Baseline: 29 open issues on 2026-09-16. Close only after implementation, relevant verification, and delivery to the repository. Feature requests retain their original scope. GitHub closure is a separate gate from a local passing test.

• Implement each request or demonstrate existing behavior with a regression/example.
• Run offline tests and relevant integration checks; record unavailable external checks honestly.
• Update affected reference and recipe documentation.
• Commit and deliver tested changes, checking CI before merge.
• Close resolved issues with links to delivered changes; verify open-issue count at the end.

• 67 offline tests passed on Python 3.12.9 with fresh dependency resolution, including Pydantic 2.13.5, Instructor 1.17.0, LiteLLM 1.101.0 and pypdfium2 5.13.0.
• Found and fixed undeclared NumPy import and removed PDFium document-render API during clean-install testing.
• Wheel and sdist built successfully; strict MkDocs build passed after removing placeholder navigation and linking existing EasyOCR docs.
• Cloud adapters were verified offline with SDK call interception; no paid provider calls were made.
• #351 still needs the pagination/input-budget follow-up before closure. Other untouched feature requests remain pending.

Delivery note: GitHub rejected workflow writes because the OAuth token lacks the workflow scope. CI edits are preserved locally and on local branch codex/2026-ci-preparation; source/docs fixes are published separately. Workflow authorization is requested; do not claim the new CI matrix is deployed yet.

• 104 offline tests passed on Python 3.12; strict MkDocs build passed.
• The same 104 tests also passed on Python 3.9 and 3.13; all six GitHub checks passed before merge.
• Pagination no longer drops failed pages, associates out-of-order results with the wrong source, fabricates missing required strings, or silently chooses unresolved conflicts. Partial schemas retain field metadata and aliases, with final contract validation. These improvements relate to #357; its original private Ollama/image reproduction has not been verified, so it remains open.
• GitHub confirms seven issues closed after merge of #364; 22 of the initial issues remain open.

• Markdown preservation/image embedding/splitting implemented with offline tests and documentation.
• Plain Markdown failures now raise instead of silently generating error comments/fallbacks; structured conversion does not mix error strings with PageContent results.
• GitHub confirms #365 merged and five more issues closed (12 total; 17 remaining before this batch).

• Docling dependency failures retain their cause, report SDK versions and provide dependency repair steps for #309.
• Fixed default OCR/table options, converter reuse, page numbering and Markdown metadata; six offline regressions added.
• Fresh Docling 2.127.0 / docling-core 2.96.1 installation passed a real, model-free HTML conversion and uv pip check (135 compatible packages).
• GitHub confirms #366 merged and three more issues closed (15 total; 14 remaining).
• Numeric classification IDs preserve identity through splitting and extraction, including duplicate display names. Failed calls and invalid page coverage now raise.
• 148 offline tests passed on Python 3.9, 3.12 and 3.13; strict documentation build passed.

• Camelot and Tabula PDF table loaders implemented with optional imports, table/page preservation, stream cleanup, image support, tests and documentation.
• GitHub confirms #367 merged and two more issues closed (17 total; 12 remaining).
• Real Camelot and Tabula SDK tests both passed on a generated ruled table plus blank page. Tabula used a portable Java runtime in the test environment.
• Ten new offline tests passed on Python 3.9, 3.12 and 3.13; strict docs build passed.

• Adobe PDF loader implemented with optional SDK, environment/injected credentials, normalized regions and CSV tables.
• Live Adobe service calls were not made; tests intercept the service while using real SDK job objects.
• GitHub confirms #368 merged (19 issues closed; 10 remaining).
• 166 offline tests passed on Python 3.12; eight Adobe regressions also passed on Python 3.9 and 3.13. Two real PDF Services SDK 4.3.0 compatibility checks passed using an offline service; strict docs build passed.

• SQLite retrieval selects relevant pages before model calls and preserves source metadata. The full document is still parsed; ranking is lexical.
• GitHub confirms #369 merged (20 issues closed; nine remaining).
• 179 offline tests passed on Python 3.12; 22 retrieval/page-selection checks also passed on Python 3.9 and 3.13. Strict docs build passed.

• ComplexityRouter selects user-configured LLMs using a customizable workload score and capability guards. No claim of automatic quality/cost prediction.
• GitHub confirms #370 merged (21 issues closed; eight remaining).
• 193 offline tests passed on Python 3.12; all 14 routing tests also passed on Python 3.9 and 3.13. Strict docs build passed.

• Partial/parallel extraction supports field annotations, model/vision policies, groups and final contract validation. Source loading occurs once per source.
• GitHub confirms #371 merged (22 issues closed; seven remaining).
• 203 offline tests passed on Python 3.12; all ten field-extraction tests also passed on Python 3.9 and 3.13. Strict docs build passed.

• Local masking supports explicit entities, regexes and optional email matching; no LLM receives originals for masking. The text wrapper rejects vision and removes page image fields.
• Reviewed legacy PR #72; its LLM-based masking approach is not reused. The new API uses local masking and has explicit coverage limitations.
• GitHub confirms #372 merged (24 issues closed; five remaining).
• 215 offline tests passed on Python 3.12; twelve masking tests passed on Python 3.9 and 3.13. Strict docs build passed.

• DocumentLoaderEvents supports handwritten/chart/image detection via a configured vision LLM, custom detectors and local rules. Detection finishes before callbacks run.
• GitHub confirms #373 merged (25 issues closed; four remaining).
• 224 offline tests passed on Python 3.12; nine event tests passed on Python 3.9 and 3.13. Strict docs build passed.

• Optional MCP 2.x service accepts JSON Schema contracts and confines file access to a configured root; PDF selection precedes rendering. Docker and Compose provide a local deployment.
• GitHub confirms #374 merged (26 issues closed; three remaining).
• Ten optional service tests passed, including the real Instructor adapter with an offline completion transport. The 224 existing core tests and four new MIME regressions passed; the MIME regressions also passed on Python 3.9 and 3.13. Strict documentation build passed.
• No paid provider request was made for the container tests.

• Native structured output is an explicit LLM option, with provider-schema forwarding and guards against dynamic/raw-router bypasses. Ollama recipe distinguishes input context from output limits.
• GitHub confirms #375 merged (28 issues closed; one remaining). Final Docker image passed HTTP health, tool discovery and structured document-listing checks.
• 231 core tests passed; three new structured-output regressions also passed on Python 3.9 and 3.13. Strict docs build passed. No live Ollama accuracy claim: original private documents/server were unavailable.

• GitHub reports zero open issues after #376: all 29 baseline issues are closed. Five pre-existing contributor PRs remain separate review work.
• README, documentation homepage, installation/quickstart, loader selection guide and 2026 compatibility notes refreshed. Removed redirect-only landing page and unsupported blanket PII/accuracy claims.
• CI workflow changes are prepared locally but cannot be pushed with the current OAuth token: GitHub requires the workflow scope. Existing GitHub checks passed before each source merge; offline tests and strict docs were run locally.