eidos-kernel 0.1.0

use std::collections::{HashMap, HashSet};

use serde::Serialize;

use crate::schema::{Edge, EdgeBasis, Graph, Kind, Node, RelationProfile, core_relation_profile};

use super::{
    Confidence, Hit, ground, is_distinct_code_symbol_name,
    scoring::{relation_phrase, reverse_relation_phrase},
    terms_of,
};

#[derive(Serialize, Clone, Debug, Default)]
pub struct Subgraph {
    pub hits: Vec<Hit>,
    pub expanded: Vec<String>,
    pub edges: Vec<Edge>,
    /// Where to ROUTE: the canonical front-door node (top ranked hit).
    pub route: String,
    /// What to READ, in order: hits re-ranked by RAW lexical relevance (most specific match
    /// first — so an exact child page leads), then the expanded neighbourhood. Route ≠ this[0]
    /// whenever a child out-matched its skill but the skill was promoted to the route.
    pub context_order: Vec<String>,
    /// Multi-anchor task interpretation. Present when the query is better explained as a set of
    /// distinct facets connected by the graph than as a single-node route.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub compound: Option<CompoundSubgraph>,
    /// Number of candidate context nodes omitted by the pack budget.
    #[serde(default, skip_serializing_if = "is_zero_usize")]
    pub omitted_context: usize,
    /// Compact explanation of the top context candidates that were considered but not selected.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub omitted_context_candidates: Vec<OmittedContextCandidate>,
}

#[derive(Serialize, Clone, Debug, PartialEq)]
pub struct CompoundSubgraph {
    pub primary: String,
    pub anchors: Vec<CompoundAnchor>,
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub omitted_anchors: Vec<CompoundOmittedAnchor>,
    pub coverage: f64,
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub missing_terms: Vec<String>,
}

#[derive(Serialize, Clone, Debug, PartialEq, Eq)]
pub struct CompoundAnchor {
    pub node_id: String,
    pub matched_terms: Vec<String>,
    pub confidence: Confidence,
    pub score: i64,
}

#[derive(Serialize, Clone, Debug, PartialEq, Eq)]
pub struct CompoundOmittedAnchor {
    pub node_id: String,
    pub matched_terms: Vec<String>,
    pub confidence: Confidence,
    pub score: i64,
    pub reason: String,
}

#[derive(Serialize, Clone, Debug, PartialEq, Eq)]
pub struct OmittedContextCandidate {
    pub node_id: String,
    pub score: i64,
    pub reason: String,
}

fn is_zero_usize(value: &usize) -> bool {
    *value == 0
}

/// A task-shaped context SEED must be lexically relevant — its raw lexical score within this
/// fraction of the best anchor's. Resolved graph edges now carry code callers/callees, so
/// multi-term tasks no longer need low-score lexical side doors to recover related code.
const TASK_SEED_LEXICAL_RATIO: f64 = 0.5;
/// Bare symbol lookups intentionally keep a wider context: agents usually need the symbol plus its
/// overview docs and nearby code, and there are few query terms to distinguish task intent.
const SYMBOL_SEED_LEXICAL_RATIO: f64 = 0.3;
/// Default context-pack budget. This is intentionally larger than a chat-sized "answer" because
/// workflow planning needs code + docs + tests, but small enough to avoid graph dumps on real repos.
const DEFAULT_CONTEXT_BUDGET: usize = 48;
/// Keep anchors tight. Too many same-query anchors usually means path/name noise rather than a
/// better read plan; expansion will bring back strongly connected support when it is useful.
const MAX_SEED_CONTEXT: usize = 16;
/// Avoid one source file monopolising the pack when a large file has many section/symbol nodes.
const MAX_ITEMS_PER_SOURCE: usize = 8;
/// Keep omission transparency useful without turning focus into a graph dump.
const MAX_OMITTED_CONTEXT_CANDIDATES: usize = 12;

/// Ground, then expand along edges (bidirectional, depth/width-bounded BFS) to surface the
/// connected evidence chain — the deterministic "associative reading" of AriGraph/SAGE.
/// `limit` anchors, `depth` hops, `width` neighbours expanded per node per hop.
pub fn ground_subgraph(
    graph: &Graph,
    query: &str,
    limit: usize,
    depth: usize,
    width: usize,
) -> Subgraph {
    let hits = ground(graph, query, limit);
    let top_is_confident = hits.first().is_some_and(|hit| {
        matches!(
            hit.confidence,
            Confidence::Exact | Confidence::Strong | Confidence::Ambiguous
        )
    });

    // Seeds = the route (always) + anchors whose RAW lexical relevance clears the ratio bar.
    // Dominance-boosted-but-lexically-empty anchors are excluded, so their subtree never seeds.
    let terms = terms_of(&query.to_lowercase());
    let route = select_subgraph_route(&hits, graph, &terms);
    let seed_ratio = if terms.len() <= 1 {
        SYMBOL_SEED_LEXICAL_RATIO
    } else {
        TASK_SEED_LEXICAL_RATIO
    };
    let walk_references = should_walk_references(graph, &terms, hits.first());
    let top_lexical = hits.iter().map(|h| h.lexical_score).max().unwrap_or(0) as f64;
    let floor = (seed_ratio * top_lexical).ceil() as i64;
    let route_is_anchored =
        node_by_id(graph, &route).is_some_and(|node| route_term_count(node, &terms) > 0);
    let mut seeds: Vec<&Hit> = hits
        .iter()
        .filter(|h| {
            (h.id == route || h.lexical_score >= floor)
                && (!top_is_confident
                    || h.confidence != Confidence::Fallback
                    || is_high_value_support_seed(graph, h, &terms))
                && should_seed_hit(graph, h, &route, route_is_anchored, &terms)
        })
        .collect();
    let mut seed_id_set = seeds
        .iter()
        .map(|hit| hit.id.as_str())
        .collect::<HashSet<_>>();
    for hit in &hits {
        if seed_id_set.contains(hit.id.as_str()) || hit.lexical_score < floor {
            continue;
        }
        if !is_connected_code_support_seed(graph, hit, &terms, &seed_id_set, walk_references) {
            continue;
        }
        seed_id_set.insert(hit.id.as_str());
        seeds.push(hit);
    }
    let seed_ids: Vec<String> = seeds.iter().map(|h| h.id.clone()).collect();

    let mut visited: HashSet<String> = seed_ids.iter().cloned().collect();
    let mut frontier: Vec<String> = seed_ids.clone();
    let mut edges: Vec<Edge> = Vec::new();
    let mut edge_seen: HashSet<(String, String, String)> = HashSet::new();
    let node_by_id_map: HashMap<&str, &Node> = graph
        .nodes
        .iter()
        .map(|node| (node.id.as_str(), node))
        .collect();

    for _ in 0..depth {
        let mut next: Vec<String> = Vec::new();
        for id in &frontier {
            let mut neighbours: Vec<(&String, &Edge)> = Vec::new();
            for e in &graph.edges {
                if !is_trusted_edge(e, walk_references, &graph.relation_profiles) {
                    continue;
                }
                if &e.from == id {
                    neighbours.push((&e.to, e));
                } else if &e.to == id {
                    neighbours.push((&e.from, e));
                }
            }
            neighbours.sort_by(|(a_id, a_edge), (b_id, b_edge)| {
                walk_neighbour_score(
                    id,
                    b_id,
                    b_edge,
                    &terms,
                    &node_by_id_map,
                    &graph.relation_profiles,
                )
                .cmp(&walk_neighbour_score(
                    id,
                    a_id,
                    a_edge,
                    &terms,
                    &node_by_id_map,
                    &graph.relation_profiles,
                ))
                .then_with(|| a_id.cmp(b_id))
            });
            for (other, e) in neighbours.into_iter().take(width) {
                if edge_seen.insert((e.from.clone(), e.to.clone(), e.relation.clone())) {
                    edges.push(e.clone());
                }
                if visited.insert(other.clone()) {
                    next.push(other.clone());
                }
            }
        }
        if next.is_empty() {
            break;
        }
        frontier = next;
    }

    let seed_set: HashSet<&String> = seed_ids.iter().collect();
    let mut expanded_all: Vec<String> = visited
        .iter()
        .filter(|id| !seed_set.contains(id))
        .cloned()
        .collect();
    expanded_all.sort();
    edges.sort_by(|a, b| (&a.from, &a.to, &a.relation).cmp(&(&b.from, &b.to, &b.relation)));

    let compiled = compile_context_pack(
        graph,
        &route,
        &seeds,
        &expanded_all,
        &edges,
        &terms,
        DEFAULT_CONTEXT_BUDGET,
    );
    let compound = detect_compound_subgraph(graph, &route, &hits, &edges, &terms);
    let selected: HashSet<&String> = compiled.context_order.iter().collect();
    edges.retain(|edge| selected.contains(&edge.from) && selected.contains(&edge.to));
    let selected_seed_ids = seed_ids.iter().collect::<HashSet<_>>();
    let expanded = compiled
        .context_order
        .iter()
        .filter(|id| !selected_seed_ids.contains(*id))
        .cloned()
        .collect();

