap (visual_line_count(&current_lines) == 2), so a heading spanning three or more visual lines was cut before its own last line: the orphaned line was then welded to the body paragraph beneath it, with no signal left to separate the two on a two-column page. The exemption no longer counts wraps; it now applies at every line while the paragraph is still nothing but the heading and its accepted continuations, so a heading is closed at the first line that genuinely fails to continue it, however many wraps came before. (GH#1740)
• (pdf): a numbered heading set as one TJ array with a kern for the tab no longer absorbs the other column's line at the top of a two-column page. When a producer places a heading's marker and title in one TJ array with a kern standing in for the tab ([(3.)-1329.5(Title )] TJ), the kern becomes a space-only span that is always regular weight regardless of the surrounding bold context, which broke the reading-order heading-run detector's clustering right at the marker/title boundary. The narrower run this produced no longer covered the marker's column position, letting an unrelated span from the other column land inside the heading. The same page set with the marker in its own text object (no kern) already read correctly; it now reads the same way regardless of how the producer set the marker. (GH#1738)
• (config): the configured thread budget is no longer silently dropped when two extractions start together. init_thread_pools built the process-wide Rayon pool outside the call_once fence that guards it, so a second concurrent caller could be released -- with the atomics already set -- before the pool existed. If that caller reached its own parallel work before the first caller's build_global() finished, it silently installed Rayon's default pool first, and the configured max_threads was never applied for the life of the process. The pool now builds inside the same fence, so no caller observes the installed limits before the pool they describe actually exists. (GH#1750)
• (pdf): the dense two-column repair no longer lets a table's own grid vote for the page's gutter. On a two-column page that also carries a table, detect_split_x counted a table row's internal cell gaps as gutter evidence -- a 5-column table's rows outvoted the page's real two-column lines and placed the split inside a column, and the same votes fed the hanging-label snap, so a numeric table column straddling the true gutter could be mistaken for a stack of hanging clause numbers and pull an already-correct split back into the table. Both now skip a line whose inked spans open four or more internal gaps, the shape of a table row of five or more columns; a two-column line with a hanging number on each margin opens three, so three would exclude the very lines that carry the gutter. Separately, both_sides_are_columns (gating the split's widened-corridor rescue) required both sides of a candidate gutter to classify as prose, stricter than the per-band reorder gate it feeds, which already accepts one non-prose (table) side when the two sides do not pair up row for row; it now applies the same test. A third gap remained even with both of those in place: corridor_is_hanging_label_indent read a table's own narrow edge column (cells stacked hard against the gutter's left wall) as a hanging clause number, so the widened corridor rescue still refused the page's real gutter and left the split inside the table. It now also requires an inked span on the far side of the corridor, on the same line, before counting a candidate -- a hanging label always has its clause text there; a table's edge cell never does, since the row's other cells sit on the label side and whatever text starts past the corridor belongs to an unrelated, unpaired line. Fixes the reporter's own page 1 (a five-column table with a narrow last column, plus a centred page number in the gutter); a table filling a whole column with no row-pairing across the gutter (the reporter's page 4) is fixed too. There the wrong split is crossed by only one line -- an introductory paragraph above the table, not the table itself -- because every table row's own evidence that the split sits inside it is its internal cell gap, never a span crossing the split, so the count that decides whether to widen the corridor search never reached its threshold and the search that would have found the true gutter never ran. That count now also includes a line whose split evidence is its own internal, multi-column cell gap rather than a literal crossing, so the search runs and the already-fixed corridor guards find the true gutter the same way they do on page 1. (GH#1742)

• (pdf/ocr): OcrQualityThresholds::enable_plausibility_ocr_routing and OcrQualityThresholds::min_reliable_language_chunk_ratio control language/dictionary-plausibility OCR routing, and PdfMetadata.implausible_text_pages reports which pages a page's decoded text failed to read as any real language. (GH#1696)
• (config): ExtractionConfig::runs_ocr_on_embedded_images and ExtractionConfig::wants_own_bytes_in_result are exposed on every binding, beside the existing needs_image_data. The first is the predicate the pipeline uses to decide whether a container's embedded images are OCR'd; the second is the pre-GH#1662 formula (extract_images, captioning or QR codes) that decides whether a standalone image's own bytes are echoed into images. (GH#1662)

• (ocr): the ocr, ocr-wasm, and ocr-pipeline Cargo features now imply language-detection. The language/dictionary-plausibility OCR-routing signal (GH#1696) is gated the same as the rest of the OCR module and would otherwise be a silent hole under a build that enables only ocr-pipeline (the VLM-only, Tesseract-free pipeline). whatlang (behind language-detection) depends only on hashbrown, so this is wasm/Android-safe. (GH#1696)

• (ocr): OCR no longer returns every embedded image's raw bytes. Image bytes read only to feed embedded-image OCR (GH#1662) are dropped once OCR has consumed them and any OCR text it produced has been rendered into content; images now carries data only when the caller asked for image extraction, captioning, QR-code detection, pdf_options.ocr_inline_images, or images.include_page_rasters; counts.images still reports how many embedded images the document had. (GH#1703)
• (pdf): a /ToUnicode mapping that resolves to the wrong-but-real letters is now routed to OCR too. GH#1667 fixed the case where a page has no usable mapping tier at all (MappingProvenance::Fallback); a page whose /ToUnicode CMap DOES resolve, but consistently to the wrong letter (e.g. a ROT-shifted mapping), passed every character-shape check unchanged and OCR never fired, because the text is structurally indistinguishable from real prose. A language/dictionary-plausibility signal now flags such a page under the default Auto strategy, and the opt-in ScannedPages strategy picks it up through the existing scanned_pages union. A page with no explicit ocr config keeps its native text, but a warning now names the defect. Closes the remaining gap in GH#1667. (GH#1696)
• (pdf): the mixed native-and-OCR path no longer drops every native table on a page OCR never touched. On a long document where only some pages needed OCR, the mixed path replaced the whole document's table list with just the OCR pages' tables whenever OCR found even one, silently dropping every native table on every other page. Only the tables of pages OCR itself produced a table for are now replaced; a page the mixed path never sent to OCR keeps its native tables. (GH#1670)
• (pdf): include_document_structure alone now triggers the structured native extraction pass. Previously the structured pass ran only for a Markdown/Djot/HTML/DocTags output format, an explicit hierarchy config, inline-image OCR, or a content filter -- never for include_document_structure itself. A caller who set only that flag, with output left at its Plain default, silently got a flat, paragraph-only document, so the structure tree it asked for came back holding nothing but paragraph nodes even though the extractor found headings, tables and page breaks. This configuration now renders its plain-text output from the structured document, so the rendered content changes with it: table rows render one per line, and headings and page footers reflow. A caller that sets only this flag and depends on the previous Plain text sees that text change. The table count is unaffected. (GH#1668)
• (docx): a paragraph inherits the list numbering that its style carries. A list whose paragraphs hold no direct numbering and take it from the paragraph style instead, which is the shape Word writes for its built-in list styles, came out as plain paragraphs with no list markers. The parser now resolves numbering through the style's basedOn chain. Direct paragraph numbering still takes precedence over the style. (GH#1663)
• (pdf): the fabricated-glyph-mapping provenance lookup is repaired, and a page it flags is sent to OCR. The extractor read each text run's mapping provenance after it had renamed the run's font to the resolved /BaseFont, but the provenance table is keyed by the raw resource alias, so the lookup missed for every font whose base name differs from its alias and the signal was empty almost everywhere. The lookup now runs in the same pass and uses the alias, and a page whose text carries no usable mapping tier at all (MappingProvenance::Fallback) now forces the OCR fallback under the default Auto strategy. A font whose mapping resolves, even to the wrong characters, is a different failure mode this signal does not cover. (GH#1667)
• (ocr): a low OCR confidence now lowers the reported quality score. The quality score measured only how clean the text looked, so a page that OCR itself had little confidence in still scored as clean whenever the characters happened to be well formed. The score is now capped by the word-count-weighted mean of per-page OCR confidence, once at least 20 words were recognized -- the same fold ExtractionConfidence::ocr_aggregate uses, shared rather than recomputed, though the aggregate itself reports on any recognized word and so can differ below the 20-word floor. A native, non-OCR extraction is unaffected. (GH#1669, GH#1694)
• (ocr): a cold start no longer builds one candle OCR engine per waiting caller. Each candle backend's engine pool looked for a cached engine, released the lock, built the engine, then stored it, so every request that arrived during the first build loaded its own copy of the model and a burst could exhaust the GPU memory. All four backends and the GLM-OCR layout pool now use the crate's engine cache, one per backend, which holds the write lock across the build: the first caller builds and the rest wait for that engine. Loads of different engines on the same backend (two TrOCR variants, say) therefore also serialise during a cold start. A failed load stores nothing, so the next caller retries. (GH#1683)
• (ocr): the aggregate OCR confidence is populated for page-level results, and both routes weight it the same way. A document whose OCR confidence arrived one score per page left the aggregate empty, so a caller that read it got nothing back. Both routes now fold the scores by recognized word count instead of by element, so a page of many short lines no longer counts for as much as a page of few long ones. This changes the number the embedded-image route already reported. (GH#1677)
• (ocr): the Tesseract language probe run at backend registration now finds the installed languages. The probe initialised Tesseract with an empty datapath, which makes the library look one tessdata directory deeper than the resolver real OCR jobs use, so on a layout where jobs found eng.traineddata directly the probe logged "couldn't load any languages" and silently fell back to a hardcoded language list. It now resolves the tessdata directory the same way a job does. (GH#1671)
• (ocr): a page whose embedded-image OCR retry came back empty says so. When page-raster OCR failed and the embedded-image retry ran but returned nothing -- a candle VLM backend reports success with empty content -- the page warning named only the first failure, with no trace that a retry ran at all. The per-page warning and the wholesale "every page failed" error now report the retry outcome alongside the first failure. (GH#1673)
• (ocr): OCR throughput scales with the thread budget. The mixed native-and-OCR route rasterised each batch's pages one at a time regardless of the configured thread budget, so a 240-page document with 80 scanned pages took the same 9 s at 4, 8, 16 and 32 threads. Pages now render in parallel across the batch (sequentially on wasm32, which has no thread pool), in the same order as before; measured 8.8 s to 3.3 s at 4 threads and 9.0 s to 1.0 s at 32. (GH#1666)
• (ocr): a DOCX, PPT, PPTX or HTML embedded image is read before OCR runs on it. The predicate that decides whether a container reads an image out of its archive did not count embedded-image OCR, so with OCR configured and run_ocr_on_images left on, the image was attached with an empty buffer and OCR reported "Could not determine image format" for zero bytes. A standalone image extraction's images output is unchanged. (GH#1662)
• (ocr): the mixed native-and-OCR route sizes its render batch against security_limits.max_content_size. The batch was sized from the resolved thread budget alone, so a wide thread budget requested a batch whose estimated PNG-encode peak crossed the fixed content limit, and every page in the rejected batch was skipped rather than the extraction failing. The batch is now capped up front from one representative page's estimated encode cost at the effective render DPI; the real peak is still checked, and still rejected if genuinely too large, afterwards. (GH#1665)
• (ocr): the candle OCR backends no longer reject every rendered PDF page. The PDF OCR route stamps source_dpi and page_rotation_degrees into the shared backend_options for every backend, and the candle backends (candle-paddleocr-vl, candle-glm-ocr, candle-deepseek-ocr) deserialise their options strictly, so every rendered page failed validation and came back empty. Exactly those two pipeline hint keys are stripped before the candle options are read; any other unknown key still fails validation. (GH#1672)
• (ruby): a tagged enum's payload readers return the stored value. The generated Data readers called super on a class that has no such method, so reading a payload field on a tagged-enum value raised NoMethodError. Regenerated on alef 0.93.1, whose Magnus emitter reads the stored member. (alef 0.93.0)
• (ocr): a candle VLM backend no longer emits implausible lines of noise. candle-paddleocr-vl and the other three candle backends would occasionally emit a CJK-script run in an otherwise English document, or bare LaTeX markup on a plain-text OCR task, most visibly on a non-text region such as a handwritten signature scribble. A line whose classified letters are mostly in a script the configured OCR language list does not cover, or whose content is mostly \command-shaped LaTeX tokens with no bare word on a plain-text task, is now dropped, with a warning naming how many lines were removed; a $$/``` fenced block and a genuine formula/table/chart task are exempt. (GH#1676)
• (ocr): candle-deepseek-ocr reads the whole page. Generation was capped at 128 new tokens (~400-650 characters, the 15-25% observed), the page was fed as a single 1024 px global view, and weights were loaded as F32 (16.5 GB on a 24 GB card). The cap is now a model configuration field defaulting to 4096 in line with the other full-page backend, pages larger than 640 px are also tiled into 640 px local crops as in the reference "Gundam" mode, the dtype follows the device (BF16 on CUDA, F16 on Metal, F32 on CPU), the weights auto-download from a checksum-pinned deepseek-ai/DeepSeek-OCR revision so model_path is optional, and a run that starts looping (the repeated-digit table on a signature block) is detected once the repeat fills a trailing detection window and truncated back to a single copy of the repeated unit, rather than burning the rest of the token budget on it. (GH#1674)
• (ocr): candle-glm-ocr no longer garbles hex identifiers such as UUIDs. The repetition penalty defaulted to 1.1 with no upstream basis (zai-org/GLM-OCR's generation_config.json sets none) and was applied once per occurrence of a token in the whole decode history, so a token repeated k times was suppressed by 1.1^k -- exponential enough that hex digits and - inside a long identifier lost probability mass until argmax drifted onto a re-emitted group. The default is now 1.0 (a no-op, matching upstream), and each token in the trailing 64-token window is now penalised exactly once regardless of how many times it recurs. (GH#1675)
• (ocr): candle-deepseek-ocr no longer runs its prefill without a causal mask. prepare_causal_attention_mask built a 0/1 keep-mask that the attention step added straight onto the logits, so a future token was nudged down by one instead of blocked and the prompt-and-image prefill (and the KV cache it seeds for every later token) saw the whole sequence; the mask is now the additive 0/-inf form GLM-OCR and PaddleOCR-VL use. This is the one DeepSeek-only forward-pass defect the GH#1701 audit found; with it fixed a full Letter page decodes without repetition on Metal at both F16 and F32, but the reported CUDA F32 loop has not been re-run. (GH#1701)
• (ocr): candle-deepseek-ocr no longer fails on the first 640 px local crop. Resizing the SAM relative-position table to the crop size multiplied each row by a one-element tensor, which candle does not broadcast, so every page large enough to be tiled failed with shape mismatch in mul before a token was decoded; the resize now scales with affine and samples at half-pixel positions like the reference F.interpolate(mode="linear"). Verified on Metal at F16: a 2,018-character Letter page comes back whole. (GH#1674)

1.2.4 was tagged but never reached a package registry: its generated binding files still named 1.2.3, so its publish run was cancelled. 1.2.5 is the first published release carrying the 1.2.4 changes below in addition to its own.

• (python): a config field holding a nested data-carrying enum is no longer dropped by from_json / to_json. The Python DTOs marked every field whose type transitively held a data enum as #[serde(skip)] -- CaptioningConfig.llm and ChunkingConfig's nested options among them -- so a JSON payload round-tripped through such a config silently lost those fields. Regenerated on alef 0.92.1, whose Python emitter no longer treats a serializable enum wrapper as opaque. (alef#394)
• (ppt): a legacy .ppt's embedded OLE objects are extracted. A Word document or Excel sheet inserted as an object -- the way PowerPoint 97-2003 decks routinely carry a table -- lives in the deck's ExOleObjStg records, which the extractor walked over as opaque bytes, so the slide came out as its title and nothing else. The .pptx path already recursed into ppt/embeddings/; the legacy path now does the same: each embedded object's storage is resolved through the persist chain (so a superseded save's copy is never read), decompressed when the record says so, identified by its root stream (WordDocument, Workbook/Book, PowerPoint Document, or an OPC Package) and extracted into children as slide<N>/oleObject<id>.bin with the slide whose shape displays it. The decompressed size is bounded by max_embedded_file_bytes (falling back to max_archive_size), the object count by max_files_in_archive, and max_archive_depth = 0 disables the recursion, as for .pptx. Metafile presentation pictures (EMF/WMF) are still not emitted as images. (GH#1660)
• (pdf): a ruled table drawn one bar per cell no longer loses all but its last two rows. Word draws cell borders as per-cell filled bars with a corner square at every crossing, so every vertical rule terminates at every horizontal rule. The section-divider split, which reads a full-width rule that no vertical rule runs through as the boundary between two stacked form sections, therefore treated every internal rule of such a table as a divider, cut the table into one-row pieces and discarded the pieces below the cell minimum -- a 12-row table came back as 11 | 290, 12 | 300. A rule that every vertical rule stops at carries no section information when the same is true of every other rule on the table, so a table whose internal rules all qualify is now left whole; a form whose sections' rules stop at a minority of the rules still splits. (GH#1656)

• (llm): LlmConfig exposes liter-llm's provider response-byte cap. liter-llm's ClientConfigBuilder::max_response_bytes bounds every HTTP response body read from a provider, but xberg's LlmConfig never forwarded it, so downstream products had no way to bound provider responses. LlmConfig::max_response_bytes mirrors it: it bounds response bodies on every non-streaming call and the error body read on a failed request (a successful streaming response keeps its own existing frame bounds), and defaults to None (unbounded), matching liter-llm. Some(0) is rejected by LlmConfig::validate rather than reaching liter-llm's own builder. (xberg-io/xberg-enterprise#1568, xberg-io/xberg-enterprise#1861)

• (pdf): a page-sized background rectangle no longer seeds a whole-page table. Many Office-to-PDF producers draw a white rectangle over the entire page before anything else. That rectangle passed the table-primitive filter (its < 1000 pt bound does not exclude an A4 or Letter page) and, because clustering unions any two primitives whose boxes intersect, pulled every rule on the page -- the real table's borders, the footer rule, a figure's frame, strokes inside a drawing -- into one cluster with the page as its box. The cluster fallback then built a table over the whole page: the heading above the real table became its first row, torn at the table's own column rule, and the first words of the prose below it became its last row, with those words missing from or doubled in the paragraphs that followed. A rectangle covering at least 90 % of the page's MediaBox in both dimensions is now treated as page furniture and dropped before clustering, so the real table's rules cluster on their own and the page comes out as it does without the background. A full-page-width rule (one dimension at page scale) is deliberately still a primitive. (GH#1656)
• (pdf): a single-column page of hanging-number headings is no longer read as two columns. Some producers (Distiller among them) emit the tab between a heading's number and its title as a space-only span in a different font, and leave space-only spans for blank lines. The dense-two-column repair counted those spans as column population and let a blank line's two whitespace spans and a footer split across both margins vote for a gutter, so five headings' tabs plus two lines of nothing met the six-line quorum and the page was emitted column-major: 6, 6.1, 6.1.1 as bare numbers and their titles as separate unnumbered headings. Whitespace-only spans no longer count toward the per-side density gate, the gutter vote, the whitespace corridors or the row-pairing guard, and a per-line gap wider than a quarter of the page width is no longer accepted as gutter evidence (a blank line's or a footer's gap; every real gutter measured here is under 10 %). (GH#1655)
• (ocr): a configured security_limits now reaches the Tesseract image decode. GH#1554 routed ExtractionConfig::security_limits onto OcrConfig before each OCR call, but the Tesseract backend's config_to_tesseract had no field to carry it into and built a fresh default ExtractionConfig for the processor, so every Tesseract decode -- standalone images, embedded images, scanned pages and the targeted page fallback alike -- ran under the 100 MiB default no matter what the caller set. A 600 DPI A4 scan was refused with an error naming 104857600 bytes while both configuration objects said 5 GiB. The internal TesseractConfig now carries the limits (outside the cache key: they gate whether a decode runs, never what it produces), the processor prefers them over its synthetic config, and the standalone-image and PDF embedded-image routes inject the caller's limits like the other routes already did. A limit set directly on OcrConfig is honoured when ExtractionConfig carries none; ExtractionConfig still wins when both are set. (GH#1651)
• (go): Register* no longer leaks its cgo.Handle when the C vtable allocation fails. Every Register* wrapper in the Go binding created the handle before allocating the C vtable, and the allocation-failure branch returned without deleting it, keeping the bridge and the caller's implementation reachable for the life of the process. The branch is only reachable when a small malloc fails, so this is a resource leak on an error path rather than an exploitable defect; it is fixed in the generator (alef 0.91.6) and regenerated here. Reported by @OvOhao (GHSA-q5pq-8g86-v9j9).
• a table's multi-word cell no longer bleeds a trailing word into the next column, a right-aligned amount column split by digit-width drift is folded back together, and a legitimately sparse but independently-headed column (or a sparse first data row) no longer gets the whole table rejected. Reconstructing a table from word boxes decided each word's column independently by nearest x-position, so a wide cell's own trailing word (e.g. a long description's last word) could resolve to a neighboring column's anchor instead of its own cell's. Column membership for a data row is now decided once per merged cell cluster and applied to every word the cluster contains; a drift-split right-aligned numeric column (mutually exclusive per row, no independent header of its own) is folded into one column; and a section caption sharing a table region with its real header row is dropped. This cell-clustering and column-assignment logic lives in the shared table_core module, so it also affects native-PDF word-position table reconstruction, not just OCR. Separately, pdf's shared table post-processor no longer treats an independently-headed but infrequently-populated column (e.g. a bank statement's DEPOSIT, populated on a minority of transaction rows) as noise, and no longer folds a sparse first data row (one missing an optional numeric field) into a bogus multi-row header merge with the row after it -- and that header-shortcut no longer stops one row early on a genuine two-row text header, which also fully populates a digit-free first row. The punctuation cell-merge gap this widens for a lone dash/colon glyph is now conditioned on the punctuation genuinely bridging two neighboring words on both sides, so an isolated punctuation cell near a real column boundary (a - placeholder, a :) no longer fuses two columns' content into one cell. (GH#1649)

