skill-inject 0.9.0

//! Hybrid ranking: cosine(query, skill-description) + context blend + file boost
//! + ambient project boost + keyword boost + phrase boost.

use crate::config::Config;
use crate::index::Index;
use crate::text::{match_tokens, norm_token, tokenize};
use std::collections::{BTreeSet, HashSet};

#[derive(Clone, Debug)]
pub struct Hit {
    pub id: String,
    pub name: String,
    /// Cosine of the *current prompt* against the skill — kept pure (never folded
    /// with the context blend) so confidence/agreement gates can still read the
    /// prompt's own signal.
    pub cosine: f32,
    /// Boost from conversational context (see [`rank_all_ctx`]). Zero when the
    /// context feature is off, the prompt is confident, or the skill is no more
    /// context-relevant than average. Kept separate from `cosine` for attribution.
    pub context: f32,
    /// Boost from a referenced file of this skill's type (see
    /// [`crate::context::file_ids`]). Zero unless a matching file was named in the
    /// prompt or recent context. Separate for attribution — the highest-precision,
    /// directly-attributable context signal.
    pub file: f32,
    /// Boost from the working directory's project ecosystem (see
    /// [`crate::context::project_terms`] / [`crate::context::skills_for_terms`]).
    /// Zero unless the channel is on and the skill's ecosystem matches; gated on
    /// `cosine >= min_similarity - PROJECT_GATE_SLACK` so this ambient signal can
    /// lift a *near*-floor ecosystem skill over the line but never rescue a
    /// clearly-irrelevant one. Separate for attribution.
    pub project: f32,
    pub keyword: f32,
    /// Boost from matched trigger phrases (see [`phrase_score`]).
    pub phrase: f32,
    pub score: f32,
}

impl Hit {
    /// The stage-1 hybrid score: the sum of every channel. The single source for
    /// the `score` field, the reranker's stage-1 agreement gate
    /// ([`crate::rerank::passes`]), and `ski why`'s breakdown display — so the
    /// channel set can never drift apart across the three (it previously did: two
    /// call sites silently omitted `project`).
    pub fn stage1_score(&self) -> f32 {
        self.cosine + self.context + self.file + self.project + self.keyword + self.phrase
    }

    /// The per-channel contributions, in summation order, for attribution display.
    pub fn breakdown(&self) -> [(&'static str, f32); 6] {
        [
            ("cos", self.cosine),
            ("ctx", self.context),
            ("file", self.file),
            ("project", self.project),
            ("kw", self.keyword),
            ("ph", self.phrase),
        ]
    }
}

/// Cosine similarity. `0.0` on a dimension mismatch — rather than silently
/// zipping to the shorter vector (a meaningless partial dot product) — since a
/// query and an index entry from different embedders/dimensions should never be
/// compared at all; the `model == id()` guard in `hook::load_or_build_index`
/// normally prevents this, but a hand-edited or same-id-different-dim index
/// should score as "no match", not a truncated garbage value.
pub fn cosine(a: &[f32], b: &[f32]) -> f32 {
    if a.len() != b.len() {
        return 0.0;
    }
    let (mut dot, mut na, mut nb) = (0f32, 0f32, 0f32);
    for (x, y) in a.iter().zip(b.iter()) {
        dot += x * y;
        na += x * x;
        nb += y * y;
    }
    if na == 0.0 || nb == 0.0 {
        return 0.0;
    }
    let c = dot / (na.sqrt() * nb.sqrt());
    // A corrupt index (a hand-edited or overflowed vector, e.g. `1e999` parsing
    // to +inf) yields a NaN here; NaN compares Equal to everything in the sort
    // below, so the corrupt entry could silently claim rank 0. Treat it as "no
    // signal" instead.
    if c.is_finite() {
        c
    } else {
        0.0
    }
}

