research(context): EDU-based structured context compression (LingoEDU)

[bug-ops]

Summary

Reformulates context compression as structure-then-select: decomposes text into Elementary Discourse Unit (EDU) relation trees, then selects query-relevant subtrees. Eliminates hallucination by anchoring EDUs strictly to source text indices.

Source: arXiv 2512.14244 — From Context to EDUs: Faithful and Structured Context Compression via Elementary Discourse Unit Decomposition
Published 2025-12-16, revised 2026-01-05.

Key Results

• State-of-the-art structural prediction accuracy on StructBench (248 manually annotated documents)
• Outperforms frontier LLMs on structured compression while reducing cost
• Zero hallucination: EDU nodes anchored to source byte offsets

Applicability to Zeph

Current Zeph compaction: summarize_tool_outputs() → token-based chunking → LLM summarization.

Problem: Chunking splits at arbitrary token boundaries, destroying discourse structure. Summaries lose causal/temporal relations between events.

Enhancement: Apply EDU decomposition as a preprocessing step in SemanticMemory::compress() before the compaction LLM call:

1. Parse tool output / conversation chunk into EDU tree (LingoEDU)
2. Score each EDU subtree by relevance to current task intent
3. Select top-K EDU subtrees (respect token budget) → pass to compaction LLM
4. Compaction LLM sees structured, non-redundant content → better summaries

Synergy: Complements #1851 (SWE-Pruner goal-guided pruning) and #1607 (structured anchored summarization). EDU tree provides the structure that #1607 assumes.

Implementation Sketch

• EduDecomposer trait in zeph-memory::compression
• Initial implementation: regex-based clause splitting (lightweight, no dependency)
• Advanced: integrate LingoEDU parser (Python subprocess or Rust port)
• Config: [memory.compression] edu_decomposition = false (opt-in, experimental)