• (ocr): the registry reports each backend's declared languages. list_ocr_backends() returned only names, so a caller validating an OCR language against the installed backends had to maintain a second language table of its own. list_ocr_backend_capabilities() returns one OcrBackendCapabilities record per registered backend — its name and the languages it reports — ordered by backend name. An empty supported_languages means the backend does not enumerate its languages, not that it supports none, so ocr_backend_supports_language() is provided for callers that need a decision rather than a list. (GH#1643)
• (ocr): mixed-page PDF extraction reports one coordinate frame per OCR'd page. Public ocr_elements carry the OCR backend's raster coordinates, but the per-page processed raster size lived only in the page-local result and was discarded before the elements reached the final document. A multi-page consumer was left with the document-level ocr_processed_image_width/ocr_processed_image_height, which is not authoritative for pages of differing size, preprocessing, or rotation, and so could not normalize a box safely. metadata.additional.ocr_page_coordinate_frames now carries one record per OCR page that has public elements — page_number, width, height, unit: "pixel", origin: "top_left" — ordered by page number and joined to an element through the element's own page_number. No record is emitted for a page whose raster dimensions are absent or invalid. (GH#1645)
• (ner): Rust callers can share xberg's process-wide GLiNER backend cache. text::ner::gline::get_or_init_backend performs model initialization on Tokio's blocking pool, returns the same Arc for the same model and thread budget, and leaves failed initializations retryable.
• (pdf): every hierarchy page now advertises the coordinate frame its boxes live in. Hierarchy bounding boxes are emitted in raw PDF user space, but nothing named that space, so a consumer could not tell whether a box needed translating by a non-zero MediaBox origin or rotating for a page's /Rotate. metadata.additional.pdf_page_coordinate_frames now carries one record per page that has hierarchy blocks — page_number, origin_x/origin_y (the MediaBox llx/lly, which may be non-zero and negative), width/height, unit: "point", origin: "bottom_left", and clockwise_rotation — ordered by page number. width/height are the MediaBox extent and are deliberately not swapped for a 90/270 rotation, because they describe raw user space rather than the displayed frame; this is the opposite convention to ocr_page_coordinate_frames, which reports an already-rotated raster, so the two must not be assumed to agree for the same page. A page is omitted entirely, rather than reported with a default, when its /Rotate is present but not a valid multiple of 90 or its MediaBox yields no usable extent. (GH#1653, GH#1654)

• (ocr): an isolated blank quantity cell in an invoice table can be recovered from its own pixels. A table whose QTY column has strong integer support — at least two recognized whole-number quantities — but exactly one blank cell now gets a single targeted retry that re-reads only that cell's bounded pixel region in single-line segmentation mode. The retry never derives a quantity from a price or total column, and any table outside this narrow shape takes the existing fast path unchanged. A header split across two words (UNIT/PRICE, LINE/TOTAL) is also reassembled into one header cell when the fragment sits nearer a data-supported column than any standalone column of its own. This OCR quality change ships without a corpus-wide A/B measurement.
• (reranker): a cancelled rerank_async no longer releases its concurrency permit while its inference is still running. The permit was held by the awaiting future rather than by the spawn_blocking task, and a blocking task cannot be cancelled — dropping its handle merely detaches it. A caller that timed out or dropped therefore returned the permit immediately while the model load and inference it was bounding continued, so repeated abandoned calls could exceed the configured reranker concurrency and keep several models resident after every caller had returned. (GH#1641)
• (embeddings, transcription): the same cancelled-permit leak fixed above for the reranker is now fixed for embed_texts_async and audio/video transcription. Both bound a spawn_blocking inference call with a concurrency permit held by the awaiting future rather than by the blocking task, so a caller that dropped its future released the permit while the ONNX embedding batch or Whisper inference it was bounding kept running. Transcription is timeout-wrapped by default (transcription.timeout_ms), so an expiry there was a live, designed-in path to the same concurrency breach, not a hypothetical one. (GH#1641)
• (pdf): a wrapped numbered heading set in a heading-size font no longer loses its second line. follows_section's two continuation exemptions only recognized a wrap via a hanging indent or via matching right edges, both of which a no-indent heading's short last line fails by construction. A third exemption, heading_continuation_at_margin, now recognizes a wrap by measuring the heading's own line against the width of the body column beneath it, the same signal the paragraph-merge pass already uses for a body-sized heading. Independently, the paragraph-gap detector could still cut a continuation these exemptions accepted, because a heading's own line pitch legitimately exceeds the body's and was read as a blank-line break; the gap check is now suppressed at exactly the boundary a continuation exemption already accepted. (GH#1650)
• (packaging): the Helm chart is published again. The chart publish was gated such that a failed Docker build leg skipped it entirely, while the container images themselves published and the release still reported success. No chart was published for 1.1.4, 1.1.5, 1.2.0, 1.2.1 or 1.2.2, so helm install --version <current> could not resolve. The chart for 1.2.2 has been published retroactively, and a release now fails loudly if its chart is missing rather than shipping without one.
• (node): a JavaScript plugin registered through the Node binding now actually runs. Every generated Node trait bridge -- OcrBackend, PostProcessor, Validator, Renderer, TokenizerBackend and the rest -- was structurally unable to call back into JavaScript. It invoked the JS callable directly from whatever Tokio worker thread the pipeline happened to be on, with no N-API handle scope; it never awaited a returned promise, so an async method's result was the string "[object Promise]"; and it coerced whatever came back to a string before parsing it as JSON, which fails for a plain object as well. The published .d.ts had always declared the working contract (processImage(imageBytes: Uint8Array): Promise<ExtractedDocument>), so no plugin could have been written against the shipped behaviour. Bridges now hold a ThreadsafeFunction created on the JS thread, await the reply through call_async_catch (a JS throw surfaces as a plugin error instead of aborting the process), and decode the value without the string round-trip. A handle reference that would otherwise be released off its owning JS thread is leaked rather than freed, since freeing it there corrupts the V8 heap. (GH#1636)

• (pdf): a page whose MediaBox is not rooted at the origin no longer reports the wrong page size. The native engine's per-page text result took its width and height from the MediaBox upper-right corner instead of its extent, so a page with MediaBox [10 -100 622 692] reported 622x692 rather than the true 612x792. That figure feeds hierarchy reading-order computation, so a non-zero MediaBox origin could mis-order content, not merely mis-scale it. Pages rooted at (0, 0) — the overwhelming majority — are unaffected, since the two expressions agree there. (GH#1653)

• Upgraded the crawlberg dependency to 1.7.0 (from 1.6.4). xberg re-exports crawlberg's crawl types -- CrawlConfig, MapResult, SitemapUrl -- so the pin is part of this crate's public surface. 1.7.0 makes no Rust API changes; its breaking changes are confined to crawlberg's own Java and Swift bindings, which xberg does not consume.

• (pdf): a two-column page is no longer split down the middle of a column because one line crosses its gutter. The corridor search treated whitespace as "no span's bbox", so a single centred footer, caption, or heading set across both columns closed the page's real gutter for every line on the page. The per-line median then placed the split inside the left column, every line it crossed became a band boundary, and the reordered text interleaved the two columns mid-word — producing spliced tokens such as activamodel and demandviating. A split that six or more lines run through now also considers corridors at most one line crosses, still bounded by the same distance cap and still refused when either side does not read as a column. Measured over 490 corpus PDFs: 7 documents change, all of them repairing cross-column splices, and none loses vocabulary. Reported and fixed by Data Polder. (GH#1619)
• (pdf): a wrapped numbered heading now closes above the run-in beneath it. The GH#1634 fix's closing term also required a right-edge test that only applies to the un-indented case. A hanging indent heading's wrap and a run-in sub-heading below it are both short lines by definition, so their right edges land within tolerance of each other by coincidence, which overrode the correct left-edge answer and held the heading open across the run-in and the body under it. (GH#1637)
• (ppt): a legacy .ppt slide's title is read from the file rather than guessed. The title came from the first line of the slide's merged text, gated on a second line existing and the first being at most 80 characters — so a title-only slide had no title, a two-line title was truncated to its first line, and a longer title was dropped. The outline collection's own TextHeaderAtom states which text is the title; that is now read, with the first-line rule kept only for decks that state no title. (GH#1635)
• (ppt): the live save's slides are read, not the oldest save's. The persist chain resolved the current revision correctly and then scanned the stream for the first SlideListWithText, which on a deck saved more than once belongs to the earliest save — so edited titles and outline text reverted to a superseded revision. (GH#1639)
• (ppt): speaker notes attach to the slide that owns them, and the notes master is no longer one. The notes master's placeholder text ("Click to edit Master text styles…") was emitted as the first slide's speaker note, and the remaining notes were zipped to slides by position, so every note after a slide without one landed on the wrong slide. Notes are now keyed by the NotesAtom's own slideIdRef. (GH#1640)
• (release): the xberg-cli archives are attached again. v1.2.0 published 12 assets and v1.2.1 published 54, both with none of the nine xberg-cli-* archives, so mise, cargo binstall and every direct download had nothing to fetch. One matrix leg (the aarch64-musl link) failed, and the upload job required all three CLI build jobs to succeed, so a single leg zeroed the whole asset class — as a skipped job, which reports no failure and left both releases looking green. The upload now attaches whatever built, and a following step fails the release when any of the nine is missing. (GH#1638)
• (pdf render): s (close-and-stroke) is no longer dropped, and no longer floods the next fill. The object-level content parser had arms for every path-painting operator in ISO 32000-1 Table 60 except s, which fell through to an operator every consumer ignores. Two things followed: the stroke was never painted, and — because only a painting arm resets the path builder — the abandoned geometry stayed in the builder and was painted by the next fill. A stroked frame followed by an ordinary white label fill therefore flooded the frame's whole interior, wiping anything already drawn inside it. Text extraction was unaffected: the byte-level fast path always decomposed s correctly. (GH#1633)
• (pdf render): a glyph is no longer discarded because its inferred Unicode is whitespace. The byte-indexed rasterizer decided whether to paint by asking what character a code represents rather than whether there was an outline to draw. For a simple TrueType font with a byte-indexed cmap and no /Encoding that character is the glyph id reverse-mapped through the font's own (1, 0) subtable and read as ASCII or Mac Roman — private glyph ordering decoded as an encoding it never expressed. Six byte values were affected (0x09–0x0D, 0x20, and 0xCA → U+00A0); re-indexed subsets numbering their glyphs from 0x01 upward walked through a whole run unpainted while the advance still applied, leaving gaps that read as spaces. (GH#1632)
• (pdf): a hanging-number heading no longer swallows body text indented to its title's edge. A line was treated as a numbered heading's continuation whenever the heading had a hanging indent and the line began at the title's left edge — which is equally the geometry of any layout that indents the whole clause, as contracts, tenders and many installation manuals are set. Wrapping now also requires that the preceding line actually ran out of room. Separately, a heading that had absorbed a genuine wrap could never be closed, so the sub-heading and entire body below were pulled in after it. (GH#1634)
• (cli): the Linux musl CLI binaries are published again. Since 1.2.0 the aarch64-unknown-linux-musl CLI build has been killed mid-link, and because the upload job requires every musl leg to succeed, no CLI assets were attached to the 1.2.0 or 1.2.1 releases. The cause was the switch to fat LTO: over the all feature set the final whole-program link exceeded what the arm64 runner could complete. That build now uses a release-musl profile with thin LTO; every other target keeps fat LTO.
• (php): setting outputFormat no longer throws. Every call that passed an output format — ExtractionConfig::from_json(json_encode(["outputFormat" => "markdown"])), and any extract/extractBatch carrying a per-input config — raised invalid type: string "markdown", expected struct OutputFormat. The generated Xberg\OutputFormat class carried a derived serde implementation that read and wrote its own storage shape, {"type":"markdown"}, while the core enum is externally tagged and writes a bare string, so no real value ever deserialized. EntityCategory and PiiCategory carried the same defect. Introduced in 1.2.0 alongside the flat-class change and shipped in three releases.
• (wasm): metadata.format fields are readable again. metadata.format.title returned '' and metadata.format.sheetCount returned NaN for every document, because the value reached JavaScript as a Map rather than the plain object index.d.ts declares — serde-wasm-bindgen renders every serde map that way, and serde's internally tagged enum representation is a map. Property access on a Map is undefined. Introduced in 1.2.0 with the FormatMetadata flatten.

• extraction::ppt::PptSlideText carries the slide's stated title and notes. The struct gains title: Option<String> and notes: Option<String>, so Rust callers constructing it with an exhaustive struct literal must add the two fields. The type is not exposed through any binding.
• Breaking (PHP binding): OutputFormat, EntityCategory and PiiCategory now serialize to the wire their Rust counterparts emit — "markdown" for a fieldless variant and {"custom":"latex"} for one carrying a label — instead of the flat {"type":"markdown"} object. from_json still accepts the object form, so existing input keeps working; output moves.
• Breaking (WebAssembly binding): metadata.format — and any other field carrying a flattened, camelCased enum — is now a plain object rather than a Map. This is the shape index.d.ts has always declared; code that worked around the discrepancy with .get() must switch to property access. Fields carrying an enum that is not flattened, such as chunking.sizing, keep the serde-wasm-bindgen bridge and are unchanged.
• Breaking (C#, Dart, Go, Java, Kotlin, Python, Swift bindings): an enum variant whose name ends in a multi-letter acronym followed by a single lowercase letter is spelled correctly now. StructuredDataType.RdFa becomes Rdfa (C#, Dart, Go, Swift), Kotlin's R_D_FA becomes RDFA and PADDLE_O_C_R becomes PADDLE_OCR, and Python's stub and runtime agree on RDFA. Wire values are unchanged — only the identifiers move. Names without that shape (IOError, HTMLParser, JSONLD) are unaffected.

• (swift): externally tagged enums now decode the wire the core types actually emit. Regenerated on alef 0.87.1. EntityCategory, PiiCategory and ConfidenceSemantics relied on Swift's synthesized Codable, which keys every variant ({"person":{}}), while serde writes a bare string for a fieldless variant and a single-keyed object whose value is the payload ({"custom":"foo"}). Every bridge-constructed value carrying one of these threw DecodingError.typeMismatch — Entity and PiiEntity decode their category into a non-optional field, so any result containing an entity failed. OutputFormat is handled separately: its Custom(String) variant carries #[serde(untagged)], so it round-trips as a bare string rather than a keyed object.

• Dependencies upgraded across the workspace, including crawlberg 1.6.1 → 1.6.3 and tree-sitter-language-pack 1.19.0 → 1.19.1. skrifa stays pinned at 0.46: 0.47 moves to read-fonts 0.44 while harfrust 0.13.3 is still on 0.43, and the PDF text rasterizer passes a harfrust FontRef to skrifa's OutlineFace.

• ServerConfig gained job_timeout_secs (default 600 seconds, override via XBERG_JOB_TIMEOUT_SECS or server.job_timeout_secs) as the configurable fallback timeout for POST /extract-async jobs whose request does not pin down extraction_timeout_secs. A per-request extraction_timeout_secs still always overrides it, and an explicit extraction_timeout_secs: null still falls back to this server cap rather than running unbounded. Previously this fallback was a hardcoded 300 seconds, inconsistent with the 600 second default used everywhere else. See the Changed section for the source-compatibility impact.

• A long-running process can now release the embedding and reranker models it no longer uses. embeddings::evict_model and reranking::evict_model (and the same functions in sparse_embeddings and late_interaction) drop one model, clear_engine_cache drops every model in a cache, and xberg::clear_engine_caches drops all of them. set_engine_cache_limit bounds the number of resident engines in a cache and drops the least recently used one first. The default stays unbounded, so existing callers see no change (GH#1626).

• OutputFormat.custom in the Python binding always returned None, even when the value genuinely was a custom format. The accessor derived a discriminator from the enum's JSON, which works for every tagged representation but not for this variant, whose payload is untagged and so carries no discriminator to find. It now matches on the variant directly and returns the label.

• The Python type stub declared a type: str attribute on 23 enum classes that have no such attribute at runtime. The stub emitted it for every data enum, while the runtime only exposes it for enums carrying an explicit serde tag -- so for externally tagged enums (EntityCategory, PiiCategory, OutputFormat) a type checker accepted category.type, which raised AttributeError on use. Stub-only change; no runtime behaviour moved.

• The Ruby binding's FormatMetadata.from_hash and DiffLine.from_hash read a _0 key that no longer exists on the wire, so every variant deserialized with a nil payload. All 24 call sites are corrected: the 21 FormatMetadata variants now build their payload from the flattened hash, and the 3 DiffLine variants read the text key the enum's #[serde(content = "text")] actually emits (#1594).

• The Python type stub declared format-metadata payloads as _0 (for example _0: ExcelMetadata), a key present neither on the wire nor on the runtime object -- the runtime per-variant getters such as .excel were always correct. The stub now names the real fields, so type checkers stop reporting valid code as an error. Runtime behaviour is unchanged (#1594).

• The Elixir Xberg.FormatMetadata typespec documented a metadata: payload key that matched neither the NIF struct nor the serialized wire. It now matches the struct the NIF actually returns (#1594).

• A Type0 (composite) font's content-stream character codes are now translated to CIDs before glyph widths and vertical metrics are looked up, instead of being used as if they already were CIDs. The two coincide only for Identity-H/Identity-V, which is presumably why this went unnoticed. A PDF using a non-Identity predefined CMap (UniCNS-UCS2-H, UniJIS-UCS2-H, UniGB-UCS2-H, UniKS-UCS2-H, or their -V/UTF16 counterparts) or an embedded /Encoding CMap stream got a wrong width for nearly every glyph -- some over-advancing by up to 4x through the /DW fallback, others under-advancing -- which rendered as stretched or overlapping text and, in extracted text, could split one sentence into a spurious extra paragraph. CIDs now resolve from an embedded /Encoding CMap stream's own begincidrange/begincidchar data when present (including a variable-width codespace), else from the font's /CIDSystemInfo character collection for the four Unicode-keyed predefined families above, else via Identity-H/Identity-V as before. Measured over the 230-document local PDF corpus: 227 byte-identical -- the expected result, since most PDFs use Identity-H -- and 2 changed, both merging text that a stale glyph position had fragmented. Legacy multi-byte predefined CMaps this crate carries no code-to-CID table for (90ms-RKSJ-H, GBK-EUC-H, B5-H, the UTF8 family, UniJIS-UCS2-HW-*, UniJISPro-*, and others) are unchanged -- still wrong, not newly broken -- and now log once per font instead of failing silently (GH#1631).

• A reconstructed PDF table cell now reads left to right instead of in the order its words happened to arrive. The reported symptom was a sub/superscript printing after the rest of the cell -- eta_S % came out as eta % S, Q_HE GJ as Q GJ HE -- because a script is drawn as its own content-stream segment a fraction of a point below the line it annotates, so every reading-order sort upstream placed it after the whole line. The same defect also transposed values between columns when two columns were merged into one cell: on a balance sheet whose header reads 2017 2016, the row beneath it emitted the 2016 figure first, silently attributing each year's number to the other year. A cell's words are now grouped into visual lines and ordered left to right within each line. Grouping first is load-bearing -- ordering by horizontal position alone interleaves the two halves of a wrapped cell. Each word's own whitespace is also collapsed, so a segment carrying a trailing space no longer stacks it on the separator. Table cells recovered by OCR go through the same ordering (GH#1628).

• Image OCR now honours a PNG's embedded pHYs pixel density instead of assuming 72 DPI. A genuine 300-DPI PNG submitted with target_dpi = 300 was resized anyway, because the extractor decoded, resized and re-encoded the image -- discarding the density chunk -- before the OCR backend, and therefore before the existing ocr.backend_options["source_dpi"] override, ever saw it. Embedded density is now resolved at the extractor boundary and at the backend from one shared implementation, with the explicit override still taking precedence over it. An image carrying no density metadata still defaults to 72 DPI and still resizes (GH#1630).

