rdar 0.6.4

//! File-granularity reference graph and PageRank ranking.
//!
//! Nodes are source files; an edge A→B (weighted) exists when A references a
//! name that B defines publicly. Ambiguous names split their weight across
//! all definers (Aider's model). Symbol importance is the repo-wide count of
//! references to the symbol's name - deterministic and cheap.

use std::collections::BTreeMap;

use crate::cache::ScanCache;
use crate::extract::{RefKind, Vis};

/// Per-file rank + per-symbol-name incoming reference counts.
#[derive(Debug, Default)]
pub struct Ranking {
    /// rel path → PageRank score (sums to ~1.0 over source files).
    pub file_rank: BTreeMap<String, f64>,
    /// public symbol name → total cross-file references to it.
    pub name_refs: BTreeMap<String, u64>,
    /// public symbol name → referencing files with per-file counts
    /// (speed pass A8: replaces an O(files × refs) rescan per Jump row;
    /// also lets Jump rows pick the HEAVIEST referencers, not the
    /// alphabetically first).
    pub name_users: BTreeMap<String, BTreeMap<String, u64>>,
}

const DAMPING: f64 = 0.85;
const ITERATIONS: usize = 30;

/// Build the graph and rank it. Pure function of the scan cache.
pub fn rank(cache: &ScanCache) -> Ranking {
    // Stable integer IDs make the iterative path contiguous while preserving
    // the cache's path ordering and therefore deterministic floating-point
    // accumulation order.
    let nodes: Vec<&str> = cache
        .files
        .iter()
        .filter_map(|(rel, entry)| entry.lang.map(|_| rel.as_str()))
        .collect();
    if nodes.is_empty() {
        return Ranking::default();
    }
    // name -> integer IDs of files that publicly define it.
    let mut definers: BTreeMap<&str, Vec<usize>> = BTreeMap::new();
    for (index, rel) in nodes.iter().enumerate() {
        let entry = &cache.files[*rel];
        let Some(lang) = entry.lang else {
            continue;
        };
        let Some(x) = cache.parses.get(&(lang, entry.hash)) else {
            continue;
        };
        for def in &x.defs {
            if def.vis == Vis::Pub {
                definers.entry(&def.name).or_default().push(index);
            }
        }
    }

    // Edges: referencing file -> defining files (weight split on ambiguity).
    // Also count repo-wide references per public name (cross-file only).
    let mut edge_weights: Vec<BTreeMap<usize, f64>> = vec![BTreeMap::new(); nodes.len()];
    let mut name_refs: BTreeMap<String, u64> = BTreeMap::new();
    let mut name_users: BTreeMap<String, BTreeMap<String, u64>> = BTreeMap::new();
    for (source, rel) in nodes.iter().enumerate() {
        let entry = &cache.files[*rel];
        let Some(lang) = entry.lang else {
            continue;
        };
        let Some(x) = cache.parses.get(&(lang, entry.hash)) else {
            continue;
        };
        for r in &x.refs {
            let Some(defs) = definers.get(r.name.as_str()) else {
                continue;
            };
            let targets: Vec<usize> = defs
                .iter()
                .copied()
                .filter(|target| *target != source)
                .collect();
            if targets.is_empty() {
                continue;
            }
            // Calls are 3x the evidence of imports (accuracy pass: import
            // counting inflated boilerplate symbols above real hubs).
            let weight = match r.kind {
                RefKind::Call => 3,
                RefKind::Import => 1,
            };
            *name_refs.entry(r.name.clone()).or_default() += weight;
            *name_users
                .entry(r.name.clone())
                .or_default()
                .entry((*rel).to_string())
                .or_default() += weight;
            let w = 1.0 / targets.len() as f64;
            let out = &mut edge_weights[source];
            for t in targets {
                *out.entry(t).or_default() += w;
            }
        }
    }

    // Normalize once. The previous implementation repeated these sums and
    // divisions during every PageRank iteration.
    let edges: Vec<Vec<(usize, f64)>> = edge_weights
        .into_iter()
        .map(|out| {
            let total: f64 = out.values().sum();
            out.into_iter()
                .map(|(target, weight)| (target, weight / total))
                .collect()
        })
        .collect();