    Subgraph {
        hits,
        expanded,
        edges,
        route,
        context_order: compiled.context_order,
        compound,
        omitted_context: compiled.omitted,
        omitted_context_candidates: compiled.omitted_candidates,
    }
}

fn detect_compound_subgraph(
    graph: &Graph,
    route: &str,
    hits: &[Hit],
    edges: &[Edge],
    terms: &[String],
) -> Option<CompoundSubgraph> {
    let terms = compound_terms(terms);
    if terms.len() < 2 {
        return None;
    }

    let node_by_id: HashMap<&str, &Node> = graph
        .nodes
        .iter()
        .map(|node| (node.id.as_str(), node))
        .collect();
    let mut anchors = hits
        .iter()
        .filter_map(|hit| {
            let node = node_by_id.get(hit.id.as_str())?;
            if matches!(node.kind, Kind::Section | Kind::Unknown) {
                return None;
            }
            let matched_terms = matched_compound_terms(node, &terms);
            if matched_terms.is_empty() {
                return None;
            }
            let relevant = !matches!(hit.confidence, Confidence::Fallback)
                || matched_terms.len() >= 2
                || hit.id == route;
            relevant.then_some(CompoundAnchor {
                node_id: hit.id.clone(),
                matched_terms,
                confidence: hit.confidence,
                score: hit.score,
            })
        })
        .collect::<Vec<_>>();
    anchors.sort_by(|a, b| {
        b.matched_terms
            .len()
            .cmp(&a.matched_terms.len())
            .then_with(|| b.score.cmp(&a.score))
            .then_with(|| a.node_id.cmp(&b.node_id))
    });

    let mut covered = HashSet::new();
    let mut selected = Vec::new();
    let mut omitted_anchors = Vec::new();
    for anchor in anchors {
        let adds_new_term = anchor
            .matched_terms
            .iter()
            .any(|term| !covered.contains(term.as_str()));
        if !adds_new_term && anchor.node_id != route {
            record_omitted_anchor(&mut omitted_anchors, anchor, "covered_terms");
            continue;
        }
        if selected.len() >= 6 {
            record_omitted_anchor(&mut omitted_anchors, anchor, "anchor_limit");
            continue;
        }
        for term in &anchor.matched_terms {
            covered.insert(term.clone());
        }
        selected.push(anchor);
    }
    if selected.len() < 2 {
        return None;
    }
    let distinct_anchor_terms = selected
        .iter()
        .flat_map(|anchor| anchor.matched_terms.iter().cloned())
        .collect::<HashSet<_>>();
    if distinct_anchor_terms.len() < 2 {
        return None;
    }

    let primary = select_compound_primary(graph, route, &selected, edges, &node_by_id);
    let missing_terms = terms
        .iter()
        .filter(|term| !distinct_anchor_terms.contains(*term))
        .cloned()
        .collect::<Vec<_>>();
    let coverage = distinct_anchor_terms.len() as f64 / terms.len() as f64;

    Some(CompoundSubgraph {
        primary,
        anchors: selected,
        omitted_anchors,
        coverage,
        missing_terms,
    })
}

fn record_omitted_anchor(
    out: &mut Vec<CompoundOmittedAnchor>,
    anchor: CompoundAnchor,
    reason: &str,
) {
    if out.len() >= MAX_OMITTED_CONTEXT_CANDIDATES {
        return;
    }
    out.push(CompoundOmittedAnchor {
        node_id: anchor.node_id,
        matched_terms: anchor.matched_terms,
        confidence: anchor.confidence,
        score: anchor.score,
        reason: reason.to_string(),
    });
}

fn compound_terms(terms: &[String]) -> Vec<String> {
    let mut seen = HashSet::new();
    terms
        .iter()
        .filter(|term| !is_compound_control_term(term))
        .filter(|term| seen.insert((*term).clone()))
        .cloned()
        .collect()
}

fn is_compound_control_term(term: &str) -> bool {
    matches!(
        term,
        "combine"
            | "compared"
            | "comparing"
            | "comparison"
            | "create"
            | "draft"
            | "explain"
            | "from"
            | "generate"
            | "make"
            | "prepare"
            | "produce"
            | "risk"
            | "summarize"
            | "anchor"
            | "anchors"
            | "candidate"
            | "candidates"
            | "target"
            | "targets"
            | "year"
            | "versus"
            | "vs"
            | "write"
    )
}

fn matched_compound_terms(node: &Node, terms: &[String]) -> Vec<String> {
    let text = normalized_route_text(&format!(
        "{} {} {} {} {}",
        node.id,
        node.title,
        node.summary,
        node.aliases.join(" "),
        node.query_examples.join(" ")
    ));
    terms
        .iter()
        .filter(|term| text.contains(normalized_route_text(term).as_str()))
        .cloned()
        .collect()
}

fn select_compound_primary(
    graph: &Graph,
    route: &str,
    anchors: &[CompoundAnchor],
    edges: &[Edge],
    node_by_id: &HashMap<&str, &Node>,
) -> String {
    let anchor_ids = anchors
        .iter()
        .map(|anchor| anchor.node_id.as_str())
        .collect::<HashSet<_>>();
    if node_by_id
        .get(route)
        .is_some_and(|node| matches!(node.kind, Kind::Agent | Kind::Skill))
        && reachable_anchor_count(route, &anchor_ids, edges) >= 2
    {
        return route.to_string();
    }

    graph
        .nodes
        .iter()
        .filter(|node| matches!(node.kind, Kind::Skill | Kind::Agent))
        .filter_map(|node| {
            let reachable = reachable_anchor_count(&node.id, &anchor_ids, edges);
            (reachable >= 2).then_some((node, reachable))
        })
        .max_by(|(a_node, a_reachable), (b_node, b_reachable)| {
            a_reachable
                .cmp(b_reachable)
                .then_with(|| {
                    compound_primary_kind_priority(a_node.kind)
                        .cmp(&compound_primary_kind_priority(b_node.kind))
                })
                .then_with(|| b_node.id.cmp(&a_node.id))
        })
        .map_or_else(|| route.to_string(), |(node, _)| node.id.clone())
}

fn reachable_anchor_count(start: &str, anchor_ids: &HashSet<&str>, edges: &[Edge]) -> usize {
    let mut reached = HashSet::new();
    if anchor_ids.contains(start) {
        reached.insert(start.to_string());
    }
    let mut visited = HashSet::from([start.to_string()]);
    let mut frontier = vec![start.to_string()];
    for _ in 0..2 {
        let mut next = Vec::new();
        for id in &frontier {
            for edge in edges {
                let other = if edge.from == *id {
                    Some(edge.to.as_str())
                } else if edge.to == *id {
                    Some(edge.from.as_str())
                } else {
                    None
                };
                let Some(other) = other else {
                    continue;
                };
                if anchor_ids.contains(other) {
                    reached.insert(other.to_string());
                }
                if visited.insert(other.to_string()) {
                    next.push(other.to_string());
                }
            }
        }
        frontier = next;
        if frontier.is_empty() {
            break;
        }
    }
    reached.len()
}

fn compound_primary_kind_priority(kind: Kind) -> usize {
    match kind {
        Kind::Skill => 3,
        Kind::Agent => 2,
        _ => 0,
    }
}

fn walk_neighbour_score(
    current_id: &str,
    id: &str,
    edge: &Edge,
    terms: &[String],
    node_by_id: &HashMap<&str, &Node>,
    profiles: &std::collections::BTreeMap<String, RelationProfile>,
) -> i64 {
    let kind_score = node_by_id.get(id).map_or(0, |node| match node.kind {
        // For task packs, type/trait references usually explain the shape of a change or
        // report. Keep them ahead of same-source helper fanout when width is bounded.
        Kind::Type | Kind::Trait => 30,
        Kind::Function => 16,
        Kind::Module => 12,
        Kind::Skill | Kind::Agent => 10,
        Kind::Doc | Kind::Section => 8,
        Kind::Unknown => 0,
    });
    relation_context_score(&edge.relation, profiles)
        + relation_query_match_score(current_id, edge, terms, profiles)
        + node_query_support_score(id, terms, node_by_id)
        + kind_score
        + container_child_walk_bonus(current_id, edge, node_by_id)
}