• A body paragraph is no longer deleted for repeating text that appears elsewhere on the same page. The second strip_repeating_text pass keyed on lowercased paragraph text with no check that a table was involved, so a sentence matching an earlier title -- differing only in case, with no table on the page at all -- was silently removed. The pass now runs only on pages that have a detected table, removes a paragraph only when that table's own cells carry the same text, and compares case-sensitively. Measured over 230 PDFs: 44 documents changed, 3472 words recovered and 32 lost, both loss cases inspected and benign (one is a restructure whose total content grew, the other two mojibake tokens) (GH#1623).

• OCR text is no longer discarded when a scanned page region is detected as a table but its cell grid cannot be recognised. recognize_single_table returned nothing whenever TATR failed, produced no rows or columns, or the grid failed validation, which threw away every OCR element that had been assigned to that region. A region that cannot be recognised as a table now falls back to emitting its text in reading order, and only when that text is not already carried by one of the page's paragraphs, so nothing is duplicated. Together with the restructuring-heuristic retention guard below, recognised OCR text is no longer silently lost on the layout path (GH#1622).

• extraction_confidence no longer reports a failed structured extraction as fully schema-valid. The pipeline passed SchemaCompliance::AllValid unconditionally, which is 40% of the combined score under the default weights, so a run whose LLM call failed -- or that was built without the liter-llm feature, or ran on wasm -- scored exactly as high as one that validated. A requested structured_extraction that leaves no structured_output now scores AllInvalid. Extractions with no structured_extraction configured are unaffected and keep their previous score; ConfidenceSignals is unchanged in shape, so no serialized form moves (GH#1624).

• detect_mime_type_from_bytes no longer refuses text that is not valid UTF-8. A byte buffer with no filename or declared type -- a Windows-1252 or ISO-8859-1 CSV export, say -- returned UnsupportedFormat even though the extractors that would receive it decode legacy encodings through encoding_rs. Such content is now reported as text/plain, the same answer the UTF-8 path already gave for the same document, so the two encodings of one file behave alike. Content holding a NUL byte, or with too few printable bytes to read as prose, is still rejected (GH#1625).

• The Go binding no longer discards the message of every error the native layer reports. Each known error code was mapped to a typed sentinel (ErrTimeout, ErrParsing, ErrOcr, and ~20 more) and returned before the message was ever read, so the detail the native layer had already produced -- observed durations, limits, plugin names, counts -- was dropped for all of them; only unrecognised codes kept their text. A timeout surfaced as the sentinel's own placeholder-stripped text, extraction timed out after ms (limit: ms), which reads as a formatting bug but is the whole message the binding ever had, and left callers unable to tell which timeout had fired. The message is now read first and returned alongside the sentinel, so errors.Is(err, xberg.ErrTimeout) still matches while err.Error() carries the real interpolated text. Go was the only binding affected; C#, Java and Zig already read the message before switching on the code. Regression in 1.1.0, when the typed sentinels were introduced.

• The Python package's public option classes regained from_json. from xberg import ExtractInput resolves to a generated dataclass that shadows the native class at the same name, and that dataclass carried none of the native class's methods, so ExtractInput.from_json(...) raised AttributeError while xberg._xberg.ExtractInput.from_json(...) worked -- the same name meaning two different things depending on the import. 134 public classes were affected. The dataclasses now delegate from_json to the native class, so both import paths behave the same. Other native-only methods on those classes (validate, is_empty, the PaddleOcrConfig.with_* builders) are still absent from the dataclass twins and are tracked separately.

• An extraction cancelled by extraction_timeout_secs now actually stops its per-page PDF OCR work. The timeout fires cancel_token.cancel() at every timeout site, but nothing in the OCR page fan-out read the token, so pages kept being OCR'd after the caller already had its Timeout error -- burning CPU and holding OCR concurrency permits, which degrades later extractions in a long-lived process (a server, or anything extracting in a loop). The token is now checked both before spawning a page and inside each spawned task, because the spawn loop finishes almost immediately while tasks queue on the OCR semaphore long after it. A cancelled run also reports Cancelled instead of tripping the all-pages-failed guard and reporting a wholesale OCR backend failure.

• PDF no longer promotes ordinary body text to a heading. Two gates decide headings independently and neither tested the line's shape, so any line past the title-length floor could be promoted. The sentence-boundary check that should have caught this looked for a literal ". " followed by a capital, but paragraph text joins a block's physical lines with a newline, so every sentence boundary landing at a line end was invisible to the gate while the renderer joined the same lines with a space and displayed it -- the gate and the output disagreed about what the text was. Boundaries are now found across any whitespace, and a line that is mostly bare numerals is treated as a flattened data row rather than a heading. Across 490 documents: 484 unchanged, 5 with fewer headings, 0 with more (GH#1599).

• Hardened the document-global heading/list heuristic's safety check on the scanned-PDF layout-markdown path (use_layout_for_markdown / layout detection, force-OCR route). That heuristic rebuilds paragraphs from bare line geometry with no knowledge of the ML layout regions the OCR path already classified, and can silently drop a line its own font-clustering pass treats as furniture or noise; the guard against this only checked that the whole document still had one non-empty element, so a single surviving word anywhere passed it even if an entire page's body vanished. The guard is now a per-restructuring canonical-character retention check against the lossless OCR assembly, and falls back to that lossless assembly whenever any content would otherwise be lost. Compares characters rather than word tokens: a restructuring pass legitimately re-wraps text across the line boundaries it reads (measured case: "list of findings" split across a line came back "list offindings", one dropped space), and a word-token comparison read that benign re-wrap as content loss and rejected legitimate heading/list promotion along with it. This closes a real gap in the guard's own logic; it was not reproduced against a specific "entire page lost" report and should not be read as a confirmed fix for one (GH#1622).

• PDF table/paragraph assembly (assemble_page_elements_with_tables) now suppresses a paragraph whose words a positioned table's own grid fully carries, so a recognized table no longer also renders its flattened source text as an ordinary paragraph immediately next to the grid -- observed directly (not inferred) on a scanned-PDF fixture with layout detection enabled, where a table's status-row text appeared once as prose and once as a correctly gridded table. Mirrors the GH#1616 precedent from the other direction: suppression requires the table's own cell/markdown content to actually account for every one of the paragraph's words (an order-insensitive multiset match, since a reconstructed grid can reassemble the same words in a different order than the source paragraph), not geometry alone, so a paragraph carrying text the grid does not represent still survives. Note: this closes the duplication for paragraph/table pairs that share a coordinate space (the native-PDF table path). Investigating this also surfaced a separate, unresolved coordinate-space mismatch between OCR/TATR-recognized table bounding boxes and OCR paragraph bounding boxes on the force-OCR + layout-detection route specifically, which currently prevents this same guard from geometrically matching on that route; fixing that is out of scope here and is not yet done (GH#1622).

• PDF no longer deletes text a table's bounding box covers but its grid leaves out. Suppression of text a table already renders was decided on geometry alone, and a reconstructed grid need not span every printed column inside its own bounding box. On a four-column fault-finding grid reconstructed with two columns, every run in the two omitted columns vanished from the document — not in a cell, not in any element, nowhere. A covered run is now suppressed only when the table actually carries its text (GH#1616).

• PDF no longer cuts a numbered heading that wraps onto a second line. The wrap exemption compared the two lines' right edges, and a wrap's last line is short by definition, so it could never fire: the heading kept only its first line and the rest of its title was emitted as body text. A heading's own continuation is now recognised by its left edge, which is the title's hanging indent rather than the margin body text returns to. Regression in 1.1.5 (GH#1615).

• An extraction that never requested OCR no longer fails when no OCR backend is registered. ocr-pipeline can be enabled without any backend — ocr implies ocr-pipeline, not the reverse — and in that build the automatic scanned-page trigger aborted an ordinary PDF extraction with OCR backend 'tesseract' not registered. Automatic triggers now check availability and skip with a warning; an explicit force_ocr, force_ocr_pages, ocr_inline_images or caller-supplied ocr config still fails loudly (GH#1610).

• Legacy binary .ppt now reports which slide each embedded picture belongs to. Pictures were read from the OLE Pictures stream, which stores blips in save order and names no slide, so every extracted image carried no page number and every image node was emitted after the last slide. A slide whose only content is a picture therefore produced nothing at all on its own number and read as a blank slide, and captions or any other data keyed on an image's page were filed against the end of the deck. The owning slide is now resolved through the drawing that references the blip; a picture no live shape references is still extracted, without a slide (GH#1620).

• Legacy binary .ppt no longer extracts deleted slide revisions or presents slides in the wrong order. The format is append-only across saves, so editing a deck leaves superseded copies in the stream; treating every Slide container as a slide produced 190 slides for a 96-slide presentation, numbered by byte order. Live slides and their order now come from the persist chain (Current User → UserEditAtom → PersistDirectoryAtom) and the document's slide list, falling back to the previous behaviour if the chain cannot be read in full. Slide numbers are the page every element and chunk of a deck is cited by, so both defects reached consumers as wrong page numbers (GH#1614).

• PDF de-hyphenation no longer welds a compound whose own hyphen falls on a line break. Two sites decide whether a trailing hyphen survives; only one consulted the lexical evidence, so long-term, cost-effective and antigen-presenting came out as longterm, costeffective and antigenpresenting — tokens that do not exist, and so unreachable by any lexical search. The assembly site now asks the same question the paragraph site already asked, weighing both the static compound list and the witnesses collected from the document itself. A hyphen the wrap genuinely inserted is still removed (GH#1613).

• Legacy binary .ppt no longer loses slide titles. PowerPoint keeps a slide's text in two places, and the extractor read only one: titles held in the document-level outline collection (SlideListWithText) landed in the loose-text bucket, which is discarded whenever any slide exists, so they were absent from the output entirely. Outline text is now attributed to its slide by persist order and merged in, skipping any line the slide's own drawing already carries so a title drawn on the canvas is not duplicated (GH#1612).

• The documented install versions for Java, Kotlin Android, Swift, Zig and the spring-ai integration no longer lag the release. These snippets sit outside task version:sync, which covers the generated API-reference badges but not hand-authored install directives, so they had been telling users to install 1.1.3 (GH#1593 covers the same class of staleness in test_apps, which is still open).

• OcrConfig no longer rejects valid Tesseract language codes such as fao (Faroese) with Invalid language code 'fao'. Use ISO 639-1 or ISO 639-3 codes.. Config validation checked the language against a general-purpose allowlist that was missing 66 codes Tesseract actually supports, while a separate, Tesseract-specific list already carried them; the two lists had never been reconciled. Config validation itself only started running for configs loaded from files, JSON overrides, or set programmatically in 1.1.0 (previously it ran only in tests), which is when this allowlist gap first became user-visible. Both validators now read from one shared list of Tesseract-supported codes, so this class of divergence cannot recur (GH#1621).

• Breaking (Java binding): enum constants now follow Java's own convention and are SCREAMING_SNAKE_CASE instead of carrying Rust's PascalCase verbatim -- LinkStyle.Inline becomes LinkStyle.INLINE, across roughly 75 generated enums. The JSON wire value is unchanged; only the Java identifier moves, so serialized documents and stored payloads are unaffected. Update references to the constants themselves; generated default values (ChunkType.Unknown, OutputFormat.Plain) moved with the declarations.

• Breaking (Ruby binding): an externally tagged enum variant carrying a single payload (EntityCategory::Custom, PiiCategory::Custom, OutputFormat::Custom) now serializes the way the Rust core always did -- {"custom" => "my-label"} -- instead of wrapping the payload in an extra object keyed by a synthesized positional name, {"custom" => {"_0" => "my-label"}}. Code reading hash[:custom][:_0] should read hash[:custom]. The previous shape matched no other binding and no core output.

• Breaking (PHP binding): EntityCategory, PiiCategory and OutputFormat change from constants-only classes to classes with static factories, because the old shape could not carry a payload at all: a caller-supplied label was silently discarded in both directions, so Custom($label) always round-tripped as an empty string. Use EntityCategory::custom($label) and EntityCategory::person() in place of the old Xberg\EntityCategory::PERSON constants; the label is readable from the readonly $custom property.

• Breaking (Node binding): the JSON surface is camelCase throughout, nested types included, and FormatMetadata is a flat discriminated union keyed by formatType whose variant payload fields sit directly on the object ({ formatType: "excel", sheetCount: 2 }). The binding previously exposed two parallel shapes for one Rust type -- an idiomatic camelCase interface beside a snake_case structural twin -- and only the latter was reachable from a result.

• Breaking (Node, Swift bindings): FormatMetadata now serializes flat in every binding -- {"format_type": "pdf", "page_count": 12, ...} -- matching what the core Rust enum has always serialized (#[serde(tag = "format_type")]), what the OpenAPI discriminator describes, and what the REST API serves. Two bindings disagreed with that wire and have been corrected:

  • Node nested the payload one level down under a property named for the variant, so doc.metadata.format.pdf.pageCount becomes doc.metadata.format.page_count. Note the field names are snake_case, unlike the camelCase Node uses elsewhere: the variants carry mutually incompatible field types (headers is string[] for text and an object array for HTML), so no single Node class can describe them and the value is passed through as serde emits it. format is typed as a discriminated union in index.d.ts, so narrowing on format_type still gives a fully typed payload. This also resolves the Node and WebAssembly bindings disagreeing with each other -- WASM was already passing serde's shape through, so the two now emit an identical format object for the same document.

  • Swift's FormatMetadata was a typealias to an opaque bridge class carrying no payload, so the JSON in Metadata.format could not be decoded into anything useful. It is now a real Codable/Sendable enum with one case per format, making the payload reachable:

    if let json = doc.metadata?.format,
       case .excel(let meta) = try formatMetadataFromJson(json) {
        print(meta.sheetCount)
    }

    Metadata.format still hands back the serialized JSON String; what changed is that formatMetadataFromJson now yields a pattern-matchable enum carrying the payload instead of an opaque handle. The wire was already correct here -- the Swift type system was the part that was missing.

  Python, Go, Java, C#, Kotlin, PHP and Ruby are unaffected on the wire: they either already emitted the flat shape or expose native per-variant accessors over it (#1594).

• Breaking (Rust source, Java): ServerConfig adds job_timeout_secs. Exhaustive Rust struct literals must set the field or use ..ServerConfig::default(), and the Java record's canonical constructor gains a sixth component, so new ServerConfig(host, port, corsOrigins, maxRequestBodyBytes, maxMultipartFieldBytes) no longer compiles -- use ServerConfig.builder(), which is unaffected. Every other binding is source-compatible: the field is last and defaulted in the Python dataclass (= 600), the Kotlin data class (= 600L) and C# ({ get; init; } = 600); a defaulted keyword in Ruby and PHP; an optional pointer with omitempty in Go; and an additive xberg_server_config_job_timeout_secs getter in the C FFI (gated on api-types). Deserializing callers are unaffected everywhere -- the field carries #[serde(default)].

• Public binding-facing structs in this crate are deliberately not #[non_exhaustive]: alef generates impl From<Mirror> for xberg::T with a struct literal in roughly ten binding crates, and #[non_exhaustive] forbids that cross-crate (E0639) -- including the ..Default::default() spread. Default plus #[serde(default)] is the forward-compatibility mechanism instead, and a field addition is recorded here as a labelled source break rather than prevented by the type system. #[non_exhaustive] is reserved for types excluded from binding generation.

• Retroactive note for 1.1.4: Metadata#format in the Ruby binding changed shape and no changelog entry recorded it at the time. The format-specific payload had been nested under a _0 key (format.fetch(:_0).fetch(:title)); since 1.1.4 the payload's fields sit directly alongside the format_type tag (format.fetch(:title)). Ruby callers written against the older shape raise KeyError on _0. The binding has emitted the flat shape since 1.1.4; the generated Ruby e2e specs were still asserting the nested one, which is why this went unnoticed for two releases. Only the Ruby binding is affected. Part of GH#1594, which also tracks the Swift binding still discarding the payload entirely -- that half is not yet fixed.

• The Java binding compiles again. A method returning Option<Vec<u8>> — Registry.sampleBytes is the only one today — was generated declaring Optional<byte[]> while returning a bare byte[], which javac rejects. 1.1.4 therefore published no Java artifact at all, and the spring-ai integration was blocked waiting on it. Fixed upstream in alef 0.85.12; this release regenerates on it.

• BREAKING (Ruby): FormatMetadata reaches Ruby as a flat hash. It previously arrived as {format_type:, _0: {...}}, where _0 was the name serde invents for an unnamed tuple field; it now arrives as {format_type: 'excel', sheet_count: 2, ...}, the canonical wire the core declares. Code reading metadata.format[:_0][:sheet_count] must read metadata.format[:sheet_count]. This shipped unannounced in 1.1.4 and is recorded here retroactively; no other binding's shape changed (GH#1594).

• PDF reading order no longer tears a subscript off the symbol it names. Spans were ordered by the top of their bounding box, but a subscript is drawn 35-40% smaller than its base, so its top sits several points lower even though its baseline is a fraction of a point away. An unrelated span from the next column could sort between a base run and its own subscript, and the symbol the subscript names no longer existed anywhere in the output. Ordering now quantises the baseline into row bands before comparing horizontally, which is what every other caller of that comparator already did (GH#1600).
• PDF table detection no longer bridges two separate tables across the graphics-free gap between them. A cell was built from intersection points alone, so a section heading printed in that gap was absorbed into one of the tables as a single-cell row. A candidate cell now also requires a drawn vertical rule spanning its own Y-range on both sides. The span tolerance is load-bearing: at the tighter X-axis value, rows of a table whose rules are inset by a few points are dropped (GH#1601).
• PDF two-column detection no longer loses the page's split to a hanging-number indent. When any span straddled a correctly detected gutter, the split was replaced outright by the midpoint of the widest whole-page whitespace corridor — on a hanging-number layout, the indent between the numbers and the text. The reorder then hoisted every clause number out of its clause. A relocation is now rejected when it would move the split more than a quarter of the page width, which leaves every legitimate corridor move in the corpus intact (GH#1603).
• PDF paragraph grouping no longer splits a numbered heading that wraps onto a shorter second line. The wrap exemption compared the two lines' right edges, but a heading fills its column on its FIRST line and the continuation is whatever is left over, so the metric was anti-correlated with the answer. The pair is now also exempt when the continuation opens lowercase AND the heading line reaches within a tolerance of the width of what would be merged onto it — the "fills its column" half the original rule stated but never measured. The lowercase test alone is not sufficient: body prose beginning lowercase under a complete numbered heading has the same signature (GH#1605).
• PDF paragraph grouping now recognises a numbered heading whose line arrives as more than one text span. The break terms tested the predicate against a single span, so a heading set with a hanging section number — 3.1.7 in one span, its title in the next, on one baseline — never looked like a numbered heading and was left to the ordinary paragraph-gap rule. That rule needs a gap wider than ordinary line pitch, so every such heading whose body starts on the next line was welded into it. The line's spans are now re-joined before the predicate runs, which is what the continuation-merge pass already did (GH#1609).
• PDF paragraph grouping now recognises a heading whose number is not its first token — ARTIKEL 1., Chapter 1, Appendix 1, Annex III, Exhibit A. The numbered-heading predicate is the only boundary signal available when a heading shares font, size, weight and leading with its neighbour, so a heading it could not see was welded onto the line above it, and a run of such headings collapsed into a single element. Recognition is by shape, not by a keyword list: one capitalised word standing in front of an enumerator. Prose that opens the same way — Artikel 12 van de wet is van toepassing. — stays prose, because behind a keyword the text after the enumerator must still be capitalised (GH#1608).
• PDF heading detection no longer skips a numbered heading that is only two words long. Promotion of a bold, body-size line to a heading required more than two words — a floor that keeps short bold fragments out — and a numbered section title such as 3. PRIJZEN or 1. INTRODUCTION falls below it. Those lines stayed plain bold paragraphs, and a run of them was then coalesced into a single bold line in the rendered output, while the element stream still reported them separately. A numbered section heading is now exempt from the word-count floor; everything else still has to clear it (GH#1611).
• OCR no longer adopts a markdown table rebuild that loses content. The rebuilt page replaced the original whenever it was merely non-empty, so a rebuild that dropped text still won. The rebuild is now rejected, with a warning naming both word counts, when it retains fewer words than the content it would replace (GH#1599).
• PaddleOCR's default model_tier of mobile now resolves to the pp-ocrv6 small detection model (9.9 MB) rather than medium (62 MB). A tier named mobile silently loading the largest available model made a 21-page document take over ten minutes. small and medium share the same 18,708-character dictionary, so recognition coverage is unchanged. The documented model sizes were also wrong and have been corrected (GH#1602).
• The PHP extension now loads on Debian 12 and other distributions built against GCC 12. The Linux publish runners ship GCC 13+, and the extension picked up a GLIBCXX_3.4.31 symbol from their libstdc++ while Debian 12 provides at most GLIBCXX_3.4.30. libstdc++ is now linked statically; the highest glibc requirement was already below Debian 12's (GH#1606).

