Holographic codebase pack generator.
Produces a cache-stable, model-readable representation of a codebase that pre-loads enough information into an LLM's context to answer most codebase questions without tool calls.
Format
Each symbol gets one entry:
## <symbol_name>
L2: <signature> [<file>:<line>]
L3: <surprise facts only>
L2 (signature) is always present — it anchors the model's prior. L3 (surprise) is present only when the function deviates from what a trained model would predict.
Cross-references
Uses the SQLite index (occurrence + block + scope_binding tables) for precise, disambiguated cross-references — not regex word matching.
Surprise detection
Uses skeleton-frequency analysis: lines whose aggressive_skeleton is rare across the codebase are "surprising"; lines whose skeleton is common are "expected" (the model's prior handles them).