fn container_child_walk_bonus(
    current_id: &str,
    edge: &Edge,
    node_by_id: &HashMap<&str, &Node>,
) -> i64 {
    if edge.from != current_id {
        return 0;
    }
    let Some(current) = node_by_id.get(current_id) else {
        return 0;
    };
    match (current.kind, edge.relation.as_str()) {
        (
            Kind::Doc | Kind::Section | Kind::Skill | Kind::Agent | Kind::Module | Kind::Type,
            crate::schema::relation::CONTAINS | crate::schema::relation::HAS_KNOWLEDGE,
        ) => 34,
        _ => 0,
    }
}

fn is_trusted_edge(
    edge: &Edge,
    walk_references: bool,
    profiles: &std::collections::BTreeMap<String, RelationProfile>,
) -> bool {
    use crate::schema::relation::{MENTIONS, REFERENCES};
    let profile = profiles
        .get(&edge.relation)
        .cloned()
        .or_else(|| core_relation_profile(&edge.relation));
    if profile.as_ref().is_some_and(|profile| !profile.traversable) {
        return false;
    }
    match edge.relation.as_str() {
        MENTIONS => false,
        REFERENCES => walk_references && edge.basis == EdgeBasis::Resolved,
        _ => edge.basis == EdgeBasis::Resolved,
    }
}

fn should_walk_references(graph: &Graph, terms: &[String], top_hit: Option<&Hit>) -> bool {
    if terms.len() > 1 {
        return true;
    }
    let Some(hit) = top_hit else {
        return false;
    };
    if matches!(hit.confidence, Confidence::Weak | Confidence::Fallback) {
        return false;
    }
    let Some(node) = graph.nodes.iter().find(|node| node.id == hit.id) else {
        return false;
    };
    matches!(
        node.kind,
        Kind::Function | Kind::Type | Kind::Trait | Kind::Module
    ) && is_distinct_code_symbol_name(&node.title.to_lowercase())
}

fn should_seed_hit(
    graph: &Graph,
    hit: &Hit,
    route: &str,
    route_is_anchored: bool,
    terms: &[String],
) -> bool {
    if hit.id == route || !route_is_anchored {
        return true;
    }
    node_by_id(graph, &hit.id).is_some_and(|node| {
        if matches!(
            node.kind,
            Kind::Function | Kind::Type | Kind::Trait | Kind::Module
        ) {
            route_term_count(node, terms) > 0
        } else {
            support_term_count(node, terms) > 0
        }
    })
}

fn is_high_value_support_seed(graph: &Graph, hit: &Hit, terms: &[String]) -> bool {
    node_by_id(graph, &hit.id).is_some_and(|node| {
        has_receipt(node)
            && match node.kind {
                Kind::Doc | Kind::Section | Kind::Skill | Kind::Agent => {
                    support_term_count(node, terms) >= 2
                }
                Kind::Function | Kind::Type | Kind::Trait | Kind::Module | Kind::Unknown => false,
            }
    })
}

fn is_connected_code_support_seed(
    graph: &Graph,
    hit: &Hit,
    terms: &[String],
    seed_ids: &HashSet<&str>,
    walk_references: bool,
) -> bool {
    let Some(node) = node_by_id(graph, &hit.id) else {
        return false;
    };
    if !matches!(
        node.kind,
        Kind::Function | Kind::Type | Kind::Trait | Kind::Module
    ) || !has_receipt(node)
        || support_term_count(node, terms) < 2
    {
        return false;
    }
    graph.edges.iter().any(|edge| {
        is_trusted_edge(edge, walk_references, &graph.relation_profiles)
            && ((edge.from == hit.id && seed_ids.contains(edge.to.as_str()))
                || (edge.to == hit.id && seed_ids.contains(edge.from.as_str())))
    })
}

fn select_subgraph_route(hits: &[Hit], graph: &Graph, terms: &[String]) -> String {
    let Some(top) = hits.first() else {
        return String::new();
    };
    if top.relation_anchor {
        return top.id.clone();
    }
    if hit_has_exact_query_example(top)
        && node_by_id(graph, &top.id).is_some_and(|node| route_term_count(node, terms) > 0)
    {
        return top.id.clone();
    }
    if let Some(code_route) = select_code_intent_route(hits, graph, terms) {
        return code_route;
    }
    if matches!(
        top.confidence,
        Confidence::Exact | Confidence::Strong | Confidence::Ambiguous
    ) && node_by_id(graph, &top.id).is_some_and(|node| route_term_count(node, terms) > 0)
    {
        return top.id.clone();
    }

    let node_by_id: HashMap<&str, &Node> = graph
        .nodes
        .iter()
        .map(|node| (node.id.as_str(), node))
        .collect();
    hits.iter()
        .filter_map(|hit| {
            let node = node_by_id.get(hit.id.as_str())?;
            let route_hits = route_term_count(node, terms);
            (route_hits > 0).then_some((hit, route_hits))
        })
        .max_by(|(a_hit, a_route_hits), (b_hit, b_route_hits)| {
            a_route_hits
                .cmp(b_route_hits)
                .then_with(|| a_hit.score.cmp(&b_hit.score))
                .then_with(|| b_hit.id.cmp(&a_hit.id))
        })
        .map_or_else(|| top.id.clone(), |(hit, _)| hit.id.clone())
}

fn select_code_intent_route(hits: &[Hit], graph: &Graph, terms: &[String]) -> Option<String> {
    if !query_has_code_action_intent(terms) {
        return None;
    }
    let subject_terms = code_subject_terms(terms);
    if subject_terms.is_empty() {
        return None;
    }
    let node_by_id: HashMap<&str, &Node> = graph
        .nodes
        .iter()
        .map(|node| (node.id.as_str(), node))
        .collect();
    let candidates = hits
        .iter()
        .filter_map(|hit| {
            let node = node_by_id.get(hit.id.as_str())?;
            if !is_code_route_kind(node.kind) {
                return None;
            }
            let subject_hits = route_term_count(node, &subject_terms);
            let exact_identifier_hits = code_identifier_term_count(node, &subject_terms);
            (subject_hits > 0).then_some((hit, *node, exact_identifier_hits, subject_hits))
        })
        .collect::<Vec<_>>();
    let min_score = candidates
        .iter()
        .map(|(hit, _, _, _)| hit.score)
        .max()
        .map(|score| ((score as f64) * 0.8).ceil() as i64)
        .unwrap_or_default();

    candidates
        .into_iter()
        .filter(|(hit, _, _, _)| hit.score >= min_score)
        .max_by(
            |(a_hit, a_node, a_exact_identifier_hits, a_subject_hits),
             (b_hit, b_node, b_exact_identifier_hits, b_subject_hits)| {
                a_exact_identifier_hits
                    .cmp(b_exact_identifier_hits)
                    .then_with(|| a_subject_hits.cmp(b_subject_hits))
                    .then_with(|| a_hit.score.cmp(&b_hit.score))
                    .then_with(|| a_hit.lexical_score.cmp(&b_hit.lexical_score))
                    .then_with(|| {
                        code_route_kind_priority(a_node.kind)
                            .cmp(&code_route_kind_priority(b_node.kind))
                    })
                    .then_with(|| b_hit.id.cmp(&a_hit.id))
            },
        )
        .map(|(hit, _, _, _)| hit.id.clone())
}

fn hit_has_exact_query_example(hit: &Hit) -> bool {
    hit.why.iter().any(|why| why == "exact query_example match")
}

fn code_identifier_term_count(node: &Node, terms: &[String]) -> usize {
    let title = normalized_route_text(&node.title);
    let id_tail = node
        .id
        .rsplit("::")
        .next()
        .map(normalized_route_text)
        .unwrap_or_default();
    terms
        .iter()
        .map(|term| normalized_route_text(term))
        .filter(|term| !term.is_empty() && (term == &title || term == &id_tail))
        .count()
}

fn query_has_code_action_intent(terms: &[String]) -> bool {
    terms.iter().any(|term| {
        matches!(
            term.as_str(),
            "call"
                | "caller"
                | "callee"
                | "change"
                | "changing"
                | "code"
                | "definition"
                | "function"
                | "implement"
                | "implementation"
                | "implemented"
                | "method"
                | "modify"
                | "source"
                | "trace"
                | "type"
                | "usage"
                | "use"
                | "used"
                | "uses"
        )
    })
}

fn code_subject_terms(terms: &[String]) -> Vec<String> {
    terms
        .iter()
        .filter(|term| !is_code_intent_control_term(term))
        .cloned()
        .collect()
}

fn is_code_intent_control_term(term: &str) -> bool {
    matches!(
        term,
        "call"
            | "caller"
            | "callee"
            | "change"
            | "changing"
            | "code"
            | "definition"
            | "function"
            | "implement"
            | "implementation"
            | "implemented"
            | "method"
            | "modify"
            | "source"
            | "trace"
            | "type"
            | "usage"
            | "use"
            | "used"
            | "uses"
    )
}