• Table.cell_styles and GridCell.heading_level / GridCell.style_name expose the paragraph style a DOCX table cell carries. A heading styled Heading1..Heading6 inside a w:tc — the banner row forms, questionnaires and datasheets use as a section title, and what Word's navigation pane and a TOC field treat as the document outline — previously reached every consumer as anonymous cell text. Cell text is deliberately unchanged: prefixing it with # would put a markdown heading inside a table cell. The style travels beside the text instead, so a caller can decide whether a heading 2 in a banner row is a section title or a column label. cell_styles is sparse and omitted entirely for tables whose cells carry no style, so ordinary tables serialise exactly as before (GH#1587).

• PDF text repair no longer welds two complete words into one. repair_ligature_spaces removes the space in …f + + i|l|f… to undo a real artefact — some PDFs decompose a ligature glyph and leave a spurious gap, so first arrives as f irst — but the same character pattern is an ordinary word boundary whenever a word ends in f and the next begins with i, l or f. The only guard was a hard-coded list of 33 short English words tested against the left token, so everything outside it welded, English included: relief for became relieffor and itself infringes became itselfinfringes. The space is now kept when either fragment is independently attested as a standalone word elsewhere in the same document, reusing the witness mechanism dehyphenation already applies. A fragment appearing only as one half of a candidate pair does not witness itself (GH#1591).

• DOCX page counting no longer collapses a table onto one page. Word writes <w:lastRenderedPageBreak/> into every cell of a row that straddles a page boundary — one physical break, one marker per cell — and the duplicated markers were reduced to a single break, losing the originals with the duplicates. A seven-page document reported two. Breaks are now identified by table depth, row and cell, so a marker echoed across the cells of one row counts once while several breaks inside a single deep cell each still count (GH#1592).

• PDF outline (bookmark) named destinations now resolve when the /Names -> /Dests name-tree key is UTF-16BE-with-BOM, the form Adobe Distiller writes. The lookup previously decoded the /Dest byte string with a lossy UTF-8 conversion before searching the tree; a name-tree key is a byte string compared by byte (ISO 32000-1 §7.9.6), not text, so the BOM was mangled into replacement characters and every such destination silently resolved to None, leaving the bookmark's dest as an unresolved Destination::Named with no page (GH#1589).

• MimeDetectionPolicy::ContentOnly no longer rejects a legacy OLE2 Office document (.doc/.xls/.ppt) passed by path when the same bytes are accepted through the bytes API. Path-based content detection only sniffs the first 4 KB of a file, but an MS-CFB compound document cannot be typed from a prefix — identifying it means following the FAT sector chain to the root directory entry, which a truncated buffer cannot do. Detection now falls back to a structure-aware read of the file for a compound-file header that a 4 KB prefix left inconclusive, the same escape hatch a ZIP-based Office document already had for the same class of failure (GH#1590).

• PDF table detection no longer invents a column boundary from a rule that stops short of the row band. BAND_RULE_SPAN_TOL was defined as SNAP_TOL, conflating two different questions: SNAP_TOL decides whether two coordinates are the same coordinate, while this one decides whether a vertical rule runs through a band. At 3pt an edge could fall short at each end and still count as spanning, so a band up to 6pt shorter than the rule beside it was cut where the drawn rule gave it no boundary. Those phantom columns are what let a band of prose inside a drawn frame split into cells and qualify as a table, which on the reported document cost page text. Now 1.0 and deliberately independent of SNAP_TOL (GH#1588).

• Tesseract psm = 0 is now rejected at configuration validation. PSM 0 is Tesseract's PSM_OSD_ONLY — orientation and script detection with no character recognition — so it cannot satisfy a text-extraction request, and Tesseract emits no hOCR for it at all. Setting it previously succeeded while returning either a zero-length document or degraded, partially dropped text, depending on the Tesseract build, in both cases with no warning and at several times the cost of a normal run. The error now names the mode and points at 3 (auto), 6 (single block), and 11 (sparse text). Valid values are 1-13; omitting psm continues to let the pipeline choose (GH#1586).

This release contains a breaking public API change. TesseractConfig.psm is now optional. Callers that read or set it as a plain integer must handle None / null — see below.

• Breaking: TesseractConfig.psm is now Option<i32> (null/None/absent in the bindings) and defaults to unset rather than to 3. This fixes supplying a TesseractConfig at all acting as a hidden behaviour switch: because several code paths keyed on the struct being absent, a caller who set one unrelated field — table detection, a preprocessing knob — silently lost the whole-image PSM 11, the vertical-language PSM 5, the layout-region PSM 6, and the sparse-text retry, and got Tesseract's PSM 3 instead. TesseractConfig() with default fields is now a no-op: the pipeline applies exactly the same automatic PSM it would with no TesseractConfig. An explicitly set psm is still honoured. Bindings that model psm as a plain integer expose a companion presence check (for example xberg_tesseract_config_has_psm in the C API), since a bare integer cannot distinguish "unset" from a real 0.

• Fixed a numbered or bulleted list on a scanned page being reconstructed as a table, replacing the list text with a mangled grid. A candidate region whose first column is list markers (1., a), •) end to end — the header cell included — is now rejected on the OCR routes. A genuine numbered table is unaffected: its first column carries a header label (Line, Item) above the numbers, which is what separates the two.
• Fixed a table detected on a scanned page having its text returned twice — once as paragraphs, once as table cells — in the document content and element tree. This affected every output_format; "plain" only appeared to avoid it.
• Fixed PdfConfig.top_margin_fraction / bottom_margin_fraction defaulting to 0.06/0.05 (6%/5%) since 1.1.0, which silently dropped OCR text — page titles included — in the top and bottom bands of every default-config scanned PDF page with no warning. Both now default to 0.0 (disabled); set them explicitly to filter header/footer content. The nonzero defaults also forced every default-config OCR page onto the lossy per-page route instead of a document-capable backend's whole-document path; that routing is restored too.
• Fixed rendered PDF pages losing the Tesseract backend's own ProcessingWarnings (including the dictionary-filter removal notice) and OCR metadata (psm, language, tesseract_dict_invalid_word_ratio), both of which reached the caller for a standalone image but were silently dropped for the same page rendered from a PDF.
• Fixed OcrConfig.language being discarded whenever a TesseractConfig was supplied, so a German document was OCR'd in English. One precedence rule now governs both Tesseract backends and the vertical-language check.
• Fixed supplying any ImageExtractionConfig suppressing document-level OCR on scanned PDFs, which returned empty pages with only a debug log.
• Fixed rendered PDF pages ignoring the configured render DPI. target_dpi, min_dpi, max_dpi, and auto_adjust_dpi are now honoured. With no configuration the default stays at 150 DPI, unchanged.
• Fixed suspended hyphens being welded during text assembly, turning onderhouds- en into onderhoudsen. A hyphen is now joined only across a genuine visual line break, matching the rule the pipeline layer already applied.
• Fixed an unruled full-width band in a ruled table being cut at column positions no rule gives it, splitting headings mid-word. A column boundary now counts only where a vertical edge actually spans the band.
• Fixed every non-header table cell having its em-dashes, en-dashes and minus signs rewritten to an ASCII hyphen, the spaces around a hyphen collapsed, E-/E+ lowercased to e-/e+, and any cell consisting solely of a dash emptied. That normalisation is correct for a numeric column (an em-dash means nil, 1.5E-05 is an exponent, - 3 is -3) but corrupted prose tables, turning Functionaliteit—12 into Functionaliteit-12 and a part code HRE - HReco into HRe-HReco. It is now applied only to columns whose data cells are predominantly numeric (#1582).
• Fixed the Windows PHP extension archives failing to publish at all. vendor-windows-native-closure.ps1 repacks a .zip with Compress-Archive, which runs no native command and so never sets $LASTEXITCODE; the release workflow gated on it, and an unset $LASTEXITCODE compares as non-zero, so every Windows archive was rejected immediately after being vendored successfully. Combined with an all-or-nothing matrix gate that withheld the release's PHP assets whenever any single leg failed, this left v1.1.0 and v1.1.1 with no PHP binaries at all. Both are fixed: the script now sets its exit contract explicitly, matching its sibling scripts, and the upload job now ships the archives from the legs that succeeded (#1585).

This release contains a breaking public API change. OutputFormat::Structured is renamed to OutputFormat::DocTags. Update any config, CLI invocation, or binding call using output_format = "structured" to "doctags".

• Breaking: renamed OutputFormat::Structured to OutputFormat::DocTags across every binding and the output_format config field. The rename shipped in 1.1.0 but was only alluded to there, with no entry describing it; the variant was renamed, not removed, and is available as "doctags". An output_format of "structured" is not rejected — it resolves to a custom renderer of that name, which is not registered.
• Breaking: the TypeScript and WebAssembly OutputFormat type is a string union again ("plain" | "markdown" | "djot" | "html" | "json" | "doctags" | ...), matching the serde wire format shared with the CLI, REST, MCP, config-file, and Go surfaces. 1.1.0 briefly published an object union ({ type: "markdown" }) for these two bindings only.

• Fixed PHP extension packaging, which produced no PIE archives for 1.1.0.
• Fixed HEIC and AVIF decoding on the Linux (glibc) Node binding, which shipped a libheif built with no HEVC or AV1 decoder at all — every .heic and .avif input failed to decode while the heic feature still reported as present. The Elixir linux-gnu NIF carries the same working codec closure.
• Fixed Elixir NIF publishing for linux-gnu and Windows, which produced no artifacts for 1.1.0 and left the Hex package at 1.0.14. The linux-gnu NIF is now built against the glibc 2.28 floor it claims to support; the artifacts published for 1.0.14 bundled HEIF codec libraries that required a newer glibc.
• Fixed the Windows Hex package, which declared the x86_64-pc-windows-gnu target while CI built and published x86_64-pc-windows-msvc. RustlerPrecompiled resolves the msvc triple on Windows and rejects any triple the package does not declare, so mix deps.get failed with "precompiled NIF is not available for this target" even though the artifact existed. Windows users had to compile the NIF from source.

• The Linux binding images now verify a pinned SHA-256 for every vendored native dependency (libde265, libheif, ONNX Runtime) before building it, instead of trusting the download. These libraries are linked into the published Node and Elixir artifacts.

This release contains breaking public API changes. Entries prefixed Breaking: below remove or change public API — notably OutputFormat::Structured, the ElementId wrapper, ExtractedDocument.formatted_content in the language bindings, and the core::batch_mode, core::formats, and core::io modules — and configuration deserialization now rejects unknown fields rather than ignoring them. Review them before upgrading.

• Added per-page OCR confidence to PageContent.ocr_confidence, reported as a PageOcrConfidence { score, word_count, backend } (#1568). The field is absent for pages that were not OCR'd. score is populated only for backends whose confidence is a calibrated legibility scale (normalised to 0.0..=1.0) and is None for uncalibrated ones, so a page OCR'd without a comparable score is still distinguishable from a page nobody scored. It is reported alongside word_count because noise filtering runs before the score is computed: a high score over very few surviving words does not mean the page read well.

• Added HWPX (Hangul Word Processor XML) extraction to the WebAssembly package. unhwp target-gates its ZIP reader to a deflate-only, LZMA-free build under wasm32, so the native-C dependency that previously kept hwpx off wasm-target does not apply there.

• Added diagram recovery from flat OpenDocument drawings (.fodg), including content-based detection of the application/vnd.oasis.opendocument.graphics-flat-xml MIME type. Connectors name their endpoints outright, so the recovered graph is exact rather than inferred from geometry (#1545 corpus fixture).

• Added structural extraction for MyST Markdown syntax and MyST text notebooks, including saved inline {eval} values in Jupyter markdown cells (#1538).

• Added extraction of Jupytext percent- and light-format notebook scripts, including text/x-python, text/x-r-source, and text/x-julia MIME aliases (#1538).

• Added bounded, cancellable SQLite and GeoPackage table extraction with schema-based GeoPackage detection, .sqlite3 and .gpkx filename support, and defensive handling for untrusted databases (#1510).

• Added configurable MIME inference policies for preferring content signatures, trusting supported filename extensions, or ignoring extensions, with per-input overrides (#1509).

• Added native KML and GeoJSON extraction with canonical MIME routing (#1508).

• Added Rust SUPPORTED_FORMAT_COUNT and SUPPORTED_EXTENSION_COUNT constants derived from the MIME registry, plus automated synchronization for published format-count claims (#1511).

• Added reusable Rust PDF render sessions for querying page counts and rendering multiple pages without reopening the document (#1485).

• Added cooperative cancellation for single and batch extraction (#1476).

• Added dynamic system linking for Tesseract and Leptonica through the tesseract-dynamic feature (#1407).

• Added managed Azure AD, Google Vertex AI, and AWS STS credential providers, with credential values redacted from debug output.

• Added reasoning-effort, provider-specific request-body, and Bedrock configuration for LLM extraction.

• Added xberg doctor and the Rust doctor() API for validating configuration and probing every compiled OCR, VLM, layout, table, formula-recognition, and cache capability without downloading models or contacting remote providers. xberg doctor --clean removes stray files only from Xberg-owned caches (#1347).

• Added the Sceptre EasyOCR Gen2 backend for desktop, mobile, and WebAssembly.

• Added sparse and late-interaction embeddings to chunk output.

• Added a Prometheus /metrics endpoint to the API server (#1391).

• Added explicit CSV delimiters and comment-line prefixes through CsvOptions.

• Added xberg tree-sitter commands for downloading, listing, and cleaning language assets, with optional configuration-file loading.

• Added VLM extraction for complex PDF regions and LaTeX formula extraction from VLM OCR.

• Added structural AsciiDoc and WebVTT extraction.

• Added Docling DocTags input and output, including tables and page geometry (#1383).

• Added formula recognition for rasterized pages and exposed formulas consistently across extracted formats (#1385).

• Added JATS, EPUB, ODT, and ODP MathML-to-LaTeX conversion.

• Added deterministic diagram recovery from SVG and PDF sources with Graphviz DOT output (#579).

• Added SecurityLimits.max_pages for PDF, presentations, Keynote, and multi-frame TIFF documents (#1451).

• Added explicit PDF backend selection through PdfConfig.backend and --pdf-backend (#1448).

• Added musllinux Python wheels and a Windows x86_64 Ruby gem.

• Added PDF and HTML extraction plus layout and transcription types to the WebAssembly package.

• Added --ocr-no-cache to bypass the Tesseract result cache.

• Added ContentFilterConfig.include_footnotes for retaining footnotes classified as page furniture.

• Added a public render_heading_breadcrumb helper for retrieval-oriented chunk content (#1393).

• Added structured-output merge, citation, and vision-fallback helpers for Rust embedders.

• Added a Tower-compatible extraction service, request type, and builder for Rust applications.

• Added typed configuration for TrOCR, PaddleOCR-VL, GLM-OCR, and DeepSeek-OCR backends.

• Added classify_chunks_owned for classifying and returning an owned document.

• Exposed chunk-classification and LLM concurrency, provider, cache, budget, and rate-limit configuration types at the Rust crate root.

• Added OcrConfig::security_limits. ExtractionConfig::security_limits is now threaded through to every OCR route — embedded images, Tesseract, PaddleOCR, and scanned PDF pages — instead of each route decoding images under a hardcoded SecurityLimits::default() (#1554).

• Added detected_language_confidences, carrying each detected language's confidence, proportion, script, and reliability alongside the existing detected_languages codes, so a document that is 95% English and 5% French is distinguishable from an even mix (#261). The existing field keeps its type and ordering.

• DOCX reviewer comments now emit their own NodeContent::Comment node instead of riding the footnote reference and definition machinery, so consumers can tell a comment from a footnote.

• PDF annotations now preserve their subtype (Ink, Square, Circle, Polygon, PolyLine, Line, Squiggly, Caret, FileAttachment, Sound, Movie) instead of collapsing to Other, carry author, modification date, colour, subject, and QuadPoints, recover the text a Highlight marks, and are emitted by the Markdown, Djot, plain, HTML, and JSON renderers — previously no renderer emitted annotations at all (#63).

• PDF extraction now reads image alt text from the structure tree, falls back to XMP for title, author, and subject when the Info dictionary is empty, surfaces /PageLabels (roman-numeral front matter, per-section numbering) through metadata.additional, excludes content on optional-content layers that are off by default, and renders filled AcroForm values. Unencodable images, annotation failures, and form failures now emit a ProcessingWarning instead of being dropped at log level (#62, #71).

• The OOXML DocSecurity bit field is decoded into named protection flags on Metadata.additional for DOCX, XLSX, and PPTX, so a password-protected or read-only-recommended document is distinguishable from an unrestricted one.

• Added PaddleOCR on the tract backend, so classical PaddleOCR (DBNet, CRNN, AngleNet) is available on wasm32 and the Android x86_64 emulator, where ONNX Runtime cannot link.

• Added top_p, stop, seed, presence_penalty, and frequency_penalty to LlmConfig, validated and applied to every outgoing request. They were previously accepted by every config file and language binding and then dropped before reaching a provider.

• Added LlmConfig.max_concurrency to bound VLM OCR and image-captioning requests in flight independently of ConcurrencyConfig.max_threads, which represents local CPU capacity (#1453).

• Every error variant now carries a stable FFI error code, so typed error handling works in the C-ABI bindings; errors.Is(err, ErrOcr) in Go, Java's checkLastError switch, and Zig's error set previously collapsed all variants to a single unknown constant.

• Exposed html_to_markdown_rs::ConversionOptions at the Rust crate root, so callers configuring ExtractionConfig::html_options no longer need a direct dependency on the upstream crate, and made DocumentNode's text and node-type accessors public so DocumentStructure.nodes can be read as documented.

• Added FormatMetadata::html(), returning the HTML metadata when the variant is Html, matching the accessors already exposed for the other formats.

• Added an opt-in Pdfium PDF extraction backend behind the pdf-pdfium feature, selectable with PdfConfig.backend or --pdf-backend pdfium, providing page count, per-page text, and Info dictionary metadata. Its scope is deliberately narrower than the native engine — no table detection, layout integration, form fields, or OCR fallback — and every result carries a ProcessingWarning naming the gap. The feature is not part of full, so it reaches source builders only.

• Added a Scoop manifest published to the xberg-io/scoop-bucket on release, so the Windows CLI can be installed with scoop install xberg.

• Extraction now reports a ProcessingWarning when a document decodes lossily or degrades silently. Decode provenance is captured before mojibake cleanup strips the replacement characters that used to be the only evidence, and archive, AsciiDoc, WebVTT, XML, and plain-text extraction warn on replaced characters. Unresolved ODT image hrefs, unparseable styles.xml, collapsed repeated table cells, skipped LaTeX, Typst, RST, and Org includes, OPML without a body, links past the per-document URI cap, truncated XML, and words Tesseract failed to extract now warn instead of failing silently (#171, #133).

• A PDF page whose raster render comes back blank now falls back to OCR'ing the page's embedded image XObjects, and that recovery preserves the tables, formulas, LLM usage records, and image preprocessing metadata the backend produced instead of keeping only the text, with every recovered payload accounted against security_limits.

• Breaking (Python binding): ExtractionConfig and DoctorReport are now frozen dataclasses rather than TypedDicts, matching the 121 option types that were already dataclasses. Passing a plain dict or a JSON string as config still works — extract() coerces both — but an ExtractionConfig object no longer supports mapping operations, so config.get("chunking") and config["chunking"] = ... now raise AttributeError/TypeError, and the instance is immutable. Build a modified config with dataclasses.replace(config, chunking=...).

• PDF parsing no longer reports recoverable input at WARN. A missing embedded font, an object outside the xref table, an unreadable CFF version, and a reading-order fallback are ordinary properties of real PDFs rather than conditions an operator can act on; they are now TRACE (or DEBUG for strategy fallbacks), and each document emits a single DEBUG summary on the xberg_native_pdf::recovery target carrying the totals instead of one event per occurrence. Measured over a 4,000-document corpus this removed 4,012,488 of 4,014,206 log events, against which 44 genuine parse failures had been sitting at a ratio of about 1 in 91,000. ERROR behaviour is unchanged — it already corresponded one to one with documents that failed (#1547).

• Breaking (Rust source): validate_mime_type no longer accepts any value with an image/ prefix. It now parses the MIME type and requires exact membership in the supported-format registry, so unregistered vendor image subtypes such as image/x-custom-format are rejected as UnsupportedFormat instead of validating (#1511).

• Per-page OCR recognition-noise detail (fragmented-word ratio, word count, mean confidence) now reaches the page accept/reject decision and is emitted at DEBUG instead of being discarded one frame earlier. No threshold is gated on it yet; the blended stage score alone cannot discriminate noise pages.

• Breaking (Rust source): ExtractionConfig adds apply_notebook_cell_tags. Notebook extraction now honors MyST and Jupyter Book remove/hide cell tags by default; set the field to false to retain all saved cell content (#1538).

• Breaking (Rust source): OcrQualityThresholds adds discard_suspected_ocr_noise; exhaustive struct literals must set the field or use ..Default::default().

