Graphenium v0.2.0: Provenance-Aware Structural Memory

Graphenium v0.2.0

Provenance-aware structural memory for AI coding agents.

This release upgrades Graphenium from a basic code indexer to an
agent-grade codebase intelligence engine. Every edge now carries
provenance metadata so agents know exactly how much to trust the graph.

New in v0.2.0

Graph provenance and trust model:

• extractor and resolution_status fields on every edge and node
• Tree-sitter edges marked [tree-sitter:resolved], LLM edges by confidence
• Cross-file import resolver marks edges resolved or unresolved
• Schema versioning in graph.json metadata (schema 0.2.0)
• graph_stats reports provenance and confidence breakdowns

Architectural analysis:

• Directed graph projections from undirected petgraph
• PageRank for architectural importance scoring
• Reverse reachability for blast radius analysis
• Community boundary crossing scores
• Rooted dominators for mandatory gateway identification
• chokepoint_report combining multiple ranking signals

Symbol diff and impact analysis:

• gm diff command for graph snapshot comparison
• Symbol inventory diff: added, removed, community-changed
• Downstream impact via reverse reachability
• Risk-sorted review order recommendations

Hybrid retrieval:

• gm query --mode flag: lexical, structural, hybrid
• TF-based text embeddings with cosine similarity
• Node2Vec structural embeddings from random walks

Runtime telemetry (experimental):

• OTEL-compatible trace import
• P50/P95/P99 latency percentile estimation
• Hot path and runtime-weighted traversal
• Import module marks telemetry as experimental

Diagnostics and documentation:

• gm doctor reports schema version, extraction modes, languages
• Competitive positioning in README with comparison table
• AI Skill updated with provenance guidance
• All 315 tests pass, 22 files changed, 5 new modules