/// Descending comparator for sorting by score. `f32::partial_cmp` returns `None`
/// only when a NaN is involved; the common `.unwrap_or(Ordering::Equal)` fallback
/// then makes a NaN compare equal to everything, which can leave it sorted into
/// rank 0 (a stable sort keeps *some* input order among "equal" elements, and a
/// NaN score should never win). This instead sorts any NaN strictly last,
/// regardless of which side of the comparison it's on.
pub fn cmp_score_desc(a: f32, b: f32) -> std::cmp::Ordering {
    match (a.is_nan(), b.is_nan()) {
        (false, false) => b.partial_cmp(&a).unwrap_or(std::cmp::Ordering::Equal),
        (true, true) => std::cmp::Ordering::Equal,
        (true, false) => std::cmp::Ordering::Greater, // a is NaN -> sorts after b
        (false, true) => std::cmp::Ordering::Less,    // b is NaN -> sorts after a
    }
}

/// Keyword channel: `boost` per keyword found in the prompt. Both sides are
/// normalized through [`norm_token`] at match time, so "make some charts" still
/// hits a `chart` keyword — the surface channels shouldn't lose a match to
/// trivial inflection the dense channel shrugs off.
pub fn keyword_score(prompt: &str, keywords: &[String], boost: f32) -> f32 {
    let toks: HashSet<String> = tokenize(prompt).iter().map(|t| norm_token(t)).collect();
    let hits = keywords
        .iter()
        .filter(|k| toks.contains(&norm_token(k)))
        .count();
    hits as f32 * boost
}

/// Phrase channel: `boost` per trigger phrase whose every content token appears in
/// the prompt. A phrase is the normalized (content-token) form produced by
/// [`crate::skill::extract_phrases`]; requiring *all* its tokens (>=2 by
/// construction) keeps the signal high-precision, so it lifts a skill on the
/// exact wording the bi-encoder dilutes without firing on incidental overlap.
/// Tokens on both sides are singular-normalized ([`norm_token`]) at match time,
/// so a stored `merge pdf files` still fires on "merge these pdf file chunks".
pub fn phrase_score(prompt: &str, phrases: &[String], boost: f32) -> f32 {
    if phrases.is_empty() {
        return 0.0;
    }
    let toks: HashSet<String> = match_tokens(prompt).into_iter().collect();
    let hits = phrases
        .iter()
        .filter(|p| {
            let mut pt = p.split_whitespace().peekable();
            pt.peek().is_some() && pt.all(|t| toks.contains(&norm_token(t)))
        })
        .count();
    hits as f32 * boost
}

/// How far below `min_similarity` a skill's own cosine may sit and still receive
/// the ambient project boost. The project signal is present on every turn, so it
/// must not resurrect arbitrary skills — but the whole point of the channel is to
/// surface the workspace's ecosystem skill on prompts that are *about* the project
/// without naming the ecosystem ("add that dependency", "set up tests"), whose
/// cosine hovers just under the floor. A small slack lets the boost carry exactly
/// those over the line (ski deliberately errs toward surfacing; the model ignores
/// a skill it doesn't need, and per-session dedup caps the cost at one showing),
/// while a clearly-off-topic skill stays gated out. Mirrors
/// [`crate::rerank`]'s `AGREEMENT_SLACK` shape.
pub const PROJECT_GATE_SLACK: f32 = 0.06;

/// Effective context-blend weight for a prompt whose best self-match cosine is
/// `prompt_top`. Scales from `cfg.context_weight` (a *fully vague* prompt,
/// `prompt_top <= vague_lo`) down to `0` (a *confident* prompt,
/// `prompt_top >= vague_hi`), linearly between. So a specific prompt ignores
/// context — avoiding the redundancy that regressed bi-encoder mean-centering
/// (see `crate::rerank` module docs) — while a vague follow-up leans on it.
/// Returns `0` whenever the feature is disabled (`context_weight <= 0` or
/// `context_depth == 0`).
pub fn context_weight(prompt_top: f32, cfg: &Config) -> f32 {
    if cfg.context_weight <= 0.0 || cfg.context_depth == 0 {
        return 0.0;
    }
    let (lo, hi) = (cfg.vague_lo, cfg.vague_hi);
    let vagueness = if hi <= lo {
        // Degenerate band: a hard step at `hi`.
        if prompt_top >= hi {
            0.0
        } else {
            1.0
        }
    } else {
        ((hi - prompt_top) / (hi - lo)).clamp(0.0, 1.0)
    };
    cfg.context_weight * vagueness
}