• Breaking: configuration deserialization now rejects unknown fields in nested Xberg configuration tables instead of silently ignoring misspelled settings.

• Breaking: PDF backend configuration now uses "native" and PdfBackend::Native instead of "pdf_oxide" and PdfBackend::PdfOxide. Update explicit configuration values; the default is unchanged.

• Breaking: EmbeddingModelType::Llm and RerankerModelType::Llm now carry their model name in the enum variant.

• Breaking: Formula.bbox and Formula.page are optional so formulas from formats without page geometry can be represented.

• Breaking: unknown multipart fields on extraction endpoints now return an error instead of being ignored.

• Chunk content now contains the exact source span; heading breadcrumbs are available separately.

• The CLI all feature now includes audio transcription.

• security_limits.max_pages now applies to presentations, Keynote, and multi-frame TIFF as well as PDF.

• create_client_with_credential_provider now returns ManagedClient, and an LLM concurrency limit of zero is rejected.

• Native PDF pages now expose their final per-page reading order.

• WebVTT cue timing is optional for blocks without a timing line.

• OpenDocument packages without content.xml now return an extraction error.

• CLI text output now includes the extraction envelope with warnings, timings, and metadata.

• CLI JSON output now reports peak resident memory.

• Windows builds now include the same supported feature set as other desktop builds.

• Breaking: Rust element identifiers now use String directly; the ElementId wrapper has been removed.

• Breaking: Public tuple fields for ranges, coordinates, dimensions, links, code blocks, and attributes now use named Rust structs and serialize as JSON objects. Legacy positional JSON arrays are still accepted when parsing, so payloads written by 1.0.x keep deserializing, but they are no longer emitted.

• Breaking: removed the duplicate xberg::llm::region_extractor::RegionKind; import xberg::RegionKind instead.

• Parsing and configuration deserialization now reject invalid region, redaction, and reranker values.

• Corrected and expanded installation, CLI, configuration, extraction, migration, integration, and cross-language API documentation.

• Corrected canonical MIME and extension routing for DBF, YAML, reStructuredText, Org, Typst, XHTML, Djot, JPEG 2000, HEIC/HEIF, MP4, and MPEG inputs.

• GeoJSON extraction now returns a bounded aggregate summary by default, including feature, geometry, property-key, position, and bounds metadata. Set geojson.include_full_coordinates = true to retain the complete document and coordinate arrays.

• quality_score now explicitly measures the cleanliness and readability of retained text, not extraction completeness; inspect processing_warnings for known partial or degraded results.

• The default security_limits.max_table_cells remains 100,000 aggregate cells per document; limit errors now explain how to raise it for trusted inputs or reduce the source table.

• TesseractConfig.language_model_ngram_on now defaults to true on both the PDF and standalone image OCR paths. Tesseract previously applied no penalty to output that does not look like a word of the target language, the dominant failure mode on scanned line art. Set the field to false to restore the previous behaviour.

• Tesseract Markdown-format OCR now drops hOCR lines whose dictionary-checkable words are more than 60% invalid, removing recognition noise such as OWATS DNDEVET while keeping labels like EXHIBIT and LEGEND. A line needs at least two checkable words to be scored, and the removed-line count is reported as a ProcessingWarning.

• Undecodable-text OCR routing is now decided per page rather than for the whole document, so a single unreadable page no longer sends every page of a PDF through OCR and discards good native text. The previous document-wide fallback still applies when page boundaries are unavailable or inconsistent.

• With max_threads unset the thread budget is min(num_cpus, 8) and now ceilings Rayon, ONNX Runtime intra-op threads, and batch workers alike. A cgroup CPU quota is honoured in place of the hardcoded 8 where one exists, and a host with more than 8 cores and no max_threads is warned once per process (#1392).

• PaddleOCR inference now uses the resolved thread budget instead of a hardcoded single thread. The session is serialized behind a mutex, so exactly one worker runs and can claim the whole budget without oversubscribing.

• Breaking: removed the inert ChunkingConfig::prepend_heading_context, breadcrumb_target, BreadcrumbTarget, and corresponding CLI and environment options; use chunk metadata or render_heading_breadcrumb when a retrieval index needs headings inline.
• Breaking: removed OutputFormat::Structured; use Plain for unrendered content or Json for a structured content tree.
• Breaking: removed ExtractedDocument.formatted_content from language bindings; use content or select the desired output format during extraction.
• Removed advertised support for troff, mdoc, POD, and DokuWiki because they did not have structural extractors.
• Removed fabricated OCR script_name and script_confidence values.
• Removed the unused public LanguageRegistry, BatchProcessor, object-pooling APIs, and unused tree-sitter re-exports.
• Removed the nonfunctional wasm-threads feature.
• Removed PDF writing, editing, building, and XFA conversion APIs from the native PDF crate; read-only XFA analysis remains available.
• Breaking: removed the inert Engine structured-policy, preset-resolver, LLM-client, and model-provider injection methods.
• Breaking: removed the inert transcription field from EnrichmentConfig; configure transcription during extraction instead.
• Breaking: embedding, reranking, sparse-embedding, late-interaction, and preset APIs are now exposed only when their required features are enabled.
• Breaking: core::batch_mode, core::formats, and core::io are now crate-private, and the public DocumentStructureBuilder has been removed.

• Fixed the Windows Ruby gem failing to build. xberg-libwpd's build script chose its zlib by operating system alone, so the gem's MinGW/UCRT toolchain was handed vcpkg's MSVC-built x64-windows-static-md archive and the link died with corrupt .drectve/ld returned 5. The vcpkg path is now taken only for genuinely MSVC targets; every other target links the static zlib libz-sys already builds from source.

• Fixed XbergLoader ignoring chunking and per-page splitting whenever the LangChain integration was given an ExtractionConfig object. Both settings were read only when the config was a dict, so after ExtractionConfig became a frozen dataclass the documented ExtractionConfig(pages=PageConfig(extract_pages=True)) and chunking=ChunkingConfig(...) forms silently produced one Document per file instead of one per page or chunk. The config is now read as an object or a mapping.

• Fixed a ruled troubleshooting page collapsing into one table, taking its section headings down with it as cell text. split_rows_by_text_positions subdivides a producer-drawn row band by the Y positions of the text inside it, and since the #1555 fix a candidate split was accepted only when EVERY resulting Y-cluster carried text in at least two columns, with the rejection all-or-nothing for the band. A band that mixes multi-column data rows with single-column lines -- a section heading, a lead-in, a wrapped continuation -- can never satisfy that, so one such line vetoed the split for the whole band and every line inside it became cell text. On one 56-page installation manual, six ~20 pt row bands became a single 522 pt table, the document went from 808 elements to 759, and four numbered headings disappeared from the outline. The band is now split once at least two of its clusters are independently evidenced, and each deficient cluster is resolved on its own terms: it folds into the cluster above only when it introduces no column that cluster left empty, which is the signature of a wrapped continuation. Anything else -- a heading, a lead-in -- stays a row of its own, one cell wide, which is what such a line inside a ruled band actually is. Two independently evidenced clusters are required rather than one because a single evidenced cluster can be coincidence, which is precisely the #1555 case (#1565).

• Fixed a word split across two touching PDF spans being rejoined with a space, so prijs extracted as pri js. The gap between the two spans measures 0.069 pt -- 0.008 em at 9 pt, against a 2.5 pt space glyph -- on an identical baseline at an identical font size, so no gap threshold produced the space: segments_need_space reached one of its unconditional return true branches first. SegmentData keeps only is_bold/is_italic/is_monospace and drops font_name, so a mid-word switch between two embedded subset fonts whose /FontDescriptors disagree on ForceBold, ItalicAngle or FixedPitch reads as a style change carrying no geometric signal at all. That is why the defect never reproduced against base-14 Helvetica, and why widening the gap to 2 pt changed nothing. A touching-spans guard now runs before those branches: two segments on the same baseline, at the same font size, with alphanumeric characters on both sides of the boundary and a gap under 0.025 em are one word and are concatenated. The guard can only join, never split, and it never fires across an explicitly drawn space. The table path needed the same test one stage earlier, in segments_to_words, because HocrWord is integer-rounded and cannot represent a sub-point gap by the time cell text is joined. Affects ordinary prose, not just tables: of 18 confirmed cases, 14 were NarrativeText, 3 ListItem and 3 Table (#1566).

• Fixed PDF table reconstruction dropping early rows when data-start inference classified more than two leading rows as headers. The two-row header cap is retained, but surplus inferred header rows are now demoted to data in source order instead of being discarded (#1558).

• Fixed native PDF top-to-bottom reading order splitting one visual table row at an absolute 3-point coordinate-band boundary, which could move an article number before its position and fuse the two identifiers. Visual rows now use an anchored, font-scaled tolerance, reconstructed lines restore left-to-right fragment order, and narrative assembly preserves a separator after a severe geometric backtrack (#1560).

• Fixed PDF dehyphenation treating inline run/style boundaries as visual line wraps. Suspended hyphens such as vracht- en verzendkosten are now preserved, while compounds genuinely split across different baselines are still rejoined (#1561).

• Fixed DOCX page attribution staying permanently low after Word omitted a rendered-page marker between vertically stacked inline images. The parser now conservatively infers missing breaks from each section's usable page height, including documents with different section geometries (#1559).

• Fixed DOCX DrawingML and VML text boxes dropping XML and numeric character references such as & and € from extracted text (#1562).

• Fixed OCR image decoding ignoring the caller's configured security_limits. Every OCR route — embedded images, Tesseract, PaddleOCR, and scanned PDF pages — decoded raw image bytes under a hardcoded SecurityLimits::default(), so raising ExtractionConfig::security_limits to accept a large scan still had it rejected at the OCR decode step. The configured limits now reach all four routes, and PaddleOCR also honors a per-call backend_options["security_limits"] override (#1554).

• Fixed a drawn PDF table row with a wrapped cell being shattered into extra rows. Splitting a row band by text Y-position now requires at least two columns to have independent text evidence for every candidate row before splitting; a band where only one column wraps to a second line now stays a single row (#1555).

• Fixed monospace font detection matching any font name containing "mono", misclassifying foundry names such as "Monotype Corsiva" as a monospace font and skewing the word-spacing heuristic and code-block detection that depend on it. "Monotype" is now excluded from the substring match, and the PDF text run buffer's separate ad hoc monospace check was replaced with the same shared helper.

• Fixed a standalone multi-line monospace paragraph not being recognized as a code block unless it had a consecutive monospace neighbor paragraph. A lone paragraph that already carries two or more monospace lines is now fenced as a code block on its own (#1557).

• Fixed PDF text extraction silently corrupting ordinary text. A contextual ligature-repair pass rewrote : to ti and an uppercase M between lowercase letters to tti on every element of every document, mangling identifiers, ratios, times, URLs, and units such as nM (for example aMb became attib). The repair was introduced for European PDFs that encode ligature glyphs at ASCII code points, but it was gated at the time on a per-font broken-CMap signal from pdfium's has_unicode_map_error(). That gate was lost when pdfium was removed as a backend and was never ported to pdf_oxide, leaving the rewrite running unconditionally. Both substitutions are removed; they can only return alongside a real document-level evidence gate (#1556).

• Fixed optional fields in the Python and PHP bindings rejecting payloads that omit them. The generated mirror structs lost their #[serde(default)] attributes, so deserializing a document whose JSON left an optional field out failed instead of falling back to the default.

• Fixed legacy .doc headings being guessed from line length rather than read from the document's own styles. A paragraph styled heading 1..heading 9 — directly or through a custom style derived from one, such as TOC Heading — now becomes a Heading at that level, instead of every detected heading being a level 2. Documents that apply no heading style keep the previous shape-based detection, because roughly half the test corpus styles its headings as bold Normal and would otherwise lose every one; the choice is made per document, not per paragraph. A heading-styled paragraph that is also list-bound stays a ListItem, matching how the DOCX path treats w:numPr (#1553).

• Fixed legacy .doc automatic list numbering being dropped entirely: a paragraph Word numbers through its list tables arrived as prose, indistinguishable from an unnumbered sentence, while the DOCX path emitted a ListItem for the same construct. Auto-numbered paragraphs now arrive as ListItems inside an ordered or bulleted list container, with their nesting depth, matching the DOCX path. The number Word paints (1.1, a.) is still not rendered — recovering it needs list-table counter state — so a document mixing automatic and hand-typed numbering shows the typed numbers as text and the automatic ones as list structure (#1550).

• Fixed legacy .doc elements being split on blank lines rather than on Word's paragraph marks, which merged every pair of consecutive paragraphs not separated by a blank line into a single element. One corpus letter returned its entire ten-paragraph body as one element. Word97 and later documents now emit one element per Word paragraph, matching what the DOCX path does with w:p. This changes element boundaries, counts and indices for most .doc documents, and alters content line spacing accordingly; consumers keying on element position will see the difference. Word 6/95 documents and those falling back to contiguous text extraction keep the previous blank-line behaviour, because they carry no paragraph properties to use.

• Fixed legacy .doc extraction reading fcClx from FibRgFcLcb97 pair 66 — an obsolete field Word writes as zero — instead of pair 33, so the piece table was never walked for any document and extraction always fell back to reading reserved5/reserved6, bytes [MS-DOC] requires a reader to ignore. Where those bytes disagreed with the real text start, whole documents were decoded as UTF-16LE and returned as glued CJK-looking code points; multi-piece and fast-saved documents could not be assembled at all. Footnote, header/footer, comment, and text-box subdocument text now also reaches the output for these files (#1551).

• Fixed the Elixir NIF's vendored Cargo.lock, shipped in the Hex package, pinning tree-sitter-language-pack 1.15.12 while the crate requires 1.16.1 — a source build of the NIF with --locked could not resolve. This affects anyone whose platform has no precompiled artifact and therefore builds from source.

• Fixed a DOCX table cell spanning several grid columns (w:gridSpan) or rows (w:vMerge) being returned once per covered column and again for every covered row, so a cell merged across 4 columns and 3 rows came back 12 times in result.tables[].cells, result.tables[].markdown, and result.content alike — a 39 KB document could extract to 232 KB. A merged/spanned cell's text is now written once, at its origin, with the columns and rows it covers left blank. This also fixes a DOCX header or footer table with a merged cell shifting every following cell one column to the left (#1549).

• Fixed PDF render diagnostics matching a captured engine warning against a hardcoded message substring to decide whether it meant a glyph actually failed to paint. The message it was built to exclude no longer reaches this capture at all (it moved to TRACE under #1547), so the match could only ever misfire: a future warning whose text happened to share that substring would have been silently dropped instead of surfacing as a ProcessingWarning. Every captured warning is now reported (#1548).

• Fixed a PDF page that places a statistics table beside a prose column being emitted in full-width Y order, which spliced the prose apart mid-sentence (more likely to be aged 35Female 51.5 ...) and welded the table's two label/value panels together on every row. The table region is now emitted whole, in row order, ahead of the prose column, and a repeated panel is emitted panel by panel (#1545).

• Fixed PDF text coming back scrambled when a short Tj run sat between two TJ arrays: the run was emitted at an earlier run's stale position and sorted into the wrong place, so within a period ... after conclusion extracted as wincthin a period ... after co lusion. Every text-showing boundary operator closed the pending run except TJ (#1544).

• Fixed every image in a DOCX reporting page_number 1 regardless of the page it sits on. The page was resolved by searching rendered Markdown for a per-image placeholder that is never written -- every drawing renders to the same link target -- so the lookup always missed. Page numbers now come from the parsed element order (#1546).

• Fixed an author's hyphen being deleted when it fell at a line break, so price- + determining joined as pricedetermining. A hyphen written mid-line elsewhere in the same document is now treated as evidence that the compound is real and its hyphen is kept. Compounds that appear only broken, with no such occurrence anywhere in the document, are still joined without the hyphen (#1543).

• Fixed OCR backends registered through register_ocr_backend being rejected before extraction started: configuration validation checked the backend name against the built-in list only, which made every custom plugin OCR backend unusable once validation was wired into extract and extract_batch.

• Fixed the native C FFI library shipping without eleven features the crate advertises, so the Java, Go, C#, Swift, Zig, and C bindings had no summarization, translation, analysis, HEIC, captioning, ML redaction, or static-embedding support. The desktop dependency hand-maintained a feature list that had drifted from full; a regression test now fails on any future omission.

• Fixed HTML pages fetched over HTTP(S) losing every format-specific metadata field: results were reported as text/html while metadata.format stayed empty, because the extraction ran over the crawler's pre-rendered Markdown and never reached the HTML extractor. Title, headings, Open Graph, Twitter card, links, and structured data are now recovered from the page HTML.

• Fixed pdf_options.hierarchy.enabled silently producing no hierarchy: headings were detected and then discarded unless the caller also set the unrelated pages.extract_pages. Requesting the heading hierarchy now enables the per-page tracking it requires.

• Fixed the bundled Tesseract build failing to configure on Windows when the MSVC developer environment is not present, which broke building Xberg from source with the default OCR features.

• Fixed URL extraction reporting internally converted HTML pages as text/markdown; results now retain a validated, canonical source MIME type.

• Fixed clear_post_processors stopping at the first failed shutdown hook and permanently removing enabled built-ins; it now attempts every shutdown, returns the first error, and restores built-ins before the next post-processed extraction while custom processors remain cleared.

• Fixed VLM concurrency limits increasing concurrent local OCR work and raster memory use (#1465).

• Fixed structured extraction forcing every caller schema to JSON Schema Draft 2020-12; validation now honors the schema's declared draft while keeping external reference resolution offline (#1539).

• Fixed hybrid PDF OCR dropping surrounding prose when a table-bearing bare-text page was restructured alongside geometry-backed pages.

• Fixed automatic PDF OCR fallback reporting an empty success when OCR failed and no native text remained; recoverable failures still return available native text with a warning.

• Fixed degraded VLM fallback output replacing denser OCR text, while abstaining from the density comparison for short text and non-space-delimited CJK or kana content.

• Fixed Windows source and Ruby package builds failing on stable Rust while validating the identity of staged Tesseract source directories.

• Fixed GCC 12+ WordPerfect builds by adding the standard header that declares size_t before compiling the pinned libwpd source.

• Fixed Ruby source-package installation by aligning the Gemfile and lockfile with the gemspec's supported rb_sys range.

• Fixed generated Ruby development commands so Bundler and its tools use the active Ruby interpreter, avoiding native-extension ABI conflicts on systems with multiple Ruby versions.

• Fixed generated Python optional constructor arguments so Pyrefly receives precise keyword types without unused helper declarations.

• Fixed generated Dart tests for nested tagged unions, nullable payloads, and Flutter Rust Bridge tuple accessors; added e2e analyzer coverage and refreshed the Dart lock file to the generated Flutter Rust Bridge version.

• Fixed compressed image inputs with oversized declared dimensions exhausting memory during OCR, layout and QR detection, image classification, re-encoding, HEIF conversion, or structured-image rasterization; decoded allocations now obey security_limits.max_content_size and are rejected from the image header before pixel decode.

• Fixed PDF OCR fallback being suppressed for image-only pages when dot leaders or other non-textual native content pushed the document below the alphanumeric-ratio threshold.

• Fixed process-global native PDF font-cache collisions that made glyph spacing, geometry, and batch output depend on document order and concurrency when fonts used indirect width tables.

• Fixed Markdown OCR metadata so word counts and confidence statistics describe only text retained after dictionary filtering; fully filtered output now reports zero words and omits confidence quantiles.

• Fixed repeated bold PDF presenter labels and same-row legend keys being promoted to headings, which could invert document hierarchy and fragment retrieval chunks.

• Fixed PDF OCR so fragmented, low-confidence, and dictionary-suspect non-empty text is retained with a processing warning by default instead of silently emptying pages. Set ocr.quality_thresholds.discard_suspected_ocr_noise = true (or the equivalent pipeline quality threshold) to opt into the previous destructive filtering behavior.

• Fixed runtime crashes in system-linked Tesseract OCR builds by linking the required native exception-safety shim.

• Fixed xberg batch so mixed-success runs emit every successful document and every attributed per-input error before returning a nonzero status; JSON and TOON timing slots remain aligned with inputs, and TOON now uses the documented batch envelope.

• Fixed xberg extract --ocr false so it authoritatively disables implicit OCR fallback, overrides conflicting loaded OCR routing, and rejects contradictory OCR flags.

• Fixed Tesseract preprocessing so deskew, denoise, contrast enhancement, and Otsu, adaptive, and Sauvola binarization settings transform the OCR raster on native and WebAssembly backends; none (with off as an alias) preserves unthresholded grayscale when deskew is disabled, sparse receipt-image fallback and faint colored text no longer lose content to global thresholding, dark labels over bright map fills still receive Otsu preprocessing without isolated or clustered dark artifacts triggering it, and WebAssembly Tesseract now rejects images exceeding 4096 × 4096 pixels before decoding.

• Fixed OCR measurement tooling so line-filter comparisons score the intended ground-truth lines and report filtering regressions accurately.