fn is_code_route_kind(kind: Kind) -> bool {
    matches!(
        kind,
        Kind::Function | Kind::Type | Kind::Trait | Kind::Module
    )
}

fn code_route_kind_priority(kind: Kind) -> usize {
    match kind {
        Kind::Function => 4,
        Kind::Type | Kind::Trait => 3,
        Kind::Module => 2,
        _ => 0,
    }
}

fn node_by_id<'a>(graph: &'a Graph, id: &str) -> Option<&'a Node> {
    graph.nodes.iter().find(|node| node.id == id)
}

fn route_term_count(node: &Node, terms: &[String]) -> usize {
    let title = normalized_route_text(&node.title);
    let alias_hits = if matches!(
        node.kind,
        Kind::Function | Kind::Type | Kind::Trait | Kind::Module
    ) {
        0
    } else {
        node.aliases
            .iter()
            .map(|alias| normalized_route_text(alias))
            .flat_map(|alias| {
                let alias = alias.clone();
                terms
                    .iter()
                    .filter(move |term| alias.contains(normalized_route_text(term).as_str()))
            })
            .count()
    };
    terms
        .iter()
        .filter(|term| title.contains(normalized_route_text(term).as_str()))
        .count()
        + alias_hits
}

fn support_term_count(node: &Node, terms: &[String]) -> usize {
    let mut count = route_term_count(node, terms);
    let supporting_text = normalized_route_text(&format!(
        "{} {} {}",
        node.summary,
        node.query_examples.join(" "),
        node.tags.join(" ")
    ));
    count += terms
        .iter()
        .filter(|term| supporting_text.contains(normalized_route_text(term).as_str()))
        .count();
    count
}

fn normalized_route_text(text: &str) -> String {
    text.chars()
        .filter(char::is_ascii_alphanumeric)
        .flat_map(char::to_lowercase)
        .collect()
}

#[derive(Debug, Clone)]
struct CompiledContextPack {
    context_order: Vec<String>,
    omitted: usize,
    omitted_candidates: Vec<OmittedContextCandidate>,
}

fn compile_context_pack(
    graph: &Graph,
    route: &str,
    seeds: &[&Hit],
    expanded: &[String],
    edges: &[Edge],
    terms: &[String],
    budget: usize,
) -> CompiledContextPack {
    if budget == 0 {
        return CompiledContextPack {
            context_order: Vec::new(),
            omitted: seeds.len() + expanded.len(),
            omitted_candidates: seeds
                .iter()
                .map(|hit| OmittedContextCandidate {
                    node_id: hit.id.clone(),
                    score: hit.score,
                    reason: "pack_budget".to_string(),
                })
                .chain(expanded.iter().map(|id| OmittedContextCandidate {
                    node_id: id.clone(),
                    score: 0,
                    reason: "pack_budget".to_string(),
                }))
                .take(MAX_OMITTED_CONTEXT_CANDIDATES)
                .collect(),
        };
    }

    let node_by_id: HashMap<&str, &Node> = graph
        .nodes
        .iter()
        .map(|node| (node.id.as_str(), node))
        .collect();
    let mut seen = HashSet::new();
    let mut context_order = Vec::new();
    let mut omitted_candidates = Vec::new();
    let mut per_source: HashMap<String, usize> = HashMap::new();

    let mut seed_order: Vec<&Hit> = seeds.to_vec();
    seed_order.sort_by(|a, b| {
        (b.id == route).cmp(&(a.id == route)).then_with(|| {
            b.lexical_score
                .cmp(&a.lexical_score)
                .then_with(|| a.id.cmp(&b.id))
        })
    });
    for (index, hit) in seed_order.into_iter().enumerate() {
        if index >= MAX_SEED_CONTEXT {
            record_omitted_candidate(&mut omitted_candidates, &hit.id, hit.score, "seed_limit");
            continue;
        }
        let result = push_context_item(
            &mut context_order,
            &mut seen,
            &mut per_source,
            &node_by_id,
            &hit.id,
            budget,
            true,
        );
        if let PushContextResult::Omitted(reason) = result {
            record_omitted_candidate(&mut omitted_candidates, &hit.id, hit.score, reason);
        }
    }

    let selected_seed_ids = context_order.iter().cloned().collect::<HashSet<_>>();
    let mut candidates = expanded
        .iter()
        .filter(|id| !seen.contains(id.as_str()))
        .map(|id| {
            let score = context_candidate_score(
                id,
                &selected_seed_ids,
                edges,
                terms,
                &node_by_id,
                &graph.relation_profiles,
            );
            (id.clone(), score)
        })
        .collect::<Vec<_>>();
    candidates.sort_by(|(a_id, a_score), (b_id, b_score)| {
        b_score.cmp(a_score).then_with(|| a_id.cmp(b_id))
    });
    for (id, score) in candidates {
        if context_order.len() >= budget {
            record_omitted_candidate(&mut omitted_candidates, &id, score, "pack_budget");
            break;
        }
        let result = push_context_item(
            &mut context_order,
            &mut seen,
            &mut per_source,
            &node_by_id,
            &id,
            budget,
            false,
        );
        if let PushContextResult::Omitted(reason) = result {
            record_omitted_candidate(&mut omitted_candidates, &id, score, reason);
        }
    }

    let total_candidates = seeds.len() + expanded.len();
    CompiledContextPack {
        omitted: total_candidates.saturating_sub(context_order.len()),
        context_order,
        omitted_candidates,
    }
}

enum PushContextResult {
    Selected,
    Omitted(&'static str),
}

fn push_context_item(
    out: &mut Vec<String>,
    seen: &mut HashSet<String>,
    per_source: &mut HashMap<String, usize>,
    node_by_id: &HashMap<&str, &Node>,
    id: &str,
    budget: usize,
    force: bool,
) -> PushContextResult {
    if out.len() >= budget || seen.contains(id) {
        return PushContextResult::Omitted("pack_budget");
    }
    if !force && !node_by_id.get(id).is_some_and(|node| has_receipt(node)) {
        return PushContextResult::Omitted("missing_receipt");
    }
    let source_key = node_by_id
        .get(id)
        .map_or_else(|| "(unknown)".to_string(), |node| source_key(node));
    let count = per_source.get(&source_key).copied().unwrap_or(0);
    if !force && count >= MAX_ITEMS_PER_SOURCE {
        return PushContextResult::Omitted("source_cap");
    }
    seen.insert(id.to_string());
    *per_source.entry(source_key).or_default() += 1;
    out.push(id.to_string());
    PushContextResult::Selected
}

fn record_omitted_candidate(
    out: &mut Vec<OmittedContextCandidate>,
    node_id: &str,
    score: i64,
    reason: &str,
) {
    if out.len() >= MAX_OMITTED_CONTEXT_CANDIDATES {
        return;
    }
    if out.iter().any(|candidate| candidate.node_id == node_id) {
        return;
    }
    out.push(OmittedContextCandidate {
        node_id: node_id.to_string(),
        score,
        reason: reason.to_string(),
    });
}

fn has_receipt(node: &Node) -> bool {
    node.span.is_some() || !node.source_files.is_empty()
}

fn context_candidate_score(
    id: &str,
    selected_seed_ids: &HashSet<String>,
    edges: &[Edge],
    terms: &[String],
    node_by_id: &HashMap<&str, &Node>,
    profiles: &std::collections::BTreeMap<String, RelationProfile>,
) -> i64 {
    let mut score = node_by_id
        .get(id)
        .map_or(0, |node| kind_context_score(node.kind));
    for edge in edges {
        let touches_candidate = edge.from == id || edge.to == id;
        if !touches_candidate {
            continue;
        }
        let other = if edge.from == id {
            &edge.to
        } else {
            &edge.from
        };
        let relation_score = relation_context_score(&edge.relation, profiles);
        let relation_query_score = relation_query_match_score(other, edge, terms, profiles);
        if selected_seed_ids.contains(other) {
            score += relation_score + relation_query_score + 40;
        } else {
            score += relation_score + relation_query_score;
        }
    }
    score += node_query_support_score(id, terms, node_by_id);
    score
}

fn node_query_support_score(id: &str, terms: &[String], node_by_id: &HashMap<&str, &Node>) -> i64 {
    let Some(node) = node_by_id.get(id) else {
        return 0;
    };
    let support = support_term_count(node, terms) as i64;
    if support == 0 {
        return 0;
    }
    let multiplier = match node.kind {
        Kind::Function | Kind::Type | Kind::Trait | Kind::Module => 14,
        Kind::Doc | Kind::Section | Kind::Skill | Kind::Agent => 10,
        Kind::Unknown => 0,
    };
    support * multiplier
}