/// All skills, scored and sorted by descending hybrid score. No threshold — for
/// `ski why` and as input to [`select`]. No conversational context.
pub fn rank_all(query: &[f32], prompt: &str, index: &Index, cfg: &Config) -> Vec<Hit> {
    rank_all_ctx(
        query,
        None,
        &BTreeSet::new(),
        &BTreeSet::new(),
        prompt,
        index,
        cfg,
    )
}

/// Like [`rank_all`], but blends an optional conversational-context vector into
/// each skill's score. The blend is gated by how *vague* the current prompt is
/// ([`context_weight`]) and is a *relative* signal: a skill is boosted only in
/// proportion to how much more context-relevant it is than the average skill
/// (`cos(context, skill) - mean`), clamped at 0. That self-normalization is what
/// keeps the anisotropic bge floor (every skill cosines ~0.5 to anything) from
/// uniformly inflating scores and manufacturing false injects.
///
/// `file_ids` carries the file-type channel: any skill whose id is in the set
/// (a file of its type was named in the prompt/context — see
/// [`crate::context::file_ids`]) gets a flat `cfg.file_boost`, *not* gated on
/// vagueness, since a named file is unambiguous.
///
/// `project_ids` carries the ambient project-type channel (see
/// [`crate::context::project_terms`] / [`crate::context::skills_for_terms`]): a
/// skill whose ecosystem matches the working directory's manifests (or a code file
/// referenced in the conversation) gets `cfg.project_boost`, but — because this
/// signal is present every turn — only when the skill's own cosine is within
/// [`PROJECT_GATE_SLACK`] of `cfg.min_similarity`. So it lifts near-plausible
/// ecosystem skills over the floor and reorders among plausible ones, but cannot
/// rescue a clearly-irrelevant skill.
///
/// With `context = None`, empty `file_ids`/`project_ids`, and the features
/// disabled, this is identical to [`rank_all`].
pub fn rank_all_ctx(
    query: &[f32],
    context: Option<&[f32]>,
    file_ids: &BTreeSet<String>,
    project_ids: &BTreeSet<String>,
    prompt: &str,
    index: &Index,
    cfg: &Config,
) -> Vec<Hit> {
    // The prompt's own cosines; their max gauges prompt specificity, which sets
    // how much (if any) context is allowed to contribute.
    let prompt_cos: Vec<f32> = index
        .skills
        .iter()
        .map(|e| cosine(query, &e.embedding))
        .collect();
    let prompt_top = prompt_cos.iter().copied().fold(0.0_f32, f32::max);
    let lambda = match context {
        Some(_) => context_weight(prompt_top, cfg),
        None => 0.0,
    };

    // Context cosines and their mean (the relative-boost baseline), computed once.
    let ctx_cos: Vec<f32> = match (lambda > 0.0, context) {
        (true, Some(c)) => index
            .skills
            .iter()
            .map(|e| cosine(c, &e.embedding))
            .collect(),
        _ => Vec::new(),
    };
    let ctx_mean = if ctx_cos.is_empty() {
        0.0
    } else {
        ctx_cos.iter().sum::<f32>() / ctx_cos.len() as f32
    };