• Fixed the OpenAPI document's dangling Djot attribute reference so schema validators and client generators can resolve every advertised component (#1505).

• XML and JSON content with unsupported specialized extensions now routes through the supported generic extractor instead of failing MIME validation (#1507).

• File extraction now falls back to bounded content sniffing when a path has an unknown or missing extension (#1506).

• Explicit application/octet-stream hints now trigger configured MIME detection instead of being treated as an authoritative document type.

• Fixed documentation-snippet fixtures that named non-existent result fields, which made the generated snippets silently drop the affected presentation block: element content is now text, table rows is now cells, and the result paths keywords, structured_data, and document_structure are now extracted_keywords, structured_output, and document.

• Fixed EPUB extraction for text/html spine items, named entities, declared non-UTF-8 encodings, navigation documents, SVG fallbacks, nested tables, MathML, headings, images, and malformed HTML (#1486, #1488-#1494).

• EPUB extraction now preserves usable chapters when another spine item fails and reports per-item warnings instead of failing the whole document (#1491).

• Fixed EPUB metadata, EPUB 2/3 cover selection, DRM detection, and font-obfuscation handling (#1492, #1494).

• Fixed PDF OCR and rendering for highly compressed scans, CCITT images, CFF fonts, maximum-size font tables, malformed embedded fonts, rotated content, missing glyph warnings, and concurrent Pdfium extraction.

• Fixed native PDF tracing so corrupt optional content is reported as a recoverable warning, while mandatory cross-reference failures emit a single operation-boundary error without changing the returned error type.

• Fixed annotation-only PDFs so visible FreeText content is recovered into page-aware document text, including when OCR replaces the page text, without exposing hidden, transparent, cropped, or disabled annotations when annotation extraction is off.

• Fixed the Swift package manifest so SwiftPM no longer warns about a nonexistent target-relative license file.

• Fixed scanned PDF extraction so CCITT parameters align with their filter in multi-filter streams, referenced JBIG2 image masks are available to OCR, and stencil-mask polarity renders text as opaque.

• Fixed PDF reading order for dense two-column layouts, hanging clause numbers, split list markers, and modest font-size changes on one baseline.

• Fixed PDF heading recovery for repeated bold section titles set at body font size while retaining short bold labels, presenter attributions, and calendar legends as body text (#1513).

• Fixed PDF table extraction so multi-word cells, rule-less prose regions, OCR-derived tables, and page-local table failures are handled correctly (#688, #1358, #1542).

• Fixed PDF Markdown and Djot output so native text is retained when structured conversion is incomplete.

• Fixed PDF configuration so metadata suppression and header/footer settings are honored by every backend; invalid or unsupported PDF and OCR settings now return configuration errors.

• Fixed OCR-backed PDFs so filtering, confidence thresholds, hierarchy, tables, formulas, lists, bounding boxes, page boundaries, and partial page results are preserved consistently across output formats and OCR backends (#1444).

• Fixed Tesseract caching, configuration, preprocessing, page segmentation, and font-size extraction.

• Tesseract Markdown extraction now reports a ProcessingWarning when dictionary filtering removes physical text lines, including the number removed.

• OCR element hierarchy output now honors build_hierarchy and contains only resolvable parent references.

• Fixed Sceptre and PaddleOCR line grouping, region ordering, per-page resizing, table validation, and font-size reporting.

• Fixed DOCX extraction for nested tables, VML images, text boxes, comments, fields, headings, hyperlinks, headers, footers, table-of-contents entries, nested lists, and page attribution; element output now preserves explicit page breaks and single-page documents report page metadata consistently (#1452, #1460, #1503).

• Fixed PPTX extraction for malformed relationships, nested image paths, equations, fallback shapes, comments, metadata, and security limits.

• Fixed spreadsheet extraction for hyperlinks, formulas, names, comments, hidden state, dates, and OpenDocument metadata.

• Fixed ODT, ODP, iWork, HWP, DBF, RTF, email, and PST extraction across nested content, metadata, binary data, folder traversal, and repeated text.

• Fixed Markdown, MDX, RST, HTML, DocBook, JATS, FictionBook, Djot, Org, YAML frontmatter, and Jupyter extraction so supported structure and content are retained.

• Fixed result.elements so headings report their level (metadata.additional["heading_level"]) instead of every ##-###### heading collapsing into indistinguishable Heading elements with empty metadata; result.document.nodes already carried the level correctly (#1504).

• Fixed CSV parsing for stray quotes and archive extraction order.

• Fixed MIME routing so HTML is detected before the generic XML fallback and supported-format lists reflect the active extractor registry.

• Fixed post-processing, chunking, enrichment, translation, NER, QR codes, captions, and caching so extracted structure is preserved consistently.

• Fixed chunking presets so standalone and pipeline APIs apply the documented size and overlap while preserving unrelated chunking settings.

• Fixed extraction timeout handling so timed-out work is cancelled.

• Fixed configuration merging so changing one CLI option no longer erases sibling settings.

• Fixed multipart API extraction to accept json and doctags values for output_format.

• Fixed cache keys to reflect only settings that affect the corresponding extraction or OCR result.

• Fixed model caching so OCR, embedding, and reranking settings no longer reuse incompatible models.

• Fixed Node.js native-library loading, Swift iOS resolution, Windows DirectML packaging, and cargo install xberg-cli (#1456).

• Fixed Docker image builds and reduced the CLI image to runtime dependencies.

• Fixed API and packaging defects in the Python, PHP, Dart, Go, Java, C#, Kotlin, Elixir, Ruby, Zig, and C packages.

• Fixed Windows wheel and gem packaging, manylinux compatibility, musl smoke tests, and dynamic Tesseract builds (#1495, #1497).

• Fixed archive and ZIP validation for small compressed entries and impossible declared sizes (#1496).

• Fixed batch extraction so configured caches are used and progress callbacks report completed items.

• Fixed extraction configuration validation so invalid nested values, including OCR quality and scanned-page thresholds, are rejected consistently by every public entry point.

• Fixed error classification so callers can distinguish all documented extraction failure categories.

• Built-in path and byte extraction now always reports a recognized extraction_method; custom extractors retain explicit provenance and otherwise leave it unspecified.

• Fixed owned document classification so detected labels are written back to the returned document.

• Fixed ContentFilterConfig.include_watermarks so enabling it retains watermark content.

• Fixed JsonExtractionConfig.flatten_nested_objects so disabling it preserves nested objects instead of flattening them.

• Fixed standalone-image and OCR-backed PDF results so preprocessing scale, dimensions, and DPI are retained.

• Fixed Candle OCR configuration so supported backend options are validated and applied.

• Fixed PaddleOCR-VL so the task selected when constructing the backend is honored unless a request explicitly overrides it.

• Fixed keyword extraction so invalid n-gram ranges return an error instead of silently producing empty results.

• Fixed builds that enable only the api or mcp feature.

• Fixed the excel-wasm feature so spreadsheet extraction builds for WebAssembly.

• Fixed WebAssembly configuration so unsupported managed credential providers are rejected explicitly.

• Fixed the Swift package failing to link on Linux. Package.swift linked libxberg_ffi.a alongside libxberg_swift.a, but the Swift static library already folds the entire compiled xberg-ffi crate in, so every Rust core, std, and alloc symbol existed twice and the linker reported hundreds of duplicate symbols. It also never asked for ONNX Runtime, leaving OrtGetApiBase undefined.

• Fixed the public clear_ocr_backends() and clear_renderers() leaving their process-global registries permanently empty. After clear_ocr_backends() every later extraction failed with "No available OCR backends"; after clear_renderers() the Custom output-format path silently downgraded DOT renders to plain text for the life of the process. Both now re-seed the built-ins non-destructively, keeping user-registered entries.

• Fixed nested lists rendering as flat, blank-line-separated bullets in pages[N].content: container list markers are never page-tagged, so a page subset dropped them and every item was rewrapped in its own single-item list. Also fixed figure alt text being dropped whenever a caption was present, the VLM OCR probe reporting availability without checking credentials, and the PDF margin filter judging rotated text runs by baseline origin.

• Fixed HEIC-enabled builds requiring a libheif newer than current stable distributions ship. The prebuilt artifacts link libheif dynamically and were built against 1.21 APIs, so the PHP extension failed to load on Debian 13 with undefined symbol: heif_image_get_plane_readonly2. The floor is now 1.19, with version-gated fallbacks (#1541).

• Fixed PDF text collapsing on itself when a font's /Widths array declares 0 for an ordinary glyph. Extraction now falls back to the embedded font's own advance for such codes, while an explicit TJ displacement stays authoritative and genuine zero-width combining marks remain overlays.

• Fixed automatic PDF OCR replacing a page's native text with a substantially poorer recognition. OCR output for a page whose native text was independently judged healthy is now rejected when it retains under half that page's alphanumeric characters.

• Fixed OCR of a single detached page image being attributed to page 1. Local image indices were used as document page numbers, so warnings named the wrong page and the rejected-page filter discarded OCR elements, tables, and formulas belonging to a different page than the one rejected.

• Fixed XML extraction narrowing element depth to u8 before clamping, so an element nested more than 255 levels deep wrapped to a low heading level in release builds and panicked in debug builds, before the configured max_xml_depth limit ever applied (#1474).

• Fixed PDF XMP metadata losing text fragments split around an entity boundary: named and numeric XML references in XMP scalar and sequence values are preserved instead of the surrounding text being truncated (#1475).

• Fixed image-level OCR running again over a page-sized PDF XObject on a page whose native text had already been extracted, which duplicated the page's content and paid for a second OCR pass (#1479).

• Fixed the musl (Alpine) native artifacts failing to load. The published Java, C#, Zig, C, and Elixir artifacts shipped without ONNX Runtime's transitive closure — libprotobuf-lite, the libabsl_* set, libre2, and libicu. Both musl images now vendor the full ldd closure and hard-fail the build if anything is unresolved. A host runtime that links libstdc++ itself still needs libstdc++ 15 or newer in the process, because a bundled copy cannot win once the soname is already mapped.

• Fixed a DOCX or PPTX relationship targeting ../media/image1.png — the ordinary OPC shape for an image at the package root — being rejected by the traversal check and dropped, so the image went missing from extraction. Container-relative names now resolve boundary-relative.

• Fixed OCR of rendered PDF pages assuming a 72 DPI raster when pages render at 150 DPI, so DPI normalisation computed a 2.48x upscale, hit the dimension clamp, and reported a resolution hint of 179 for what was really a 372 DPI image. Also fixed image DPI normalisation being skipped entirely in candle-backend and VLM-only builds.

• Fixed layout detection marking real figure and drawing text as page furniture, which the renderer then discarded, so labels such as SITE PLAN and LEGEND disappeared from scanned documents. A Picture hint now means a figure was detected, not that the text is decoration, and furniture hints only match short text.

• Fixed the Docling-compatible endpoint discarding OpenWebUI's extraction parameters. OpenWebUI sends one form field per key rather than a JSON blob, so settings made in its admin UI produced identical output with or without them (#1462).

• Fixed CLI flags being silently discarded. --ocr-backend, --ocr-language, --ocr-auto-rotate, and --ocr-backend-options were dropped unless --ocr true was also passed, so --ocr-scanned-pages --ocr-backend sceptre ran Tesseract with no error; --ocr-scanned-pages alone returned an empty document at exit status 0; and --chunk-size was a no-op without --chunk true.

• Fixed legacy .doc extraction emitting every field's instruction — its URL, switches, and screen-tips — verbatim as prose, and the non-breaking hyphen being dropped with the other control characters, fusing twenty-one into twentyone.

• Fixed paragraph grouping only breaking when a line starts a numbered section and never when the previous line was one, so a subsection heading followed by unnumbered lines at the same size and weight was merged into the following prose (#1467). Consecutive numbered headings are likewise no longer welded into a single paragraph (#1386).

• Fixed the PDF pipeline stripping a list item's printed marker and discarding it, leaving renderers to synthesize a position, so a document whose clauses are cross-referenced by their printed label was renumbered — B. rendering as 1. and (a) as 1..

• Fixed candle-trocr accepting a whole page and returning invented text. TrOCR is trained on single cropped lines and force-resizes any input, so a multi-page document exited successfully with text appearing nowhere in it. Input taller than a plausible line crop is now rejected.

• Fixed inline <svg> elements being discarded during HTML extraction even with extract_images enabled (#745).

• Fixed an explicitly requested GPU execution provider silently running on CPU. is_available() reports only compile-time support and ORT's session builder defaults to not erroring on failure, so an explicit CUDA, TensorRT, or CoreML request that failed to load was swallowed. Explicit requests now fail; Auto keeps its silent fallback.

• Fixed DOCX documents with legacy VML picture markup being rejected as NestingTooDeep, and the inverse hole where content inside drawings, table property helpers, the table grid, and streaming section properties was never measured against the depth cap at all. A flat 600-row table of real depth 8 previously leaked over a thousand levels and was rejected outright (#1395).

• Fixed XBERG_LLM_API_KEY and XBERG_LLM_BASE_URL fabricating a structured-extraction config with an empty model and schema, so any deployment that merely had an LLM key in its environment ran the post-processor on every document and failed every one (#1421).

• Fixed two PDF paths aborting or failing the whole request: a /ModDate whose raw bytes decode to a replacement character sliced a str off a char boundary and panicked, which across the Go FFI boundary aborts the process before any catch_unwind frame is consulted; and a rasterizer panic on a page with damaged content streams unwound through the async boundary and lost every other page's text (#1422, #1408).

• Fixed keyword extraction panicking on a language hint whose first character is multi-byte.

• Fixed legacy .ppt slide numbering and image extraction. Slide numbers were the ordinal of a text block in a joined string, so a trailing paragraph mark cut one slide into several; they now come from the slide containers in persist order. The OLE /Pictures stream was never opened, so .ppt extraction never produced an image (#1418, #1417).

• Fixed PPTX slides without a title losing their page number (#1413).

• Fixed URL extraction reporting no crawled URLs, because the result field is no longer populated upstream. The URLs are now derived from the crawled pages, deduped in first-seen order.

• Fixed PDF page-number stripping deleting real table data. The decision was made from one paragraph's text, so any short numeric cell matched; it now requires a margin band, a stable horizontal slot across pages, and a progressive sequence to agree (#1411).

• Fixed PDF paragraph breaks never being detected on a normally-set page, so a whole memo — date, salutation, body, sign-off — came back as one line. The vertical advance is now compared against the body leading, which is scale-free.

• Fixed detected PDF tables being injected on top of native text that already contained them, so the same content was rendered twice.

• Fixed non-HTML raw blocks being written verbatim into styled HTML output. ODP speaker notes and master-page text, Org source, script and style bodies, and Djot raw blocks all reached the page unescaped, so any < in them corrupted the document structure.

• Fixed the PyPI xberg-cli wheels shipping without their native libraries. The build hook force-included siblings with a macOS-only glob, so every Linux shared object staged beside the binary was dropped, and the musl wheel shipped only the launcher script. An incomplete platform payload now fails the build instead of publishing a wheel that installs and cannot run.

• Fixed OCR'd PDF pages reporting bounding boxes in raster pixels while digital pages report PDF points, with nothing in the response distinguishing the two spaces. Node, hierarchy block, chunk page span, and table bounding boxes are now converted to page points with a bottom-left origin (#1423).

• Fixed OCR on pages carrying a /Rotate entry. Backends now declare how they cope with a rotated raster, so a backend that requires an upright page is handed one with its geometry mapped back, and PaddleOCR receives the page rotation as a sort key. Auto-rotation composes with the page hint instead of double-correcting it.

• Fixed PDF text and tables on rotated pages. Rotated-text repair reconstructs the reading frame but only when rotated spans are at least 20% of a page's characters, so a single rotated caption no longer costs the upright majority of the page its whitespace structure, and heuristic table reconstruction clusters cells on the table's own axes rather than raw page space (#1358).

• Fixed the OpenAPI document omitting types that client generators need: second-order nested component schemas are now registered, along with the PDF, office, and transcription schema groups and the 415 and 429 responses the extraction endpoints can return (#1424).

• Fixed code_intelligence being hardcoded to None, so the documented metrics, imports and exports, comments, docstrings, symbols, and diagnostics never reached callers (#259).

• Fixed Whisper timestamp tokens leaking into transcripts as literal text. They are not marked special in the tokenizer vocabulary, so they survived decoding; they are now paired into segments, emitting one paragraph per segment with start and end times.

• Fixed cargo add xberg --features full failing to link on Windows MSVC, where a transitive build script forces /MT while Rust defaults to /MD, killing the build with LNK2038 (#1389).

• Fixed show_download_progress having no readers anywhere on the embedding, sparse-embedding, reranker, and late-interaction model configs, so the documented option did nothing.

• Fixed split_and_extract rebuilding each segment from a handful of fields, dropping keywords, entities, summaries, chunks, warnings, and the rest of the enrichment that extraction produced, and an off-by-one in the chunk image-index remap that pointed chunks at the wrong image.

• Fixed target_dpi, max_image_dimension, auto_adjust_dpi, min_dpi, and max_dpi having no readers: every preprocessing config was built with defaults, so these settings were dropped (#209).

• Fixed declared telemetry that never emitted. The cache-hit, cache-miss, and batch instruments were declared but never recorded, and the pipeline and batch operations, five of the eight pipeline stage spans, and the extractor-priority and batch attributes were likewise never recorded, so filtering on them returned nothing (#332, #282).

• Fixed an injected cache backend never being consulted and ProgressSink::emit having no caller on single extraction; extract_batch was already correct. A bytes-input cache hit now short-circuits extraction and coarse start, complete, error, and cache-hit events are emitted.

• Fixed renderer output completeness: JSON silently dropped page breaks, footnote references and definitions, citations, slides, definition terms, admonitions, raw blocks, and metadata blocks through a catch-all arm; styled HTML opened a section for each slide that was never closed and never rendered the slide title; and formulas rendered as preformatted code, which KaTeX and MathJax cannot pick up, and are now delimited display math.

• Fixed footnote definitions never appearing in JSON output, and a definition present in the document but never referenced being dropped from rendered output entirely (#68).

• Fixed plugin-produced documents losing content at the bridge. The conversion into the internal document dropped uris, children, annotations, processing_warnings, llm_usage, pages, and ocr_elements; native renderers reached through the public entry point emitted an empty shell; and pre_rendered_content was ignored for HTML and JSON output.

• Fixed CRLF documents collapsing into a single paragraph. Ten call sites split paragraphs on a bare double newline without normalising line endings first, affecting email and PST bodies, OCR backend output, plain text, and Djot conversion (#227).

• Fixed MIME aliases that were advertised as supported and then failed as UnsupportedFormat, because the registry looks up by exact string with no alias resolution. application/wordperfect, application/x-quarto, and four audio and video transcription aliases now route to the same extractor as their canonical type.

• Fixed three internal OCR plumbing keys being copied into user-visible document metadata.

• Bounded DOCX image and iWork archive member reads by the member's declared uncompressed size instead of trusting that declaration. A crafted document could forge a small declared size in the ZIP central directory while carrying a deflate stream that inflated to multiple gigabytes, exhausting memory during DOCX image extraction (images.extract_images) or .pages/.numbers/ .key extraction. Reported by Syed Anas Mohiuddin (GHSA-85w9-wqcq-x48r).

• Pinned downloaded Tesseract, Leptonica, and English tessdata inputs to immutable revisions with verified sizes and SHA-256 digests, race-safe content-addressed caches, private build directories, and bounded fail-closed archive extraction.

• Structured extraction now resolves caller-provided JSON Schemas strictly offline and rejects external HTTP and file references without performing I/O.

• REST and MCP requests can no longer override LLM credentials, provider registrations, or other server-controlled settings.

• Hardened ZIP accounting against overflow, impossible sizes, and compression-ratio bypasses.

• Hardened DOCX, PPTX, and EPUB relationship resolution against container traversal, malformed UTF-8, NUL bytes, drive-letter paths, UNC paths, and symlink escapes.

• Added bounded EPUB traversal and retained-content accounting to prevent resource-limit bypasses.

• Cache namespaces are validated before directories are created.

• Redaction now reports only content that was actually removed, never exposes pre-redaction element text, and rejects invalid strategies instead of silently falling back to masking.

• Hardened the native PDF engine against crafted documents that abort or hang the host process. A self-referencing /Names /EmbeddedFiles tree and deeply nested array or dictionary brackets each recursed until the stack overflowed, which is an abort no catch_unwind can contain; a negative /W element in an xref stream, a reversed bfrange, a non-hex ToUnicode destination, an all-NaN font-size set, and unchecked /Widthx/Height, /N, and /VerticesPerRow products each panicked or allocated without bound; and decode_stream_with_params, the entry point every production call site uses, applied no ratio or size guard at all. All were reachable from extract_bytes under default configuration.