fn relation_query_match_score(
    current_id: &str,
    edge: &Edge,
    terms: &[String],
    profiles: &std::collections::BTreeMap<String, RelationProfile>,
) -> i64 {
    if terms.is_empty() {
        return 0;
    }
    let phrase = if edge.from == current_id {
        relation_phrase(&edge.relation, profiles)
    } else if edge.to == current_id {
        reverse_relation_phrase(&edge.relation, profiles)
    } else {
        return 0;
    };
    let phrase_terms = terms_of(&phrase.to_lowercase());
    if phrase_terms.is_empty() {
        return 0;
    }
    let matched = phrase_terms
        .iter()
        .filter(|phrase_term| terms.iter().any(|term| term == *phrase_term))
        .count();
    if matched == 0 {
        return 0;
    }
    if matched == phrase_terms.len() {
        36
    } else {
        (matched as i64) * 12
    }
}

fn kind_context_score(kind: Kind) -> i64 {
    match kind {
        Kind::Function | Kind::Type | Kind::Trait => 16,
        Kind::Module => 10,
        Kind::Skill | Kind::Agent => 8,
        Kind::Doc | Kind::Section => 6,
        Kind::Unknown => 0,
    }
}

fn relation_context_score(
    relation: &str,
    profiles: &std::collections::BTreeMap<String, RelationProfile>,
) -> i64 {
    if let Some(profile) = profiles.get(relation) {
        return profile.context_score;
    }
    if let Some(profile) = core_relation_profile(relation) {
        return profile.context_score;
    }
    match relation {
        "depends_on" => 30,
        "produces" => 30,
        "consumes" => 28,
        "requires_contract" => 22,
        "uses_tool" => 22,
        "dispatches" => 22,
        "related_to" => 18,
        _ => 4,
    }
}

fn source_key(node: &Node) -> String {
    if let Some(span) = &node.span {
        return span.path.clone();
    }
    node.source_files
        .first()
        .cloned()
        .unwrap_or_else(|| node.id.clone())
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn focus_walks_resolved_references_but_not_lexical_references() {
        let graph = Graph {
            nodes: vec![
                test_node("fn.config::load_settings", "load_settings", Kind::Function),
                test_node(
                    "fn.worker::apply_settings",
                    "apply_settings",
                    Kind::Function,
                ),
                test_node("fn.local::maybe_settings", "maybe_settings", Kind::Function),
            ],
            edges: vec![
                Edge {
                    from: "fn.worker::apply_settings".to_string(),
                    to: "fn.config::load_settings".to_string(),
                    relation: crate::schema::relation::REFERENCES.to_string(),
                    evidence: "resolved reference".to_string(),
                    basis: EdgeBasis::Resolved,
                    ..Default::default()
                },
                Edge {
                    from: "fn.local::maybe_settings".to_string(),
                    to: "fn.config::load_settings".to_string(),
                    relation: crate::schema::relation::REFERENCES.to_string(),
                    evidence: "local lexical reference".to_string(),
                    basis: EdgeBasis::Lexical,
                    ..Default::default()
                },
            ],
            ..Default::default()
        };

        let sg = ground_subgraph(&graph, "where is load_settings used", 1, 1, 8);

        assert!(
            sg.expanded
                .contains(&"fn.worker::apply_settings".to_string())
        );
        assert!(
            !sg.expanded
                .contains(&"fn.local::maybe_settings".to_string())
        );
    }

    #[test]
    fn focus_context_is_budgeted_but_keeps_high_value_neighbours() {
        let mut nodes = vec![
            test_node("fn.config::load_settings", "load_settings", Kind::Function),
            test_node("type.config::Settings", "Settings", Kind::Type),
        ];
        let mut edges = vec![Edge {
            from: "fn.config::load_settings".to_string(),
            to: "type.config::Settings".to_string(),
            relation: crate::schema::relation::REFERENCES.to_string(),
            evidence: "resolved type reference".to_string(),
            basis: EdgeBasis::Resolved,
            ..Default::default()
        }];
        for i in 0..80 {
            let id = format!("fn.noise::helper_{i}");
            nodes.push(test_node(&id, &format!("helper_{i}"), Kind::Function));
            edges.push(Edge {
                from: "fn.config::load_settings".to_string(),
                to: id,
                relation: crate::schema::relation::CONTAINS.to_string(),
                evidence: "wide container fanout".to_string(),
                basis: EdgeBasis::Resolved,
                ..Default::default()
            });
        }
        let graph = Graph {
            nodes,
            edges,
            ..Default::default()
        };

        let sg = ground_subgraph(&graph, "load_settings", 1, 1, 100);

        assert!(sg.context_order.len() <= DEFAULT_CONTEXT_BUDGET);
        assert!(sg.omitted_context > 0);
        assert!(
            sg.context_order
                .contains(&"type.config::Settings".to_string()),
            "budgeting should preserve high-value resolved neighbours"
        );
        assert!(
            sg.omitted_context_candidates
                .iter()
                .any(|candidate| candidate.reason == "pack_budget"
                    && candidate.node_id.starts_with("fn.noise::helper_")),
            "bounded focus should expose which candidate lost the pack budget: {:#?}",
            sg.omitted_context_candidates
        );
    }

    #[test]
    fn focus_walk_prioritizes_type_references_over_lexical_neighbour_order() {
        let mut nodes = vec![
            test_node(
                "fn.engine::run_package_evals",
                "run_package_evals",
                Kind::Function,
            ),
            test_node(
                "type.engine::PackageEvalReport",
                "PackageEvalReport",
                Kind::Type,
            ),
        ];
        let mut edges = vec![Edge {
            from: "fn.engine::run_package_evals".to_string(),
            to: "type.engine::PackageEvalReport".to_string(),
            relation: crate::schema::relation::REFERENCES.to_string(),
            evidence: "resolved return type".to_string(),
            basis: EdgeBasis::Resolved,
            ..Default::default()
        }];
        for i in 0..12 {
            let id = format!("fn.alpha::helper_{i}");
            nodes.push(test_node(&id, &format!("helper_{i}"), Kind::Function));
            edges.push(Edge {
                from: "fn.engine::run_package_evals".to_string(),
                to: id,
                relation: crate::schema::relation::REFERENCES.to_string(),
                evidence: "resolved helper call".to_string(),
                basis: EdgeBasis::Resolved,
                ..Default::default()
            });
        }
        let graph = Graph {
            nodes,
            edges,
            ..Default::default()
        };

        let sg = ground_subgraph(&graph, "run_package_evals", 1, 1, 4);

        assert!(
            sg.context_order
                .contains(&"type.engine::PackageEvalReport".to_string()),
            "bounded expansion should keep the directly referenced report type"
        );
    }

    #[test]
    fn focus_walk_prioritizes_container_children_before_external_references() {
        let nodes = vec![
            test_node("type.policy::RetryPolicy", "RetryPolicy", Kind::Type),
            test_node(
                "fn.policy::RetryPolicy::should_retry",
                "should_retry",
                Kind::Function,
            ),
            test_node("type.policy::Backoff", "Backoff", Kind::Type),
            test_node("fn.queue::Task::new", "new", Kind::Function),
            test_node(
                "fn.scheduler::Scheduler::submit_fire_and_forget",
                "submit_fire_and_forget",
                Kind::Function,
            ),
        ];
        let edges = vec![
            Edge {
                from: "type.policy::RetryPolicy".to_string(),
                to: "fn.policy::RetryPolicy::should_retry".to_string(),
                relation: crate::schema::relation::CONTAINS.to_string(),
                evidence: "method".to_string(),
                basis: EdgeBasis::Resolved,
                ..Default::default()
            },
            Edge {
                from: "type.policy::RetryPolicy".to_string(),
                to: "type.policy::Backoff".to_string(),
                relation: crate::schema::relation::REFERENCES.to_string(),
                evidence: "field type".to_string(),
                basis: EdgeBasis::Resolved,
                ..Default::default()
            },
            Edge {
                from: "fn.queue::Task::new".to_string(),
                to: "type.policy::RetryPolicy".to_string(),
                relation: crate::schema::relation::REFERENCES.to_string(),
                evidence: "constructor argument".to_string(),
                basis: EdgeBasis::Resolved,
                ..Default::default()
            },
            Edge {
                from: "fn.scheduler::Scheduler::submit_fire_and_forget".to_string(),
                to: "type.policy::RetryPolicy".to_string(),
                relation: crate::schema::relation::REFERENCES.to_string(),
                evidence: "policy constructor".to_string(),
                basis: EdgeBasis::Resolved,
                ..Default::default()
            },
        ];
        let graph = Graph {
            nodes,
            edges,
            ..Default::default()
        };

        let sg = ground_subgraph(&graph, "RetryPolicy", 1, 1, 2);

        assert!(
            sg.context_order
                .contains(&"fn.policy::RetryPolicy::should_retry".to_string()),
            "bounded expansion from a type should keep its own methods before external references"
        );
    }