    let mut hits: Vec<Hit> = index
        .skills
        .iter()
        .enumerate()
        .map(|(i, e)| {
            let cosine = prompt_cos[i];
            let context = ctx_cos
                .get(i)
                .map(|&c| lambda * (c - ctx_mean).max(0.0))
                .unwrap_or(0.0);
            let file = if cfg.file_boost > 0.0 && file_ids.contains(&e.id) {
                cfg.file_boost
            } else {
                0.0
            };
            // Ambient project signal: gated on the skill's own cosine sitting
            // within PROJECT_GATE_SLACK of the injection floor, so it lifts the
            // workspace's ecosystem skill on near-plausible prompts but never
            // rescues a clearly-irrelevant one (the failure mode the keyword
            // channel can hit on incidental mentions).
            let project = if cfg.project_boost > 0.0
                && cosine >= cfg.min_similarity - PROJECT_GATE_SLACK
                && project_ids.contains(&e.id)
            {
                cfg.project_boost
            } else {
                0.0
            };
            let keyword = keyword_score(prompt, &e.keywords, cfg.keyword_boost);
            let phrase = phrase_score(prompt, &e.trigger_phrases, cfg.phrase_boost);
            let mut hit = Hit {
                id: e.id.clone(),
                name: e.name.clone(),
                cosine,
                context,
                file,
                project,
                keyword,
                phrase,
                score: 0.0,
            };
            hit.score = hit.stage1_score();
            hit
        })
        .collect();
    hits.sort_by(|a, b| cmp_score_desc(a.score, b.score));
    hits
}

/// Apply the injection guardrails: drop below `min_similarity`, cap at `max_skills`.
pub fn select(hits: Vec<Hit>, cfg: &Config) -> Vec<Hit> {
    hits.into_iter()
        .filter(|h| h.score >= cfg.min_similarity)
        .take(cfg.max_skills)
        .collect()
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::index::{Entry, Index};

    fn no_files() -> BTreeSet<String> {
        BTreeSet::new()
    }

    /// Context enabled with a known vague band, everything else default.
    fn ctx_cfg() -> Config {
        Config {
            context_depth: 1,
            context_weight: 0.3,
            vague_lo: 0.55,
            vague_hi: 0.65,
            file_boost: 0.0, // context-only baseline; file tests opt the channel in
            project_boost: 0.0, // likewise for the (default-on) project channel
            ..Default::default()
        }
    }

    fn idx2() -> Index {
        let entry = |id: &str, emb: Vec<f32>| Entry {
            id: id.to_string(),
            name: id.to_string(),
            description: String::new(),
            path: String::new(),
            keywords: Vec::new(),
            trigger_phrases: Vec::new(),
            body_head: String::new(),
            hash: String::new(),
            embedding: emb,
        };
        Index {
            model: "m".into(),
            dim: 2,
            skills: vec![entry("a", vec![1.0, 0.0]), entry("b", vec![0.0, 1.0])],
        }
    }

    #[test]
    fn context_weight_scales_with_vagueness() {
        let cfg = ctx_cfg(); // lo 0.55, hi 0.65, weight 0.3
        assert!((context_weight(0.50, &cfg) - 0.30).abs() < 1e-6); // <= lo: full
        assert_eq!(context_weight(0.65, &cfg), 0.0); // >= hi: none
        assert!((context_weight(0.60, &cfg) - 0.15).abs() < 1e-6); // midpoint: half
    }

    #[test]
    fn context_weight_zero_when_disabled() {
        let off_weight = Config {
            context_depth: 1,
            context_weight: 0.0,
            ..Default::default()
        };
        let off_depth = Config {
            context_depth: 0,
            context_weight: 0.3,
            ..Default::default()
        };
        assert_eq!(context_weight(0.10, &off_weight), 0.0);
        assert_eq!(context_weight(0.10, &off_depth), 0.0);
    }