• Bounded every ZIP, TAR, and 7z member read against SecurityLimits rather than against the size the archive declares for itself, since a declared uncompressed size is not a bound and the aggregate check previously ran only after the member was fully resident. Covers generic archives, ODT, ODP, EPUB, HWPX, PPTX, XLSX, and OOXML embedded objects, and adds the compression-ratio and aggregate-size validation that PPTX, XLSX, and DOCX were missing. A nested ZIP no longer overflows the stack.

• Clamped or rejected document-declared counts that reached an allocation or a slice unchecked: HWP table row and column counts, HTML and EPUB colspan/rowspan, DOCX w:ilvl, w:gridSpan, and w:outlineLvl, PPTX a:pPr lvl, RST simple-table column ranges, JATS date-type, EPUB link-label offsets, PPTX relationship targets, and the hOCR parser's and annotated-text renderer's byte-offset slices. Each was an out-of-bounds or char-boundary panic, or an allocation abort, on ordinary untrusted input.

• security_limits.max_files_in_archive is now enforced by every OOXML container. XLSX never checked it, DOCX enforced a hardcoded 10,000-entry cap instead of the configured one, PPTX had no entry check at all, and embedded-object extraction walked embeddings uncapped (#1449).

• EPUB packaging XML now counts real OPF nesting depth against the configured limit and accepts legacy DTD declarations without resolving external or amplified entities, so a crafted package can neither bypass the depth budget nor pull in outside content (#1477, #1478).

• Native PDF tracing no longer carries document content. Decoded page text was emitted verbatim at TRACE, embedded font names appeared in trace events and in the glyph-drop ProcessingWarning message, and parser, xref, and recovery failures were logged by formatting the underlying error string. Failure paths now emit a structured error_code with an optional byte error_offset, and font names are redacted in the warning text.

• Bounded the native PDF reader's internal caches so a malformed or hostile document cannot grow them without limit: the object-stream cache evicts to a byte budget and rejects oversized entries, font identity hashing stops at a byte budget and a reference-depth cap (both recorded in the hash so distinct fonts stay distinct), and the xref recovery-marker set is capped.

• OpenWebUI-compatible endpoints (PUT /process and POST /v1/convert/file) now honor extraction configuration. They previously cloned the server default, forced Markdown output, and ignored all inbound parameters, so configuration passed through OpenWebUI had no effect. They now use the server's configured defaults as the base and merge a per-request config — a multipart config/parameters field, or the X-Config header — matching the /extract endpoint, keeping Markdown as the default only when neither the server config nor the request selects a format.
• Image captioning is now included in the official Docker images (--features all), and the server emits a ProcessingWarning when a captioning config is supplied but the feature is compiled out, instead of silently doing nothing (#1382).
• Release builds no longer check out the test_documents benchmark submodule, so a benchmark-only submodule update can no longer fail every publish build and ship a release with no assets (#1380).

• Embedded-image captioning now runs with bounded concurrency (mirroring the image-OCR path) instead of one VLM request at a time, reducing wall-clock time on image-heavy documents (#1378).

• OCR-backed PDF extraction now keeps consecutive Tesseract paragraphs grouped within their shared hOCR text area instead of splitting them, including pages replaced by mixed native/OCR extraction. This is paragraph/block grouping only; font-clustering headings and list-marker detection for OCR-backed pages are tracked separately (see Unreleased).
• PDF table reconstruction now rejects sparse, short-wide contact blocks that were previously misclassified as tables.
• Standalone-image Tesseract OCR now defaults to sparse-text segmentation, while cropped layout regions use single-block segmentation and explicit user settings remain unchanged. Vertical language packs such as Japanese (jpn_vert) use vertical-block segmentation.
• Standalone image extraction now reports successful OCR through metadata.ocr_used and the OCR extraction method, including layout-aware OCR results.
• Tesseract now applies its default image preprocessing only to clean, near-white document pages; shadowed receipts and photographic images keep their source pixels, avoiding quality loss from destructive DPI upscaling, background normalization, sharpening, and grayscale conversion.
• Sparse, low-confidence standalone Tesseract results now retry the previous automatic page segmentation with explicit preprocessing and use it only when word confidence is consistently strong, recovering difficult receipts and scene text without replacing reliable sparse output.
• CSV and TSV plaintext now use the canonical table renderer instead of lossy Row N and header-value prose.
• Extracted EML and MSG attachment text is now included in the parent document while the structured attachment children remain available.
• DOCX extraction now emits a tab character for an in-run <w:tab/> instead of dropping it, so tab-separated fields — most visibly Word table-of-contents rows — no longer weld adjacent words together (Alpha<tab>Beta was extracted as AlphaBeta). Tab-stop definitions remain invisible. (#1377)
• The Swift package builds and publishes again. The cross-compiled desktop xberg-ffi dependency no longer pulls in HEIC (libheif-sys, which has no cross-compile support) or the Candle OCR backends, which had broken Swift package publishing in 1.0.12.
• The NuGet runtime packages for macOS and Linux (osx-x64, osx-arm64, linux-x64, linux-arm64) now publish at the current version instead of being stuck at an older one; previously only the Windows runtime package was updated. (#1375)
• The public in-browser (WASM) demo now attributes its file-size limit to the browser sandbox and points to the CLI and API for large or multi-page documents, instead of implying the document itself is at fault. (#1376)

• The xberg mcp extract and extract_batch tools no longer emit structured output that fails their own declared output schema. The schema required errors and the crawl_* fields, but a normal extraction omits them when empty, so MCP clients (e.g. Claude Code) rejected the result. Those fields are now optional in the schema, matching the serialized output. (#1372)
• The install.sh script no longer creates a self-referential xberg symlink that shadowed the installed binary, and it now selects the glibc (-gnu) build on standard Linux distributions instead of always downloading the musl build — which failed to run on glibc systems such as Ubuntu. musl systems (e.g. Alpine) still get the musl build. (#1371)
• CSV header inference no longer misclassifies all-text tables as headerless. A first row such as Name,City is now treated as the header (the dominant CSV convention) instead of rendering a broken blank header row with the real header pushed down into the data. A numeric-looking first row is still treated as data. (#1369)

• The extraction HTTP server now bounds in-flight request concurrency so a burst of large uploads can no longer exhaust memory and OOM-kill the process in memory-limited containers. The limit defaults to 2 × CPU count clamped to [4, 32]; override it with XBERG_MAX_CONCURRENT_REQUESTS (set 0 to disable). (#1368)
• PaddleOCR output now keeps consecutive visual text lines in the same Markdown paragraph instead of turning every detected line into a separate paragraph.
• PaddleOCR and Tesseract automatic image rotation now use the document-orientation model's RGB input and existing probability output correctly, and recover sparse edge-aligned text that the model's standard center crop omitted.

• cargo install xberg-cli now succeeds on a stock Windows toolchain. HEIC/HEIF decoding links native libheif, which has no default build path on Windows, so it is no longer part of the CLI's default features and the install no longer fails building libheif-sys. Enable HEIC with --features heic; the prebuilt release binaries, Docker all image, and Homebrew bottle continue to ship it. (#1361)
• The cargo binstall xberg-cli static musl builds now compile. The #1355 image-fallback OCR helpers were gated on the ocr feature but are reachable under the ocr-pipeline-only binstall profile, which failed to build both musl targets in the 1.0.9 release.
• brew install xberg-io/tap/xberg installs a working binary again instead of an empty bottle; the 1.0.9 bottle rebuild had been skipped when the CLI asset upload cascaded from the failed binstall build. (#1356)
• The hosted demo page (docs.xberg.io/demo.html) no longer 404s its toolbar and file-picker icons. (#1360)
• Dart native-library loading now propagates download, filesystem, and checksum failures instead of silently falling back to an unverified default library resolution path.
• PaddleOCR concurrent cold starts now run off async worker threads and share one engine initialization per model and accelerator, with distinct cache entries for different GPU device IDs.
• Benchmark text F1 now segments CJK around embedded Latin and numeric text while ignoring OCR line wrapping, preventing mixed-script output formatting from distorting quality comparisons.

• cargo binstall xberg-cli now installs a self-contained, fully static musl CLI binary with no ONNX/Tesseract/libheif runtime dependencies. The x86_64-unknown-linux-musl build additionally bundles the pure-Rust Candle VLM OCR backends (TrOCR and PaddleOCR-VL); aarch64-unknown-linux-musl ships extraction-only. ONNX/Tesseract/HEIC OCR remain available via Homebrew and the bundled per-target release tarballs.

• PaddleOCR now exposes the PaddleOcrEngine name and detailed word-level quadrilaterals; the former OcrLite name remains available as a deprecated compatibility alias.
• Dense XLSX extraction now scans worksheet bounds without cloning every cell before normal range parsing, while oversized sparse sheets materialize their cells only once.
• Layout-enabled image table recognition now shares its decoded RGB raster with the TATR worker, avoiding one full image allocation and pixel-buffer copy per qualifying image.
• Multi-stage PDF OCR now shares rendered page rasters across pipeline tasks instead of copying each pixel buffer, reducing peak memory by roughly one RGB raster per concurrent page.
• Batch DOCX extraction reuses one owned input buffer and avoids rebuilding discarded document structure, reducing memory copies and structure-processing overhead for large files.

• Canonical PaddleOCR benchmark presets no longer force optional whole-image auto-rotation, avoiding confident but incorrect 180-degree rotations that suppressed scene-text quality.
• PaddleOCR now preserves native resolution for 1024-pixel images by default, improving scene-text accuracy while retaining explicit detector-size overrides.
• Layout-enabled image OCR now reuses successful single-frame whole-image text when structured assembly is unavailable, avoiding repeated region OCR and redundant Tesseract table analysis.
• PaddleOCR-only CLI builds no longer compile PDF Markdown layout reuse code when layout detection is disabled.
• The prebuilt macOS CLI tarballs (aarch64-apple-darwin, x86_64-apple-darwin) now bundle the full libheif dynamic-library closure beside the xberg binary and rewrite its load commands to @loader_path, so the binary no longer fails with a libheif.1.dylib not-loaded error on machines that lack Homebrew's libheif at the baked-in path (#1357).
• PaddleOCR layout and table consumers now use projected CTC word boxes while preserving line-level semantic text and caller-requested element granularity, avoiding mixed-level duplicate table text.
• PaddleOCR detection now honors its configured DB threshold and matches upstream dilation, perspective-crop, and visual-line ordering behavior, improving small, skewed, and jittered text.
• Apple Keynote packages containing only slide archives now route to the Keynote extractor, and Numbers extraction reconstructs tables instead of emitting raw protobuf fragments.
• AsciiDoc, NXML/JATS, and WebVTT files now route through their registered text or JATS extractors instead of being reported as unsupported.
• Standalone excel and excel-wasm feature builds now include the XML parsing and table-capacity support required by XLSX extraction.
• Org-mode extraction now distinguishes separator-defined table headers from headerless tables, preserving every data row in rendered Markdown.
• EPUB extraction now resolves epub:switch branches per output renderer, preserving supported XHTML and MathML cases while retaining readable plain-text fallbacks.
• Typst extraction now emits marker-free headings and distinguishes explicit table headers from bare table rows, preserving correct Markdown structure.
• MSG extraction now reads the canonical binary PidTagHtml stream with the Internet codepage, preserving HTML-only message bodies alongside attachments.
• TATR table reconstruction now assigns each selected OCR word exactly once, using the nearest cell when predicted cells do not overlap, preventing both duplicated and silently dropped text.
• Layout-enabled image extraction now recognizes TATR table structure from cached OCR elements while preserving non-table line structure and requiring complete OCR token retention before accepting the reconstructed layout.
• Layout-enabled OCR now preserves detected image headings without losing or reordering fallback text, and regroups adjacent PDF OCR lines without collapsing distant paragraphs or separate layout regions.
• Rotated PDF OCR now avoids reusing display-coordinate Markdown layout rasters and reruns layout on inverse-/Rotate-normalized images, keeping OCR upright without desynchronizing detections.
• Benchmark text F1 treats OCR-inserted line breaks within CJK text as layout whitespace, preventing semantically identical Chinese, Japanese, and Korean output from scoring zero.
• Pipeline quality benchmarks allow forced OCR inference enough time to finish instead of recording slow but valid OCR documents as zero-quality timeout failures.
• Benchmark fixture validation now accepts descriptor filenames without an explicit parent path.
• Pipeline benchmarks now preserve exact ordered cohort fixture paths, use explicit PP-OCR model identities and fixture OCR languages, and score structural image ground truth.
• PaddleOCR now reports processed image dimensions and applied orientation corrections, keeping OCR geometry aligned with optional layout detection on rotated documents.
• PaddleOCR now selects the Japanese model for vertical Japanese and prefers Korean or Japanese recognition for mixed Latin-script requests those models can cover.
• PP-OCRv6 requests containing Korean now use PaddleOCR's script-specific Korean recognizer, recovering Hangul text that the unified recognition model omitted.
• PaddleOCR now preserves the right-to-left column order and contiguous text of traditional vertical Chinese and Japanese documents.
• Image OCR now preserves blank-line paragraph boundaries instead of flattening every recognized text block into one paragraph.
• Tesseract vertical CJK OCR now removes artificial spaces between adjacent script characters while preserving Latin-word and paragraph whitespace.
• Jupyter notebook paths retain application/x-ipynb+json routing when generic JSON content detection runs, and extracted notebook content no longer exposes diagnostic cell/output markers; cell identity, execution, tag, output-type, and MIME details remain available as structured metadata.

• Candle VLM OCR backends now ship in the published packages. The pure-Rust Candle OCR backends — TrOCR, PaddleOCR-VL, GLM-OCR, and DeepSeek-OCR — are compiled into the published packages by default (Python, Node, Go, Java, C#, Ruby, PHP, Elixir, Kotlin/JVM, Zig, and the CLI / Docker image) on Linux, macOS, and Windows. Select one with ocr.backend = "candle-glm-ocr" (or candle-trocr / candle-paddleocr-vl / candle-deepseek-ocr); model weights download from Hugging Face on first use. Previously these backends were excluded from the full feature and reachable only via a custom source build. Not available on WebAssembly, Android, iOS, Dart, or Swift.

• #1355 — force_ocr no longer emits a silently blank page when the PDF rasterizer cannot draw an image XObject. When a force_ocr page renders blank but carries image XObjects, OCR is retried directly on the embedded image bytes (decoded pixels, or the raw JPEG/JP2 stream) and a processing warning is recorded, so the page content is recovered instead of dropped without notice.
• Swift artifact-bundle cross-compile: the cross-compiled Swift binary bundle builds again — the HEIC path (which shells out to pkg-config and cannot cross-compile) is dropped from the Swift / Intel-macOS cross-build feature set (full-no-heic), restoring the x86_64-apple-darwin and Linux Swift builds. The native C FFI distribution keeps HEIC.
• XLSX extraction on Windows: the excel feature is enabled in the Windows feature set, so .xlsx files extract on Windows instead of returning UnsupportedFormat for a format the registry advertises as supported.
• Benchmark CI validates ground truth for every format family plus the exact 101-cell workflow matrix and harness contracts before expensive jobs, and the local benchmark task now delegates to the same run wrapper.
• Benchmark quality rankings and Pareto SF1 multiply successful-extraction medians by accountable coverage exactly once, so partial framework failures cannot retain a perfect rank while harness and setup failures remain excluded.
• Benchmark runs abort instead of dropping task errors, verify exact eligible-document cardinality before writing artifacts, reject contradictory failure states and unknown pipeline names, and report extension success rates with the same accountable-failure semantics as the aggregate.
• Benchmark CI invalidates its prebuilt harness cache for harness build scripts, workspace and toolchain configuration, compiler/codegen environment, and every transitive workspace crate, preventing stale binaries. Release tokens now default to the current repository installation.
• Present best-effort benchmark artifacts receive the same provenance, supported-format cardinality, failure-accounting, and aggregate integrity validation as required artifacts; only absence and framework-accountable extraction failures remain optional. Consolidated provenance, metadata, failure summaries, and rankings are cross-checked against validated groups and rows, with ranking optionality derived from the active cohort rather than a global framework union.
• Subprocess benchmark results record the framework's declared supported extensions, preserving the capability context needed to interpret historical multi-format aggregates.
• Benchmark quality guardrails fail on missing contracted documents or pipeline results instead of reporting a vacuous pass, and reject unknown pipelines, empty predicates, and invalid thresholds before execution.
• Unstructured benchmark cells advertise only their supported plaintext output, and pipeline benchmarks reject unknown sort metrics instead of silently falling back to SF1.
• Benchmark fixtures reject document and ground-truth paths that escape the repository or standalone fixture trust boundary, including symlink escapes, and derive repository boundaries from runtime fixture locations so cached binaries remain portable across CI runners. Artifact provenance hashing uses the same validated path resolution.
• Benchmark CI records declared per-framework format support, validates partial-run thresholds before execution, evaluates them independently per framework, and excludes harness/setup errors from framework success rates while retaining strict extraction coverage for every xberg pipeline.
• Benchmark comparisons include formats with text-only ground truth, report structural scores as unavailable instead of zero when Markdown ground truth is absent, and identify guardrails by file type so same-named fixtures cannot be matched across formats. Guardrails are rebased against the active corpus, removing retired PDF contracts and covering every actionable current result.
• Image layout extraction reuses safely positioned whole-image OCR elements, falls back when region-based OCR drops substantial text or quality, and preserves warnings without redundant OCR retries.
• EPUB extraction removes duplicated serialized MathML and embedded-media fallback content, and avoids emitting a cover image twice when the spine already references it.
• Email extraction preserves sender display names alongside addresses, and asynchronous attachment and nested-message extraction reuses the initial parse instead of parsing messages twice.
• FB2 and DocBook files with generic XML signatures retain their extension-specific MIME types, so they route through the semantic FictionBook and DocBook extractors instead of the generic XML fallback.
• Nested objects and arrays in JSON documents render as structured Markdown headings and lists instead of opaque compact-JSON strings, preserving readable nested keys and values.

• libwpd (Windows/MSVC): link the vcpkg-provided static zlib so librevenge's inflate* symbols resolve at the final link. Windows binding builds previously failed with undefined symbol: inflate because the MSVC path emitted no usable zlib link directive.
• #1344 follow-up: Automatic PDF layout inference retries once on CPU only for runtime inference failures, keeps explicitly selected non-Auto providers and recognized XBERG_ORT_EP values authoritative, ignores blank or unrecognized environment values, and propagates the effective or recovered CPU provider to downstream TATR and OCR table reconstruction.
• Side-by-side PDF TATR tables match source words that narrowly cross a detected outer edge to the outermost cell without changing the inference crop or center seam, preserving financial-table row prefixes that previously fell just outside the recognized cell bounds.

• Upgrade sibling dependencies: crawlberg 1.0.11 → 1.1.0, html-to-markdown-rs 3.9 → 3.10, liter-llm 1.11 → 1.12. liter-llm 1.12 makes tracing an always-on dependency and removed its tracing Cargo feature, so it is dropped from the dependency declaration (no behavior change — liter-llm spans are always emitted now).
• The otel feature now forwards to crawlberg and liter-llm (weak, crawlberg?/otel / liter-llm?/otel), so enabling xberg/otel compiles those siblings' direct OpenTelemetry integration (crawlberg's semconv/propagation, liter-llm's gen_ai.* metrics); their spans and metrics are exported by the host's provider (e.g. xberg-enterprise). html-to-markdown-rs and tree-sitter-language-pack are pure tracing emitters with no otel feature — their spans reach the collector through the consumer's tracing-opentelemetry layer, so nothing is forwarded to them.

• xberg-libwpd Windows build: the WordPerfect extractor now compiles and links on x86_64-pc-windows-msvc, unblocking the full-feature Windows binary of downstream consumers. Two first-ship gaps in the vendored C++ build are fixed: (1) zlib (needed by librevenge's RVNGZipStream) is now built from source via libz-sys on Windows too — as it already was on Linux/macOS — instead of relying on a vcpkg-installed zlib that CI did not reliably provide (fatal error C1083: Cannot open include file: 'zlib.h'); and (2) a narrowing std::make_shared<WP6SubDocument>(…, m_streamData.size()) call in WP6GeneralTextPacket.cpp — a 64-bit size() into a 32-bit const unsigned param — is patched to cast (unsigned) (matching every sibling subdocument site), which the newest MSVC toolchain (14.5x) otherwise rejects as a hard error. The vcpkg zlib probing in build.rs is removed.
• #1345: Sparse native two-column PDFs preserve column-block reading order instead of interleaving their four text lines row-by-row across the gutter.
• #1346: PaddleOCR emits a ProcessingWarning when requested languages are not covered by the single selected recognition model (previously their text was silently dropped), and OCR metadata now reports the recognition model actually used instead of joining every requested language.
• #1344: Layout inference no longer silently degrades to no-layout output when a hardware execution provider fails. macOS auto acceleration resolves RT-DETR to CPU up front (its current export cannot execute under CoreML), so the common path never attempts a failing provider. When an explicit accelerated provider does fail at inference (for example a CoreML ExecuteKernel error), both the markdown and OCR layout paths retry once on the always-available CPU provider and recover the layout, and either way surface a ProcessingWarning (recovered-on-CPU, or lost entirely if CPU also fails) instead of returning byte-identical no-layout output with empty processing_warnings.
• #1349: Successful TATR table reconstruction no longer writes source cell content and coordinates to stderr; the debug output is removed.
• #1350: The Markdown hierarchy no longer merges a distant header and footer into one block — paragraph continuation now rejects merges across a large vertical baseline gap and recomputes the merged block's bounding box.
• #1351: The published Node package ships the alef-generated index.d.ts (clean, consistent types) rather than the raw napi build output, which emitted references to undefined Js* types.
• #1353: The install script copies nested runtime library directories (for example lib/libheif) with cp -R, instead of failing with cp: -r not specified; omitting directory on Linux musl installs.