    #[test]
    fn focus_pack_skips_unreceipted_expanded_nodes() {
        let graph = Graph {
            nodes: vec![
                test_node("fn.config::load_settings", "load_settings", Kind::Function),
                Node {
                    id: "trait.Default".to_string(),
                    kind: Kind::Trait,
                    subkind: None,
                    title: "Default".to_string(),
                    summary: String::new(),
                    aliases: Vec::new(),
                    tags: Vec::new(),
                    query_examples: Vec::new(),
                    source_files: Vec::new(),
                    span: None,
                    partition: None,
                },
            ],
            edges: vec![Edge {
                from: "fn.config::load_settings".to_string(),
                to: "trait.Default".to_string(),
                relation: crate::schema::relation::IMPLEMENTS.to_string(),
                evidence: "synthetic trait implementation".to_string(),
                basis: EdgeBasis::Resolved,
                ..Default::default()
            }],
            ..Default::default()
        };

        let sg = ground_subgraph(&graph, "load_settings", 1, 1, 8);

        assert!(
            !sg.context_order.contains(&"trait.Default".to_string()),
            "expanded task context should stay receipt-backed"
        );
        assert!(sg.omitted_context > 0);
        assert!(
            sg.omitted_context_candidates
                .iter()
                .any(|candidate| candidate.node_id == "trait.Default"
                    && candidate.reason == "missing_receipt"),
            "focus should explain receipt-backed omissions: {:#?}",
            sg.omitted_context_candidates
        );
    }

    #[test]
    fn focus_explains_source_cap_omissions() {
        let mut nodes = vec![test_node("doc.route", "Route", Kind::Doc)];
        let mut edges = Vec::new();
        for i in 0..12 {
            let id = format!("doc.route.section-{i}");
            nodes.push(Node {
                id: id.clone(),
                kind: Kind::Section,
                subkind: None,
                title: format!("Route Section {i}"),
                summary: "Route support section.".to_string(),
                aliases: Vec::new(),
                tags: Vec::new(),
                query_examples: Vec::new(),
                source_files: vec!["fixtures/route.md".to_string()],
                span: Some(crate::schema::Span {
                    path: "fixtures/route.md".to_string(),
                    start_line: i + 1,
                    end_line: i + 1,
                }),
                partition: None,
            });
            edges.push(Edge {
                from: "doc.route".to_string(),
                to: id,
                relation: crate::schema::relation::CONTAINS.to_string(),
                evidence: "section".to_string(),
                basis: EdgeBasis::Resolved,
                ..Default::default()
            });
        }
        let graph = Graph {
            nodes,
            edges,
            ..Default::default()
        };

        let sg = ground_subgraph(&graph, "Route", 1, 1, 20);

        assert!(
            sg.omitted_context_candidates
                .iter()
                .any(|candidate| candidate.reason == "source_cap"
                    && candidate.node_id.starts_with("doc.route.section-")),
            "focus should explain when same-source candidates are capped: {:#?}",
            sg.omitted_context_candidates
        );
    }

    #[test]
    fn relation_profiles_can_disable_focus_traversal() {
        let graph = Graph {
            nodes: vec![
                test_node("doc.seed", "Seed", Kind::Doc),
                test_node("doc.external", "External", Kind::Doc),
            ],
            edges: vec![Edge {
                from: "doc.seed".to_string(),
                to: "doc.external".to_string(),
                relation: "catalogs".to_string(),
                evidence: "frontmatter".to_string(),
                basis: EdgeBasis::Resolved,
                ..Default::default()
            }],
            relation_profiles: [(
                "catalogs".to_string(),
                RelationProfile {
                    forward_phrase: "catalogs".to_string(),
                    reverse_phrase: "cataloged by".to_string(),
                    context_score: 30,
                    searchable: true,
                    traversable: false,
                    domain_kinds: Vec::new(),
                    range_kinds: Vec::new(),
                },
            )]
            .into_iter()
            .collect(),
        };

        let sg = ground_subgraph(&graph, "Seed", 1, 1, 8);

        assert!(!sg.context_order.contains(&"doc.external".to_string()));
    }

    #[test]
    fn focus_walk_prefers_query_supporting_code_neighbour_when_width_is_tight() {
        let graph = Graph {
            nodes: vec![
                test_node(
                    "fn.claims::Engine::aether_claim_state",
                    "aether_claim_state",
                    Kind::Function,
                ),
                test_node_with_summary(
                    "fn.claims::State::ledger_parse_errors",
                    "ledger_parse_errors",
                    Kind::Function,
                    "Count of failed ledger lines; a broken ledger leaves the claim bridge blind.",
                ),
                test_node(
                    "fn.claims::Engine::vault_root",
                    "vault_root",
                    Kind::Function,
                ),
            ],
            edges: vec![
                Edge {
                    from: "fn.claims::Engine::aether_claim_state".to_string(),
                    to: "fn.claims::State::ledger_parse_errors".to_string(),
                    relation: crate::schema::relation::REFERENCES.to_string(),
                    evidence: "resolved reference".to_string(),
                    basis: EdgeBasis::Resolved,
                    ..Default::default()
                },
                Edge {
                    from: "fn.claims::Engine::aether_claim_state".to_string(),
                    to: "fn.claims::Engine::vault_root".to_string(),
                    relation: crate::schema::relation::REFERENCES.to_string(),
                    evidence: "resolved reference".to_string(),
                    basis: EdgeBasis::Resolved,
                    ..Default::default()
                },
            ],
            ..Default::default()
        };

        let sg = ground_subgraph(&graph, "debug claim ledger aether_claim_state", 1, 1, 1);

        assert!(
            sg.context_order
                .contains(&"fn.claims::State::ledger_parse_errors".to_string()),
            "task-matching code support should beat generic code neighbours under tight width"
        );
    }

    #[test]
    fn connected_code_support_hit_can_seed_without_title_match() {
        let graph = Graph {
            nodes: vec![
                test_node(
                    "fn.claims::Engine::aether_claim_state",
                    "aether_claim_state",
                    Kind::Function,
                ),
                test_node_with_summary(
                    "fn.claims::State::ledger_parse_errors",
                    "ledger_parse_errors",
                    Kind::Function,
                    "Count of failed ledger lines; a broken ledger leaves the claim bridge blind.",
                ),
                test_node_with_summary(
                    "fn.misc::run",
                    "run",
                    Kind::Function,
                    "Debug claim ledger helper with no relation to the route.",
                ),
            ],
            edges: vec![Edge {
                from: "fn.claims::Engine::aether_claim_state".to_string(),
                to: "fn.claims::State::ledger_parse_errors".to_string(),
                relation: crate::schema::relation::REFERENCES.to_string(),
                evidence: "resolved reference".to_string(),
                basis: EdgeBasis::Resolved,
                ..Default::default()
            }],
            ..Default::default()
        };

        let sg = ground_subgraph(&graph, "debug claim ledger aether_claim_state", 3, 0, 0);

        assert!(
            sg.context_order
                .contains(&"fn.claims::State::ledger_parse_errors".to_string()),
            "connected support code should seed beside the chosen route"
        );
        assert!(
            !sg.context_order.contains(&"fn.misc::run".to_string()),
            "unconnected body-only code should remain out of the seed set"
        );
    }

    #[test]
    fn focus_route_keeps_relation_anchored_source_over_named_target() {
        let graph = Graph {
            nodes: vec![
                test_node("doc.contract", "Deploy Contract", Kind::Doc),
                test_node(
                    "fn.release::check_deploy_readiness",
                    "check_deploy_readiness",
                    Kind::Function,
                ),
            ],
            edges: vec![Edge {
                from: "doc.contract".to_string(),
                to: "fn.release::check_deploy_readiness".to_string(),
                relation: "verifies".to_string(),
                evidence: "frontmatter".to_string(),
                basis: EdgeBasis::Resolved,
                ..Default::default()
            }],
            relation_profiles: [(
                "verifies".to_string(),
                RelationProfile::new("verifies", "verified by", 30),
            )]
            .into_iter()
            .collect(),
        };

        let sg = ground_subgraph(&graph, "what verifies check_deploy_readiness", 5, 1, 4);

        assert_eq!(
            sg.hits.first().map(|hit| hit.id.as_str()),
            Some("doc.contract")
        );
        assert_eq!(
            sg.route, "doc.contract",
            "focus routing should preserve a typed relation anchor instead of rerouting to the named endpoint"
        );
    }

