Release v1.6.1: Context Optimization and Structural Edits

This release focuses on improving agent reasoning by refining how document context is presented, drastically reducing payload sizes, and stabilizing structural formatting edits.

Context & Agent UX Improvements

• Structural Appendix Decoupling: The read_docx tool no longer forces the structural appendix (defined terms, cross-references, diagnostics) onto every page. Body pages now include a lightweight pointer, and agents can fetch the appendix on demand using mode="appendix". This prevents massive context bloat and token waste on large legal documents.
• Smart Ambiguity Resolution: When an agent attempts an edit with an ambiguous target, the validation error now explicitly displays the surrounding text for up to 5 matching locations. This gives the LLM the exact context it needs to instantly self-correct and retry.
• Annotated Revision Bubbles: Tracked change IDs within the CriticMarkup bubble are now explicitly tagged by their action (e.g., [Chg:12 insert], [Chg:12 delete]). This eliminates agent confusion when accepting or rejecting specific halves of a text substitution.
• Revision ID Coalescing: Edits that span across internal XML formatting boundaries (such as a partially bolded phrase) now share a single logical Chg ID in the UI, rather than fragmenting into multiple confusing IDs.

Core Redlining Engine

• Clean Heading Swaps: Changing a heading's level via text replacement (e.g., # Heading to ## Heading) now performs a clean, atomic paragraph replacement. It accurately transfers the style without leaving behind empty "orphan" paragraphs in the document.
• Surgical Initialization: The engine no longer performs global document normalization on startup. This strictly preserves untouched paragraphs, preventing the silent destruction of unrelated metadata like spellcheck markers or untargeted comments.
• Live Word Stabilization: Major reliability improvements for the Live Word COM mode. Multi-paragraph insertions utilize a new anchoring algorithm to prevent formatting drift, and revision targeting now uses semantic proximity to survive Word's background index shifting.
• Asynchronous Sanitization: The sanitize_docx tool now offloads processing to a background thread. This prevents heavy document scrubbing from blocking the MCP event loop and dropping client heartbeats.
• Performance: Implemented a single-pass extraction engine for domain metadata and an O(1) XML style cache, significantly reducing the time it takes to parse and project large documents.