• Raw println!/eprintln!/print!/eprint!/dbg! are now denied in production code across the whole workspace (clippy print_stdout/print_stderr/dbg_macro); tracing is the sole diagnostic surface. The CLI's machine-readable result output to stdout opts back in per call site (#[expect(clippy::print_stdout)]), and the regenerated language bindings route their FFI-bridge diagnostics through tracing instead of eprintln!.
• Internal diagnostics that previously wrote to stderr via eprintln! (per-page OCR gate decisions, GLM-OCR debug tensor stats, the CLI --output-format deprecation notice) now emit through tracing at the appropriate level, so verbosity is controlled with RUST_LOG / --log-level instead of ad-hoc XBERG_DEBUG_OCR / XBERG_GLM_DEBUG environment variables.
• Repeated per-page and per-backend warnings from external dependencies (OCR engines, layout models) are now de-duplicated by (source, message), so an N-page document surfaces one warning per distinct problem rather than N copies. The paddle-ocr uncovered-language warning is also logged.

• MCP clients can run extract, extract_batch, and cache_warm as cancellable SEP-2663 tasks when they advertise task support; synchronous clients remain compatible.
• MCP cache_clear and cache_warm return typed structured results with cleared-file totals and model availability separated from confirmed cache-hit and download status.

• Dependency bumps: crawlberg 1.0.11, tree-sitter-language-pack 1.13.6, base64 0.23 (xberg-jni).

• #1338: Default OcrStrategy::Auto extraction OCRs scanned PDFs with no native text layer instead of returning empty content; explicit OCR disablement remains authoritative.
• #1341: Synthesized VLM fallback pipelines run for mixed native/OCR PDFs, preserve skipped and failed-stage diagnostics, and retain the last non-empty fallback when every stage scores below threshold.
• #1340: PDF images and generated captions render at bounding-box-aware reading-order positions, remain within the correct layout column, preserve source order, and stay consistent through chunking, translation, and redaction.
• #1343: Archive extraction skips macOS/tooling metadata entries (__MACOSX/, AppleDouble ._*, .DS_Store, Thumbs.db, desktop.ini, __pycache__/, .pyc/.pyo) instead of emitting them as text/plain children, and unsniffable extensionless members default to application/octet-stream; a single aggregated warning records what was filtered.
• Per-file OCR language overrides now also apply to explicit Tesseract pipeline stages, preserving override precedence.
• PDF plain-text extraction repairs detached subscripts, phone suffixes, and final glyphs while preserving RTL, rotated, vertical-writing, and mathematical span order.
• PDF Markdown atomically replaces adjacent native side-by-side table cohorts with validated layout table cohorts, avoiding mixed grids and dropped financial-table structure.
• OCR Markdown applies layout hints to line-local geometry while preserving soft-wrapped body paragraphs and merging multi-line headings, code, pictures, and wrapped list items by hint.
• Tesseract OCR Markdown aligns layout hints and table-cell matching with DPI-normalized and auto-rotated image coordinates, restoring semantic structure on scanned PDFs.
• OCR Markdown recovers missing ordered-list successors only when an existing numeric list item anchors a complete, bounded three-item sequence across pages.
• PDF Markdown preserves strong native headings when a lower-confidence layout Code hint lacks structured code evidence.
• OCR Markdown recovers a title from a guarded first-block logo/title pattern when the layout model emits no semantic heading region.
• PDF Markdown preserves native heading, list, code, and formula semantics while using layout geometry for reading order, grouping, and tables, tolerates minor crop jitter in side-by-side cohorts, merges sparse currency-affix columns without dropping markers, and folds wrapped financial-table lines into logical records; table-dominant pages also discard bbox-confirmed crop spill while retaining surrounding prose and annotations.
• PDF Markdown reconstructs paired wrapped financial tables as semantic three-column grids and repairs consistently merged numeric columns from native PDF table detection.
• #1342: PDF table reconstruction retains short numeric grids when a small number of inferred columns make the principal data row nearly complete instead of fully populated.
• PDF Markdown recognizes repeated large-font heading tiers across sparse multi-page documents while retaining the single-page sparse-document safeguard against display-text false positives.

• PDF benchmark fixtures can pin Tesseract OCR languages. The benchmark harness validates language codes, checks required packs before timed extraction, and preserves the effective OCR backend and cache settings when applying per-file batch overrides.

• Upgraded rmcp to 3.0.0 and migrated the MCP server to its 3.0 API (schema output, the new cache-scope/result-type/TTL list-result fields, and the GetPromptResponse/ReadResourceResponse handler enums). The exposed tools, prompts, and resources are unchanged.
• OCR now emits tracing logs when it materializes a Tesseract language pack at runtime: an info line naming the language, destination, and source before the download, one per candidate URL as it is tried, and one on success. Previously a runtime language-pack download was silent, making a first-use OCR stall on a missing pack hard to diagnose. English is unaffected on builds with the bundle-tessdata-eng feature (embedded, no download).
• Dependency bumps: liter-llm 1.11.4, toml 1.1.4.

• #1333: A sparse continuation row no longer dilutes the numeric ratio used to classify a grid, so numeric line-item tables with a trailing partial row are kept as tables instead of being flattened to prose.
• #1336: Tesseract no longer creates OCR cache directories when caching is disabled; the cache directory is created lazily, only when a result is written.
• #1337: Light-text-on-dark-background scans are auto-inverted before OCR via mean-luminance polarity detection, and the previously-dead invert_colors config is honored as an explicit override (Some forces, None auto-detects).
• #1338: NER and summarization processors are now compiled into the container and CLI builds — they were feature-gated out, so ner/summarization config was silently dropped. Under OcrStrategy::ScannedPages, a whole-document text failure now OCRs every page instead of discarding the signal.
• #1339: VLM OCR forwards XBERG_LLM_* env credentials to ocr.vlm_config when a custom base_url is set, normalizes openai.com model names, routes bare images through the OCR pipeline so vlm_fallback on_low_quality fires, and surfaces per-stage OCR failures as processing warnings.
• Linux builds without CUDA or TensorRT no longer fail under strict warning settings because of an unused ONNX Runtime execution-provider trait import.
• Per-file OCR language overrides (CLI and benchmark) now reach nested Tesseract configurations.
• PDF plain-text extraction repairs detached text spans so words are no longer split mid-token.
• PDF plain-text extraction retains table assets without rendering native table text twice.
• PDF extraction recovers and stitches label-heavy financial tables without merging independent aligned tables.
• PDF Markdown preserves explicit word boundaries and changelog heading hierarchy.
• OCR Markdown prefers validated semantic layout hints over broad text regions at comparable overlap.

1.0.2 is a packaging release. It completes the 1.0.1 rollout — the PHP/Packagist binding failed to build for 1.0.1 — and adds a first-party coding-agent plugin. No core extraction behavior changed.

• Coding-agent plugin. A first-party xberg plugin for Claude Code, Codex, Cursor, and OpenCode, with a Hermes variant, ships extraction skills (batch extraction, chunking, OCR, tables, keywords, format selection) that drive xberg through its MCP/CLI surface. Published as @xberg-io/opencode-xberg (npm) and xberg-hermes-plugin (PyPI).

• The PHP binding now builds against ort 2.0.0-rc.13. rc.13 moved the CoreML/CUDA/TensorRT execution-provider types behind matching Cargo features; a fresh dependency resolution (as on the PHP build) picked up rc.13 and failed to compile. Those EP features are now enabled unconditionally — a compile-time #[cfg] unlock only, with no SDK dependency or runtime change — so the PHP/Packagist package publishes again.

• Drops the Node @xberg-io/xberg-win32-arm64-msvc sub-package. It was declared as an optional platform dependency but never built — no xberg binding targets Windows on ARM64 — leaving an unresolvable optional dependency. The target is removed from the package manifest and loader for parity with the other bindings.
• Republishes every binding at 1.0.2 to close the 1.0.1 gaps (notably PHP/Packagist).

• #1321: Borderless, text-heavy tables are recovered on pages that also contain an ML-detected table. The geometric-table fallback now runs per region instead of per page, so a single ML Table hint no longer suppresses borderless-grid recovery across the rest of the page; words already inside an existing table hint are excluded so regions are not detected twice.
• #1326: RTF hex byte escapes now decode through the active font's \fcharsetN charset (mapped to a Windows codepage), falling back to \ansicpgNNNN and then Windows-1252. Documents that declare a Cyrillic or other non-ANSI font in the font table now decode as readable text instead of Windows-1252 mojibake, and font switches mid-document are tracked across nested groups.
• #1328: Page markers now appear verbatim in Markdown and Djot output. Flat documents no longer backslash-escape the marker (\<\!-- PAGE 1 --\>), and structured native documents no longer drop it entirely.
• #1323: RTF hex byte escapes now honor \ansicpgNNNN via the shared Windows-codepage table, so CP1251 Cyrillic and other non-1252 ANSI byte runs decode as readable text instead of Windows-1252 mojibake; adjacent escapes decode as one multi-byte run, surviving line wraps, and formatting spans stay aligned with the decoded text.

• LayoutStrategy enum on LayoutDetectionConfig (strategy field, default always). auto pre-screens each PDF page with cheap geometry signals and runs the layout model only on pages likely to benefit; existing configs keep the every-page behavior bit-for-bit. On the OCR path only inference is skipped, since OCR consumes the layout pass's rasters. Skipped pages are auditable via metadata.format.layout_gated_pages and layout_gate_reasons, and the CLI gains --layout-strategy (#1322).

• Republishes xberg-libwpd with the static zlib link fix so xberg-cli links against a working release. The 1.0.0 xberg-libwpd crate was published before the fix and left the librevenge inflateInit2_/inflate/inflateEnd symbols undefined at final link, breaking xberg-cli builds from crates.io. No source API changes.

xberg 1.0.0 is the first stable release of the document-intelligence engine previously developed as Kreuzberg. It is the direct successor to Kreuzberg v4.9 and carries the same Rust core and extraction-API lineage forward under the xberg name. The Kreuzberg v4 line continues as LTS at kreuzberg-dev/kreuzberg-lts. This entry summarizes everything that changed relative to Kreuzberg v4.9.

Beyond the rename, 1.0.0 is a large release: the PDF stack moved to a pure-Rust backend, the OCR story grew from a single engine to a family of classical and vision-language models, and whole new capabilities landed — audio/video transcription, named-entity recognition, structured LLM extraction, sparse/late-interaction retrieval, and four new language bindings.

For a step-by-step upgrade, see the migration guide.

• Packages are renamed kreuzberg → xberg across every ecosystem (crates.io, PyPI, npm, Maven, NuGet, Composer, RubyGems, Hex, Go).
• The Rust error type KreuzbergError is now XbergError.
• Environment variables are re-prefixed KREUZBERG_* → XBERG_*, and config files are discovered as xberg.{toml,yaml,yml,json}.
• Breaking API changes: extracted URIs are returned as ExtractedUri (formerly Uri); document metadata drops the untyped additional/serde-flatten bag in favour of typed fields plus a custom residual map.
• The R binding, the EasyOCR backend, and the bundled pdfium fork are removed (see Removed). Existing Kreuzberg v4 installs keep working under their original names.

The full identifier mapping is in the migration guide.

• A family of OCR backends. Alongside Tesseract, 1.0.0 adds a native PaddleOCR backend (PP-OCRv6, with medium/small/tiny tiers) and a pure-Rust Candle OCR/VLM stack — TrOCR, GLM-OCR, GOT-OCR, DeepSeek-OCR, and PaddleOCR-VL — that runs without ONNX Runtime or native Tesseract. Model weights are self-hosted on the xberg-io Hugging Face org.
• A second, ONNX-Runtime-free inference path (tract). CNN classifiers, layout detection (RT-DETR), and auto-rotation run through a pure-Rust tract backend on targets without ONNX Runtime — this is what makes in-browser (WASM) and mobile inference possible.
• Structured (LLM) extraction. extract_structured and split_and_extract drive a vision-LLM client with rasterization, chunking, citations, caching, and configurable CallMode / MergeMode / VLM-fallback policies.
• Audio and video transcription. A Whisper ONNX encoder/decoder engine extracts text from .mp3, .wav, .m4a, .mp4, and .webm.
• Named-entity recognition. GLiNER2-based entity extraction, including an in-browser WASM NerModel that detects entities locally with no server round-trip.
• Retrieval building blocks. Sparse embeddings (SPLADE), ColBERT late-interaction retrieval, and a cross-encoder reranking / semantic-search stage alongside dense embeddings, with self-hosted model presets pinned by sha256 manifests.
• Text intelligence. Redaction with reversible rehydration and per-entity erasure, summarization, translation, VLM image captioning, QR-code detection, document diffing (revisions on ExtractionResult), and page/chunk classification.
• URL and web ingestion. map_url discovers URLs from sitemaps and a shared crawl engine batches multi-URL extraction, over a URI-based ExtractInput / ExtractionOutput envelope.
• New document formats (98 total). WordPerfect .wpd/.wp/.wp5 (via a vendored xberg-libwpd), HEIC/HEIF/AVIF images (via a vendored libheif), OpenDocument Presentation .odp, Quarto/R Markdown, configurable Jupyter cell rendering, and the audio/video formats above.
• Four new language bindings. Dart/Flutter, Swift, Kotlin/Android, and Zig — for 15 language bindings over one engine, with Android/iOS cross-compilation.
• First-party integrations, consolidated into the monorepo. LangChain.js, LlamaIndex, an n8n community node, CrewAI, and a Spring AI document reader.
• Richer chunking and API surface. Caller-supplied tokenizers, TableChunkingMode::RepeatHeader, RAG chunking with heading-path breadcrumbs, multi-label chunk classification, per-page spans with bounding boxes, a list_supported_formats() call in every binding, cheap pdf_page_count, and a DELETE /jobs/{job_id} cancellation endpoint on the API server.
• Wider code intelligence. tree-sitter coverage grows from 248 to 306 programming languages.

• PDF backend replaced. pdfium is gone; pdf_oxide, a pure-Rust engine, is now the sole PDF backend — no native pdfium dependency.
• Layout-aware PDF pipeline. Reading order is reconstructed with ONNX layout detection (PP-DocLayoutV3 / RT-DETR) and Docling-style predecessor-graph reordering; scanned PDFs are detected and OCR'd selectively per page; AcroForm/XFA form fields and outline-based headings are extracted.
• Public API stabilized and frozen for 1.0, with a Rust-only Engine and extension seams.
• Renamed from Kreuzberg to xberg across packages, namespaces, and the KreuzbergError → XbergError type (see Migration).
• Environment variables use the XBERG_ prefix; new layout, OCR model-tier, CoreML, and ORT execution-provider variables are available.
• Config discovery now also accepts the .yml extension (xberg.{toml,yaml,yml,json}) with an XDG config-directory fallback.
• Models and cache live under the xberg cache segment and the xberg-io Hugging Face org; the project domain is xberg.io.
• Python support widens to 3.10–3.14; the SurrealDB connector moves to v3 (dropping mem://); the default extraction_timeout_secs is 60s.
• License. Relative to the Kreuzberg 4.8/4.9 line (Elastic License 2.0), xberg 1.0.0 is MIT.

More than 150 bugs were resolved during the 1.0 cycle. Highlights by area:

• PDF text fidelity: text inside Marked-Content (MCID) blocks is no longer dropped from markdown/HTML output (#917); ligature glyphs no longer map to control characters (#1135); glyph-spaced text no longer extracts one character per line (#962); spurious intra-word spaces in native extraction are fixed (#1291, #1222); JPEG 2000 images no longer render blank and silently break OCR (#1158); XML entity references (&/</>) are preserved (#1242).

• Tables: bordered / graphical-line tables are detected reliably instead of silently skipped (#964, #1097, #1213); rotated full-page tables no longer extract as word salad (#1220, #1221); duplicate table emission and double-counting are fixed (#1288); physically fragmented per-row tables are merged back with their header row (#1290, #1100); borderless and text-heavy grids keep their row associations (#1316, #1319).

• Reading order & structure: two-column reading order no longer scrambles headings (#1170); stale page boundaries after reordering no longer panic on multibyte text or drop documents (#1270, #1272); numbered and cover-page headings are classified correctly (#961, #966, #1096, #1098); filled form field values are placed correctly (#1120).

• OCR: an explicit PaddleOCR backend no longer silently falls back to Tesseract (#801, #1071, #1088, #1102); scanned-page OCR text and page provenance surface consistently across content, pages, and chunks (#1095, #1110, #1281); spurious auto-OCR on born-digital PDFs is suppressed (#1176); a SIGBUS crash and a NaN-sort panic in the OCR pipeline are fixed (#1057, #1179); the Candle VLM OCR backends are stabilized (#1174, #1175, #1208–#1214); model downloads handle TLS-MITM CAs, IPv6 blackholes, and connect timeouts (#1146, #1249).

• Chunking & provenance: chunk firstPage/lastPage and byte ranges are correct across output formats and long PDFs (#1013, #1074, #1105, #1294); markdown chunks retain markdown (#1073, #1094); split-table chunks keep their header and context (#1100).

• Bindings & packaging: fixed Go embed symbols, Java UnsatisfiedLinkError, missing C# config types, wrong Node/PHP embedding shapes, Android .so loading, macOS wheel floors, and musl/ONNX Runtime runtime deps (#871, #965, #991, #998, #1008, #1055, #1131, #1257, #1304, #1307); plus Homebrew 404s, Docker stop-signal handling, and multi-arch -core images (#1081, #1147, #1247, #1315).

• Formats: EML HTML <table> bodies, DOCX hyperlink/bold overlap and markdown conversion, archived markdown/CSV escaping, and Korean-charset EML detection are fixed (#942, #1086, #1212, #1237, #1278).

• Config & robustness: extraction_timeout_secs is honoured on every path (#830, #911, #1273); cancel_token, custom LLM base URLs, and page-classification config all validate correctly (#937,

  #944, #1076).

• R binding — the Kreuzberg v4 LTS line is the last to ship it.
• EasyOCR backend — the Python/torch-only backend did not survive the Rust rewrite; use Tesseract, PaddleOCR, a Candle backend, or a VLM backend instead.
• Hunyuan-OCR Candle backend — ported during development, then dropped before 1.0.0.
• Bundled pdfium-render fork and its KREUZBERG_PDFIUM_BUNDLED_PATH variable, and the standalone @kreuzberg/core npm package.

• OCR memory discipline: concurrent Tesseract sessions are capped, Leptonica/Pix/page buffers are released early, decoded RGB buffers are reused, and images are resized without copies.
• Layout inference: model sessions are pooled, batch inference threads are balanced, and an unnecessary PNG raster round-trip is bypassed.
• Engine and batch: bounded batch scheduling, no per-item config clone, single PDF parse per structured rasterization, streamed batch JSON output, and base64 hosted embeddings.
• PDF and text: streamed RGB conversion, reused OCR render document, skipped redundant compatibility parses, and a regex→scanner rewrite that removes backtracking from text/quality cleanup.

• An untrusted RTF size field in the email extractor could allocate up to 4 GB — now bounded (#1058).

• A redaction path could leak PII across roughly a dozen output fields — fixed (#1223).

• PDF embedded streams are guarded by a decompression-ratio limit and per-embedded-file size caps, and a SecurityBudget is wired through the PDF and email extractors.

• Excel DDE / external-call formulas raise warnings during extraction.

• FFI image and attachment buffers now carry explicit lengths so callees never read past the buffer (#1056, #1059), and panics on malformed input are replaced with recoverable errors (#907, #1057,

  #1198).

• pdf_oxide replaces the pdfium native dependency; libheif (LGPL, documented) and xberg-libwpd are vendored; retrieval and OCR model presets are self-hosted on xberg-io with sha256 manifests.
• Distribution hardening across all 15 targets: ONNX Runtime bundling, glibc/musl floors (musl via Alpine images), NuGet runtime-package size splits, Homebrew bottles, Go module tags, Swift C++ linkage, and Dart/Swift/Kotlin-Android/Zig release matrices. Published to crates.io, PyPI, npm, Maven Central, NuGet, RubyGems, Packagist, Hex, pub.dev, Go, Swift Package Manager, Homebrew, Docker (ghcr.io/xberg-io/xberg), and a Helm chart.