    #[test]
    fn focus_expansion_prefers_query_named_relation_when_width_is_tight() {
        let graph = Graph {
            nodes: vec![
                test_node("doc.gate", "Release Gate", Kind::Doc),
                test_node("doc.runbook", "Deploy Runbook", Kind::Doc),
                test_node("fn.release::check", "check", Kind::Function),
            ],
            edges: vec![
                Edge {
                    from: "doc.gate".to_string(),
                    to: "doc.runbook".to_string(),
                    relation: "blocks".to_string(),
                    evidence: "frontmatter".to_string(),
                    basis: EdgeBasis::Resolved,
                    ..Default::default()
                },
                Edge {
                    from: "doc.gate".to_string(),
                    to: "fn.release::check".to_string(),
                    relation: "verifies".to_string(),
                    evidence: "frontmatter".to_string(),
                    basis: EdgeBasis::Resolved,
                    ..Default::default()
                },
            ],
            relation_profiles: [
                (
                    "blocks".to_string(),
                    RelationProfile::new("blocks", "blocked by", 31),
                ),
                (
                    "verifies".to_string(),
                    RelationProfile::new("verifies", "verified by", 29),
                ),
            ]
            .into_iter()
            .collect(),
        };

        let sg = ground_subgraph(&graph, "release gate verifies", 1, 1, 1);

        assert!(
            sg.context_order.contains(&"fn.release::check".to_string()),
            "the relation named in the query should win bounded focus expansion"
        );
        assert!(
            !sg.context_order.contains(&"doc.runbook".to_string()),
            "a slightly higher static context_score should not beat the query's explicit relation intent"
        );
    }

    #[test]
    fn focus_route_prefers_title_anchored_fallback_over_body_only_noise() {
        let graph = Graph {
            nodes: vec![
                test_node_with_summary(
                    "fn.tests::run",
                    "run",
                    Kind::Function,
                    "trace package enforcement eidos harness helper",
                ),
                test_node_with_summary(
                    "fn.engine::check_eval_gates",
                    "check_eval_gates",
                    Kind::Function,
                    "enforces configured package gates",
                ),
            ],
            edges: Vec::new(),
            ..Default::default()
        };

        let sg = ground_subgraph(&graph, "trace package eval gate enforcement eidos", 2, 1, 8);

        assert_eq!(sg.route, "fn.engine::check_eval_gates");
        assert_eq!(
            sg.context_order.first().map(String::as_str),
            Some("fn.engine::check_eval_gates")
        );
        assert!(
            !sg.context_order.contains(&"fn.tests::run".to_string()),
            "body-only fallback hits should stay in hits but not seed the focus pack"
        );
    }

    #[test]
    fn focus_route_prefers_code_anchor_for_code_action_query() {
        let graph = Graph {
            nodes: vec![
                test_node_with_summary(
                    "skill.retry-policy",
                    "retry-policy",
                    Kind::Skill,
                    "Guide for retry flow changes and delay_ms safety.",
                ),
                test_node_with_summary(
                    "fn.policy::RetryPolicy::delay_ms",
                    "delay_ms",
                    Kind::Function,
                    "Computes the retry delay.",
                ),
            ],
            edges: Vec::new(),
            ..Default::default()
        };
        let hits = vec![
            Hit {
                id: "skill.retry-policy".to_string(),
                score: 120,
                lexical_score: 120,
                confidence: Confidence::Strong,
                why: Vec::new(),
                relation_matches: Vec::new(),
                anchor: 1.0,
                relation_anchor: false,
            },
            Hit {
                id: "fn.policy::RetryPolicy::delay_ms".to_string(),
                score: 60,
                lexical_score: 60,
                confidence: Confidence::Weak,
                why: Vec::new(),
                relation_matches: Vec::new(),
                anchor: 1.0,
                relation_anchor: false,
            },
        ];

        let route = select_subgraph_route(
            &hits,
            &graph,
            &[
                "trace".to_string(),
                "delayms".to_string(),
                "used".to_string(),
                "retry".to_string(),
                "flow".to_string(),
            ],
        );

        assert_eq!(route, "fn.policy::RetryPolicy::delay_ms");
    }

    #[test]
    fn code_action_route_prefers_explicit_identifier_over_related_helper_terms() {
        let graph = Graph {
            nodes: vec![
                test_node(
                    "fn.billing::sync_invoice_batch",
                    "sync_invoice_batch",
                    Kind::Function,
                ),
                test_node(
                    "fn.billing::build_idempotency_key",
                    "build_idempotency_key",
                    Kind::Function,
                ),
            ],
            edges: Vec::new(),
            ..Default::default()
        };
        let hits = vec![
            Hit {
                id: "fn.billing::sync_invoice_batch".to_string(),
                score: 81,
                lexical_score: 81,
                confidence: Confidence::Strong,
                why: Vec::new(),
                relation_matches: Vec::new(),
                anchor: 1.0,
                relation_anchor: false,
            },
            Hit {
                id: "fn.billing::build_idempotency_key".to_string(),
                score: 36,
                lexical_score: 64,
                confidence: Confidence::Fallback,
                why: Vec::new(),
                relation_matches: Vec::new(),
                anchor: 0.8,
                relation_anchor: false,
            },
        ];

        let route = select_subgraph_route(
            &hits,
            &graph,
            &[
                "trace".to_string(),
                "sync_invoice_batch".to_string(),
                "build".to_string(),
                "idempotency".to_string(),
                "key".to_string(),
                "billing".to_string(),
                "sync".to_string(),
            ],
        );

        assert_eq!(route, "fn.billing::sync_invoice_batch");
    }

    #[test]
    fn focus_route_keeps_exact_query_example_front_door_before_code_override() {
        let mut doc = test_node("doc.mcp-prompts", "MCP Prompt Workflows", Kind::Doc);
        doc.query_examples = vec!["how should coding agents use Eidos MCP prompts".to_string()];
        let graph = Graph {
            nodes: vec![
                doc,
                test_node(
                    "fn.eidos::mcp::start_coding_task",
                    "start_coding_task",
                    Kind::Function,
                ),
            ],
            edges: Vec::new(),
            ..Default::default()
        };
        let hits = vec![
            Hit {
                id: "doc.mcp-prompts".to_string(),
                score: 171,
                lexical_score: 171,
                confidence: Confidence::Ambiguous,
                why: vec!["exact query_example match".to_string()],
                relation_matches: Vec::new(),
                anchor: 1.0,
                relation_anchor: false,
            },
            Hit {
                id: "fn.eidos::mcp::start_coding_task".to_string(),
                score: 146,
                lexical_score: 146,
                confidence: Confidence::Strong,
                why: Vec::new(),
                relation_matches: Vec::new(),
                anchor: 1.0,
                relation_anchor: false,
            },
        ];

        let route = select_subgraph_route(
            &hits,
            &graph,
            &[
                "code".to_string(),
                "agent".to_string(),
                "use".to_string(),
                "eidos".to_string(),
                "mcp".to_string(),
                "prompt".to_string(),
            ],
        );

        assert_eq!(route, "doc.mcp-prompts");
    }

    #[test]
    fn code_action_route_prefers_ranked_hit_before_loose_subject_count() {
        let graph = Graph {
            nodes: vec![
                test_node(
                    "fn.eidos::mcp::EidosMcpServer::deep_code_review",
                    "deep_code_review",
                    Kind::Function,
                ),
                test_node(
                    "fn.eidos::mcp_consent_guard::mcp_source_never_references_the_sigil_write_surface",
                    "mcp_source_never_references_the_sigil_write_surface",
                    Kind::Function,
                ),
            ],
            edges: Vec::new(),
            ..Default::default()
        };
        let hits = vec![
            Hit {
                id: "fn.eidos::mcp::EidosMcpServer::deep_code_review".to_string(),
                score: 160,
                lexical_score: 160,
                confidence: Confidence::Ambiguous,
                why: Vec::new(),
                relation_matches: Vec::new(),
                anchor: 1.0,
                relation_anchor: false,
            },
            Hit {
                id: "fn.eidos::mcp_consent_guard::mcp_source_never_references_the_sigil_write_surface"
                    .to_string(),
                score: 123,
                lexical_score: 123,
                confidence: Confidence::Strong,
                why: Vec::new(),
                relation_matches: Vec::new(),
                anchor: 1.0,
                relation_anchor: false,
            },
        ];

        let route = select_subgraph_route(
            &hits,
            &graph,
            &[
                "review".to_string(),
                "eidos".to_string(),
                "mcp".to_string(),
                "prompt".to_string(),
                "surface".to_string(),
                "implementation".to_string(),
            ],
        );

        assert_eq!(route, "fn.eidos::mcp::EidosMcpServer::deep_code_review");
    }

    #[test]
    fn high_value_receipted_docs_can_seed_beside_confident_code() {
        let mut doc = test_node_with_summary(
            "doc.billing",
            "Billing Sync Runbook",
            Kind::Doc,
            "The idempotency key protects provider calls.",
        );
        doc.source_files.push("docs/billing.md".to_string());
        let graph = Graph {
            nodes: vec![doc],
            edges: Vec::new(),
            ..Default::default()
        };
        let hit = Hit {
            id: "doc.billing".to_string(),
            score: 37,
            lexical_score: 56,
            confidence: Confidence::Fallback,
            why: Vec::new(),
            relation_matches: Vec::new(),
            anchor: 0.5,
            relation_anchor: false,
        };

        assert!(is_high_value_support_seed(
            &graph,
            &hit,
            &[
                "idempotency".to_string(),
                "key".to_string(),
                "billing".to_string(),
            ]
        ));
    }