    #[test]
    fn context_none_matches_plain_rank() {
        // With no context vector, scores are exactly cosine+keyword+phrase and the
        // context term is zero — identical to the pre-feature path.
        let q = [0.5, 0.5];
        let hits = rank_all_ctx(&q, None, &no_files(), &no_files(), "", &idx2(), &ctx_cfg());
        for h in &hits {
            assert_eq!(h.context, 0.0);
            assert!((h.score - h.cosine).abs() < 1e-6);
        }
    }

    #[test]
    fn vague_prompt_lets_context_break_a_tie() {
        // Prompt sits symmetrically between the two skills (cosine 0.707 to each),
        // and is vague (top 0.707 >= hi 0.65 -> NOT vague). Widen the band so it
        // counts as vague, then context pointing at `a` lifts `a` above `b`.
        let cfg = Config {
            vague_lo: 0.80,
            vague_hi: 0.90,
            ..ctx_cfg()
        };
        let q = [0.5, 0.5]; // equal cosine to a and b
        let ctx = [1.0, 0.0]; // points at a
        let hits = rank_all_ctx(&q, Some(&ctx), &no_files(), &no_files(), "", &idx2(), &cfg);
        assert_eq!(hits[0].id, "a"); // context broke the tie
        assert!(hits[0].context > 0.0);
        // `b` is no more context-relevant than average, so it gets no boost.
        let b = hits.iter().find(|h| h.id == "b").unwrap();
        assert_eq!(b.context, 0.0);
    }

    #[test]
    fn confident_prompt_suppresses_context() {
        // Prompt is exactly `a` (cosine 1.0 >= hi): context (pointing at `b`) must
        // not contribute, so no skill carries a context boost.
        let q = [1.0, 0.0];
        let ctx = [0.0, 1.0];
        let hits = rank_all_ctx(
            &q,
            Some(&ctx),
            &no_files(),
            &no_files(),
            "",
            &idx2(),
            &ctx_cfg(),
        );
        assert!(hits.iter().all(|h| h.context == 0.0));
        assert_eq!(hits[0].id, "a");
    }

    #[test]
    fn file_boost_lifts_named_skill_ungated() {
        // A referenced file boosts its skill even when the prompt is *confident*
        // about a different skill (file channel is not vagueness-gated). Prompt is
        // exactly `a`; a file of `b`'s type is named.
        let cfg = Config {
            file_boost: 0.2,
            ..ctx_cfg()
        };
        let q = [1.0, 0.0]; // confident about `a`
        let files: BTreeSet<String> = ["b".to_string()].into_iter().collect();
        let hits = rank_all_ctx(&q, None, &files, &no_files(), "", &idx2(), &cfg);
        let b = hits.iter().find(|h| h.id == "b").unwrap();
        assert!((b.file - 0.2).abs() < 1e-6); // b carries the file boost
        let a = hits.iter().find(|h| h.id == "a").unwrap();
        assert_eq!(a.file, 0.0); // a does not (no file of its type)
    }

    #[test]
    fn file_boost_off_when_zero() {
        let q = [1.0, 0.0];
        let files: BTreeSet<String> = ["b".to_string()].into_iter().collect();
        // file_boost defaults to 0.0 in ctx_cfg -> no file term anywhere.
        let hits = rank_all_ctx(&q, None, &files, &no_files(), "", &idx2(), &ctx_cfg());
        assert!(hits.iter().all(|h| h.file == 0.0));
    }

    #[test]
    fn project_boost_gated_on_cosine_floor() {
        // The ambient project signal lifts a plausible skill but is gated on the
        // skill's own cosine sitting within PROJECT_GATE_SLACK of `min_similarity`
        // (default 0.30): it can lift a near-floor ecosystem skill, but never
        // rescues a clearly-irrelevant one.
        let cfg = Config {
            project_boost: 0.2,
            ..ctx_cfg()
        };
        let proj: BTreeSet<String> = ["b".to_string()].into_iter().collect();

