v2.6.0

Minor Changes

• #310 d368eb9 Thanks @shtse8! - Add the v3 agent document map and visual evidence path for PDF intelligence workflows. include_document_map now returns linked pages, structured elements, citation chunks, layout diagnostics, safety findings, routing signals, page geometry, and summary counts while keeping image bytes out of JSON. This release batch also adds include_text_layer, a deterministic line and word text layer with page-level character ranges, best-effort bounding boxes, and provenance. It adds include_document_ast, a semantic tree with page, section, paragraph, list item, table, and image nodes linked back to element and chunk evidence. It adds include_trust_report, a consolidated local risk report for content safety, layout uncertainty, sparse/scanned pages, table quality warnings, and external-link routing. It adds include_accessibility_report, a deterministic accessibility report for tagged-PDF coverage, structure tree availability, heading roles, image alt-text verifiability, form labels, link labels, mark info, and accessibility permissions without claiming PDF/UA certification. It adds search_pdf for bounded literal search with snippets, match offsets, text-item bounding boxes, and provenance before heavier read/render/crop workflows. It adds render_page, which renders selected PDF pages as bounded PNG MCP image parts with JSON provenance, evidence IDs, pixel budgets, and page-level metadata for visual inspection and OCR routing. It adds extract_regions for PDF-coordinate bbox crops as focused PNG MCP image parts with crop metadata and provenance. It adds analyze_regions, an optional env-configured local visual-region provider pipeline that passes focused crops to a local command and normalizes table, chart, formula, figure, image-description, confidence, warning, crop evidence, and provenance fields without bundling a vision model. It adds ocr_pages, an optional env-configured local OCR provider pipeline that renders selected pages, passes temporary PNGs to a local command or MCP_PDF_OCR_PRESET=tesseract, and returns normalized OCR text, confidence, word boxes, language, render evidence IDs, and provenance without bundling an OCR model. inspect_pdf now reports safe optional-provider readiness for ocr_pages and analyze_regions without exposing local command paths or arguments. It also adds table quality diagnostics with inferred cell span/header hints, sparse-grid warnings, merged-cell candidate signals, and repeated-header continuation candidates. It also includes optional include_layout_diagnostics output with page layout profiles, reading-order confidence, column signals, and warnings for agent routing. The release batch updates bun run benchmark to run a reproducible local PDF intelligence benchmark with JSON output.