## Problem The code indexer generates deterministic node IDs by hashing component strings (path, name, etc.) using FxHasher and casting to `i64`: ```rust fn compute_id(components: &[&str]) -> i64 { let mut hasher = FxHasher::default(); components.hash(&mut hasher); hasher.finish() as i64 } ``` FxHasher produces 64-bit hashes. With the birthday paradox, collision probability reaches 50% at around 4 billion nodes. For large monorepos with millions of definitions across many branches, this becomes a real concern. The current implementation also affects edges, which reference these node IDs as `source_id` and `target_id`. ## Proposed Solution Investigate switching to Int128 for node and edge IDs. This would push the 50% collision threshold to \~10^19 nodes. Key areas that need changes: 1. **ID generation** - Update `compute_id()` to return `i128`, possibly using a 128-bit hash like xxHash3-128 or SipHash-128 2. **Schema types** - Add `Int128` support to Arrow type mappings and ClickHouse DDL generation 3. **Query engine** - Update `InputNode::node_ids`, `InputIdRange`, and ID filter functions to use `i128` 4. **JSON serialization** - JSON doesn't have native Int128; need to serialize as strings to avoid precision loss 5. **Result formatting** - Add `ColumnValue::Int128` variant and handle in response formatting The JSON serialization is the trickiest part since `serde_json` will lose precision for values exceeding 2^53. String representation is the safest approach but requires client-side parsing.