        // Query aligned with `b`: cosine(q,b) = 1.0 >= gate -> boost applies.
        let hits = rank_all_ctx(&[0.0, 1.0], None, &no_files(), &proj, "", &idx2(), &cfg);
        let b = hits.iter().find(|h| h.id == "b").unwrap();
        assert!((b.project - 0.2).abs() < 1e-6);

        // Query aligned with `a`: cosine(q,b) = 0.0, far below the gate -> gated
        // out despite `b` being in the project set.
        let hits = rank_all_ctx(&[1.0, 0.0], None, &no_files(), &proj, "", &idx2(), &cfg);
        let b = hits.iter().find(|h| h.id == "b").unwrap();
        assert_eq!(b.project, 0.0);
    }

    #[test]
    fn project_boost_lifts_near_floor_skill_over_the_line() {
        // A prompt about the project without naming the ecosystem: the skill's own
        // cosine sits just *under* the floor but within the slack, so the project
        // boost applies and can carry it over the injection floor. This is the
        // uv-in-a-python-repo case the channel exists for.
        let cfg = Config {
            project_boost: 0.2,
            min_similarity: 0.30,
            ..ctx_cfg()
        };
        let proj: BTreeSet<String> = ["b".to_string()].into_iter().collect();
        // cosine(q,b) ~= 0.28: sub-floor (0.30) but within the 0.06 slack.
        let q = [0.9578, 0.2873];
        let hits = rank_all_ctx(&q, None, &no_files(), &proj, "", &idx2(), &cfg);
        let b = hits.iter().find(|h| h.id == "b").unwrap();
        assert!(b.cosine < cfg.min_similarity, "cosine {}", b.cosine);
        assert!(b.cosine >= cfg.min_similarity - PROJECT_GATE_SLACK);
        assert!((b.project - 0.2).abs() < 1e-6);
        assert!(b.score >= cfg.min_similarity); // boosted over the floor
    }

    #[test]
    fn project_boost_off_when_zero() {
        // project_boost defaults to 0.0 in ctx_cfg -> no project term anywhere.
        let proj: BTreeSet<String> = ["b".to_string()].into_iter().collect();
        let hits = rank_all_ctx(
            &[0.0, 1.0],
            None,
            &no_files(),
            &proj,
            "",
            &idx2(),
            &ctx_cfg(),
        );
        assert!(hits.iter().all(|h| h.project == 0.0));
    }

    #[test]
    fn corrupt_infinite_embedding_cannot_claim_rank() {
        // An overflowed vector in a hand-edited/corrupt index (`1e999` parses to
        // +inf) used to produce a NaN cosine, which compared Equal to every score
        // and could land anywhere — including rank 0. It must score 0 instead.
        let entry = |id: &str, emb: Vec<f32>| crate::index::Entry {
            id: id.to_string(),
            name: id.to_string(),
            description: String::new(),
            path: String::new(),
            keywords: Vec::new(),
            trigger_phrases: Vec::new(),
            body_head: String::new(),
            hash: String::new(),
            embedding: emb,
        };
        let idx = Index {
            model: "m".into(),
            dim: 2,
            skills: vec![
                entry("corrupt", vec![f32::INFINITY, 0.0]),
                entry("real", vec![1.0, 0.0]),
            ],
        };
        let hits = rank_all(&[1.0, 0.0], "", &idx, &Config::default());
        assert_eq!(hits[0].id, "real");
        assert_eq!(hits.iter().find(|h| h.id == "corrupt").unwrap().score, 0.0);
        assert!(hits.iter().all(|h| h.score.is_finite()));
    }

    #[test]
    fn cosine_bounds() {
        let a = [1.0, 0.0, 0.0];
        let b = [1.0, 0.0, 0.0];
        let c = [0.0, 1.0, 0.0];
        assert!((cosine(&a, &b) - 1.0).abs() < 1e-6);
        assert!(cosine(&a, &c).abs() < 1e-6);
    }