    // Plain power iteration - instant at file granularity (doctrine §2.4).
    let n = nodes.len() as f64;
    let mut rank = vec![1.0 / n; nodes.len()];
    let mut next = vec![0.0; nodes.len()];
    for _ in 0..ITERATIONS {
        next.fill((1.0 - DAMPING) / n);
        let mut dangling = 0.0f64;
        for (source, out) in edges.iter().enumerate() {
            if out.is_empty() {
                dangling += rank[source];
                continue;
            }
            for &(target, weight) in out {
                next[target] += DAMPING * rank[source] * weight;
            }
        }
        let spread = DAMPING * dangling / n;
        for v in &mut next {
            *v += spread;
        }
        std::mem::swap(&mut rank, &mut next);
    }

    Ranking {
        file_rank: nodes
            .into_iter()
            .zip(rank)
            .map(|(path, score)| (path.to_string(), score))
            .collect(),
        name_refs,
        name_users,
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::cache::FileEntry;
    use crate::extract::{Extraction, RefName, SymKind, Symbol};
    use crate::lang::Lang;

    fn file(cache: &mut ScanCache, rel: &str, defs: &[(&str, Vis)], refs: &[&str]) {
        let hash = *blake3::hash(rel.as_bytes()).as_bytes();
        cache.files.insert(
            rel.into(),
            FileEntry {
                mtime: (1, 0),
                size: 1,
                ino: 0,
                hash,
                lang: Some(Lang::Python),
            },
        );
        cache.parses.insert(
            (Lang::Python, hash),
            Extraction {
                defs: defs
                    .iter()
                    .enumerate()
                    .map(|(i, (name, vis))| Symbol {
                        line: i as u32 + 1,
                        end_line: i as u32 + 2,
                        name: name.to_string(),
                        kind: SymKind::Fn,
                        vis: *vis,
                        sig: format!("def {name}()"),
                        terms: Vec::new(),
                    })
                    .collect(),
                refs: refs
                    .iter()
                    .enumerate()
                    .map(|(i, name)| RefName {
                        line: i as u32 + 10,
                        name: name.to_string(),
                        kind: RefKind::Call,
                    })
                    .collect(),
            },
        );
    }

    #[test]
    fn heavily_referenced_file_ranks_highest() {
        let mut cache = ScanCache::default();
        file(&mut cache, "core.py", &[("gean_core", Vis::Pub)], &[]);
        file(&mut cache, "a.py", &[("a_fn", Vis::Pub)], &["gean_core"]);
        file(&mut cache, "b.py", &[("b_fn", Vis::Pub)], &["gean_core"]);
        file(
            &mut cache,
            "c.py",
            &[("c_fn", Vis::Pub)],
            &["gean_core", "a_fn"],
        );
        let r = rank(&cache);
        let core = r.file_rank["core.py"];
        assert!(core > r.file_rank["a.py"], "core outranks a");
        assert!(core > r.file_rank["b.py"]);
        assert_eq!(r.name_refs["gean_core"], 9, "3 callers x call-weight 3");
        assert_eq!(r.name_refs["a_fn"], 3, "1 caller x call-weight 3");
        let total: f64 = r.file_rank.values().sum();
        assert!((total - 1.0).abs() < 1e-6, "rank mass conserved: {total}");
    }

    #[test]
    fn private_defs_attract_no_edges_and_self_refs_ignored() {
        let mut cache = ScanCache::default();
        file(
            &mut cache,
            "x.py",
            &[("_hidden", Vis::Priv), ("shown", Vis::Pub)],
            &["shown"],
        );
        file(&mut cache, "y.py", &[], &["_hidden"]);
        let r = rank(&cache);
        assert!(
            !r.name_refs.contains_key("_hidden"),
            "private never counted"
        );
        assert!(
            !r.name_refs.contains_key("shown"),
            "self-ref not cross-file"
        );
    }
}