    #[test]
    fn code_action_route_requires_subject_match_not_just_control_word() {
        let graph = Graph {
            nodes: vec![
                test_node("doc.retry-policy", "Retry Policy", Kind::Doc),
                test_node("fn.misc::change", "change", Kind::Function),
            ],
            edges: Vec::new(),
            ..Default::default()
        };
        let hits = vec![
            Hit {
                id: "doc.retry-policy".to_string(),
                score: 80,
                lexical_score: 80,
                confidence: Confidence::Strong,
                why: Vec::new(),
                relation_matches: Vec::new(),
                anchor: 1.0,
                relation_anchor: false,
            },
            Hit {
                id: "fn.misc::change".to_string(),
                score: 70,
                lexical_score: 70,
                confidence: Confidence::Strong,
                why: Vec::new(),
                relation_matches: Vec::new(),
                anchor: 1.0,
                relation_anchor: false,
            },
        ];

        let route = select_subgraph_route(
            &hits,
            &graph,
            &[
                "change".to_string(),
                "retry".to_string(),
                "policy".to_string(),
            ],
        );

        assert_eq!(route, "doc.retry-policy");
    }

    #[test]
    fn route_term_matching_ignores_identifier_separators() {
        let skill = test_node("skill.retry-policy", "retry-policy", Kind::Skill);
        let type_node = test_node("type.policy::RetryPolicy", "RetryPolicy", Kind::Type);

        assert_eq!(route_term_count(&skill, &["retrypolicy".to_string()]), 1);
        assert_eq!(
            route_term_count(&type_node, &["retry_policy".to_string()]),
            1
        );
    }

    #[test]
    fn docs_can_seed_from_supporting_text_when_route_is_anchored() {
        let graph = Graph {
            nodes: vec![
                test_node("skill.retry-policy", "retry-policy", Kind::Skill),
                test_node_with_summary(
                    "doc.architecture",
                    "System Architecture",
                    Kind::Doc,
                    "The retry flow explains where delay_ms is used.",
                ),
            ],
            edges: Vec::new(),
            ..Default::default()
        };
        let hit = Hit {
            id: "doc.architecture".to_string(),
            score: 30,
            lexical_score: 30,
            confidence: Confidence::Fallback,
            why: Vec::new(),
            relation_matches: Vec::new(),
            anchor: 0.0,
            relation_anchor: false,
        };

        assert!(should_seed_hit(
            &graph,
            &hit,
            "skill.retry-policy",
            true,
            &["retry".to_string(), "flow".to_string()]
        ));
    }

    #[test]
    fn compound_terms_ignore_vocabulary_target_control_words() {
        let terms = vec![
            "agent".to_string(),
            "workflow".to_string(),
            "target".to_string(),
            "candidates".to_string(),
            "anchor".to_string(),
        ];

        assert_eq!(
            compound_terms(&terms),
            vec!["agent".to_string(), "workflow".to_string()]
        );
    }

    #[test]
    fn compound_focus_surfaces_multi_anchor_primary_skill() {
        let graph = Graph {
            nodes: vec![
                test_node_with_summary(
                    "doc.forecast-2026",
                    "2026 Revenue Forecast",
                    Kind::Doc,
                    "Baseline forecast assumptions for fiscal year 2026.",
                ),
                test_node_with_summary(
                    "doc.seattle-clients",
                    "Seattle Client Segment Map",
                    Kind::Doc,
                    "Seattle client segmentation and expansion probability.",
                ),
                test_node_with_summary(
                    "skill.client-forecast",
                    "Client Forecast Analysis",
                    Kind::Skill,
                    "Combine forecast assumptions with client segment evidence.",
                ),
            ],
            edges: vec![
                Edge {
                    from: "skill.client-forecast".to_string(),
                    to: "doc.forecast-2026".to_string(),
                    relation: "depends_on".to_string(),
                    evidence: "frontmatter".to_string(),
                    basis: EdgeBasis::Resolved,
                    ..Default::default()
                },
                Edge {
                    from: "skill.client-forecast".to_string(),
                    to: "doc.seattle-clients".to_string(),
                    relation: "depends_on".to_string(),
                    evidence: "frontmatter".to_string(),
                    basis: EdgeBasis::Resolved,
                    ..Default::default()
                },
            ],
            ..Default::default()
        };

        let sg = ground_subgraph(
            &graph,
            "forecast for year 2026 from Seattle clients",
            5,
            2,
            4,
        );
        let compound = sg.compound.expect("multi-facet query should be compound");

        assert_eq!(compound.primary, "skill.client-forecast");
        assert!(compound.coverage >= 0.75, "compound={compound:#?}");
        assert!(
            compound
                .anchors
                .iter()
                .any(|anchor| anchor.node_id == "doc.forecast-2026"
                    && anchor.matched_terms.contains(&"2026".to_string()))
        );
        assert!(
            compound
                .anchors
                .iter()
                .any(|anchor| anchor.node_id == "doc.seattle-clients"
                    && anchor.matched_terms.contains(&"seattle".to_string()))
        );
    }

    #[test]
    fn compound_focus_explains_covered_omitted_anchors() {
        let graph = Graph {
            nodes: vec![
                test_node_with_summary(
                    "doc.forecast-2026",
                    "2026 Revenue Forecast",
                    Kind::Doc,
                    "Baseline forecast assumptions for fiscal year 2026.",
                ),
                test_node_with_summary(
                    "doc.forecast-2026-alt",
                    "2026 Forecast Appendix",
                    Kind::Doc,
                    "Secondary forecast assumptions for fiscal year 2026.",
                ),
                test_node_with_summary(
                    "doc.seattle-clients",
                    "Seattle Client Segment Map",
                    Kind::Doc,
                    "Seattle client segmentation and expansion probability.",
                ),
                test_node_with_summary(
                    "skill.client-forecast",
                    "Client Forecast Analysis",
                    Kind::Skill,
                    "Combine forecast assumptions with client segment evidence.",
                ),
            ],
            edges: vec![
                Edge {
                    from: "skill.client-forecast".to_string(),
                    to: "doc.forecast-2026".to_string(),
                    relation: "depends_on".to_string(),
                    evidence: "frontmatter".to_string(),
                    basis: EdgeBasis::Resolved,
                    ..Default::default()
                },
                Edge {
                    from: "skill.client-forecast".to_string(),
                    to: "doc.forecast-2026-alt".to_string(),
                    relation: "depends_on".to_string(),
                    evidence: "frontmatter".to_string(),
                    basis: EdgeBasis::Resolved,
                    ..Default::default()
                },
                Edge {
                    from: "skill.client-forecast".to_string(),
                    to: "doc.seattle-clients".to_string(),
                    relation: "depends_on".to_string(),
                    evidence: "frontmatter".to_string(),
                    basis: EdgeBasis::Resolved,
                    ..Default::default()
                },
            ],
            ..Default::default()
        };

        let sg = ground_subgraph(
            &graph,
            "forecast for year 2026 from Seattle clients",
            8,
            2,
            4,
        );
        let compound = sg.compound.expect("multi-facet query should be compound");

        assert!(
            compound
                .omitted_anchors
                .iter()
                .any(|anchor| anchor.node_id == "doc.forecast-2026-alt"
                    && anchor.reason == "covered_terms"
                    && anchor.matched_terms.contains(&"2026".to_string())),
            "compound focus should expose skipped duplicate-facet anchors: {compound:#?}"
        );
    }

    fn test_node(id: &str, title: &str, kind: Kind) -> Node {
        test_node_with_summary(id, title, kind, "")
    }

    fn test_node_with_summary(id: &str, title: &str, kind: Kind, summary: &str) -> Node {
        Node {
            id: id.to_string(),
            kind,
            subkind: None,
            title: title.to_string(),
            summary: summary.to_string(),
            aliases: Vec::new(),
            tags: Vec::new(),
            query_examples: Vec::new(),
            source_files: vec![format!("fixtures/{title}.rs")],
            span: None,
            partition: None,
        }
    }
}