    #[test]
    fn cosine_rejects_dimension_mismatch() {
        // A shorter/longer vector must score 0.0 (no match), not a truncated
        // partial dot product silently computed over the shared prefix.
        let a = [1.0, 0.0, 0.0];
        let b = [1.0, 0.0];
        assert_eq!(cosine(&a, &b), 0.0);
    }

    #[test]
    fn cmp_score_desc_sorts_nan_last_either_side() {
        let mut v = [f32::NAN, 0.5, 2.0, -1.0];
        v.sort_by(|a, b| cmp_score_desc(*a, *b));
        assert_eq!(&v[..3], &[2.0, 0.5, -1.0]);
        assert!(v[3].is_nan());
    }

    #[test]
    fn cmp_score_desc_regular_values_descend() {
        let mut v = vec![1.0, 3.0, 2.0];
        v.sort_by(|a, b| cmp_score_desc(*a, *b));
        assert_eq!(v, [3.0, 2.0, 1.0]);
    }

    #[test]
    fn keyword_boost_counts_matches() {
        let kw = vec!["uv".to_string(), "setup".to_string()];
        assert!((keyword_score("set up with uv", &kw, 0.1) - 0.1).abs() < 1e-6); // only "uv"
        assert!((keyword_score("uv setup now", &kw, 0.1) - 0.2).abs() < 1e-6); // both
    }

    #[test]
    fn keyword_boost_matches_across_plural_inflection() {
        // "charts" in the prompt must hit a "chart" keyword (and vice versa): the
        // surface channels normalize both sides through `norm_token` at match time.
        let kw = vec!["chart".to_string(), "dependencies".to_string()];
        assert!((keyword_score("make some charts", &kw, 0.1) - 0.1).abs() < 1e-6);
        assert!((keyword_score("add a dependency", &kw, 0.1) - 0.1).abs() < 1e-6);
    }

    #[test]
    fn phrase_matches_across_plural_inflection() {
        // Stored phrase tokens and prompt tokens are singular-normalized at match
        // time, so trivial inflection doesn't defeat a full-phrase match.
        let ph = vec!["merge pdf files".to_string()];
        assert!((phrase_score("merge these pdf file chunks", &ph, 0.2) - 0.2).abs() < 1e-6);
        assert!((phrase_score("merging is off topic here", &ph, 0.2) - 0.0).abs() < 1e-6);
    }

    #[test]
    fn phrase_fires_only_when_all_tokens_present() {
        let ph = vec!["screen reader support".to_string()];
        // Full phrase present (any order, extra words around) -> boost.
        assert!(
            (phrase_score("does my form have screen reader support today", &ph, 0.2) - 0.2).abs()
                < 1e-6
        );
        // Reordered, still all tokens present -> boost.
        assert!((phrase_score("support for a screen reader", &ph, 0.2) - 0.2).abs() < 1e-6);
    }

    #[test]
    fn phrase_does_not_fire_on_partial_overlap() {
        // Precision guard: a partial token overlap must NOT boost, or the phrase
        // channel would manufacture false positives on unrelated prompts.
        let ph = vec!["screen reader support".to_string()];
        assert_eq!(
            phrase_score("split this screen into two panes", &ph, 0.2),
            0.0
        );
        assert_eq!(
            phrase_score(
                "implement a debounce function in vanilla javascript",
                &ph,
                0.2
            ),
            0.0
        );
    }

    #[test]
    fn phrase_score_sums_distinct_phrases() {
        // Phrases are stored already normalized to content tokens (no stopwords),
        // the form `extract_phrases` produces.
        let ph = vec![
            "convert markdown pdf".to_string(),
            "merge two pdf files".to_string(),
        ];
        assert!(
            (phrase_score(
                "convert this markdown to pdf and merge two pdf files",
                &ph,
                0.2
            ) - 0.4)
                .abs()
                < 1e-6
        );
    }
}