//! Nyx Benchmark Evaluation Framework
//!
//! Run with: `cargo test benchmark_evaluation -- --ignored --nocapture`
//!
//! Filter with env vars:
//!   NYX_BENCH_LANG=python
//!   NYX_BENCH_CLASS=sqli
//!   NYX_BENCH_CASE=js-sqli-001
//!   NYX_BENCH_POSITIVE_ONLY=1
//!   NYX_BENCH_NEGATIVE_ONLY=1
//!   NYX_BENCH_TAG=express

mod common;

use common::test_config;
use nyx_scanner::commands::scan::Diag;
use nyx_scanner::evidence::Confidence;
use nyx_scanner::patterns::FindingCategory;
use nyx_scanner::utils::config::AnalysisMode;
use serde::{Deserialize, Serialize};
use std::collections::BTreeMap;
use std::path::{Path, PathBuf};

// ── Ground-truth schema ──────────────────────────────────────────────

#[derive(Deserialize)]
struct GroundTruth {
    #[allow(dead_code)]
    schema_version: String,
    #[allow(dead_code)]
    metadata: Metadata,
    cases: Vec<Case>,
}

#[derive(Deserialize)]
struct Metadata {
    #[allow(dead_code)]
    description: String,
    #[allow(dead_code)]
    created: String,
    #[allow(dead_code)]
    corpus_size: usize,
}

#[derive(Deserialize)]
struct Case {
    case_id: String,
    file: String,
    language: String,
    is_vulnerable: bool,
    vuln_class: String,
    #[allow(dead_code)]
    cwe: String,
    #[allow(dead_code)]
    provenance: String,
    #[allow(dead_code)]
    equivalence_tier: String,
    #[allow(dead_code)]
    match_mode: String,
    expected_rule_ids: Vec<String>,
    allowed_alternative_rule_ids: Vec<String>,
    forbidden_rule_ids: Vec<String>,
    #[allow(dead_code)]
    expected_severity: Option<String>,
    #[allow(dead_code)]
    expected_category: Option<String>,
    expected_sink_lines: Option<Vec<[usize; 2]>>,
    #[allow(dead_code)]
    expected_source_lines: Option<Vec<[usize; 2]>>,
    /// Optional: line ranges where the *call site* leading to the sink
    /// appears. When present, at least one `flow_step` in the finding's
    /// trace must fall within ±2 of one of these ranges. When absent,
    /// the check is skipped (forward-compatible with existing fixtures).
    #[serde(default)]
    expected_call_site_lines: Option<Vec<[usize; 2]>>,
    #[allow(dead_code)]
    tags: Vec<String>,
    #[serde(default)]
    disabled: bool,
    #[allow(dead_code)]
    notes: String,
}

// ── Result types ─────────────────────────────────────────────────────

#[derive(Clone, Serialize)]
enum Outcome {
    TP,
    FP,
    FN,
    TN,
}

impl std::fmt::Display for Outcome {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            Outcome::TP => write!(f, "TP"),
            Outcome::FP => write!(f, "FP"),
            Outcome::FN => write!(f, "FN"),
            Outcome::TN => write!(f, "TN"),
        }
    }
}

#[derive(Serialize)]
struct CaseOutcome {
    case_id: String,
    file: String,
    language: String,
    vuln_class: String,
    is_vulnerable: bool,
    outcome_file_level: Outcome,
    outcome_rule_level: Outcome,
    outcome_location_level: Option<Outcome>,
    matched_rule_ids: Vec<String>,
    unexpected_rule_ids: Vec<String>,
    all_finding_ids: Vec<String>,
    security_finding_count: usize,
    non_security_finding_count: usize,
    #[serde(skip)]
    diags: Vec<Diag>,
}

#[derive(Serialize)]
struct Metrics {
    tp: usize,
    fp: usize,
    fn_: usize,
    tn: usize,
    precision: f64,
    recall: f64,
    f1: f64,
}

impl Metrics {
    fn compute(tp: usize, fp: usize, fn_: usize, tn: usize) -> Self {
        let precision = if tp + fp == 0 {
            1.0
        } else {
            tp as f64 / (tp + fp) as f64
        };
        let recall = if tp + fn_ == 0 {
            1.0
        } else {
            tp as f64 / (tp + fn_) as f64
        };
        let f1 = if precision + recall == 0.0 {
            0.0
        } else {
            2.0 * precision * recall / (precision + recall)
        };
        Metrics {
            tp,
            fp,
            fn_,
            tn,
            precision,
            recall,
            f1,
        }
    }
}

#[derive(Serialize)]
struct ScannerConfig {
    analysis_mode: String,
    taint_enabled: bool,
    ast_patterns_enabled: bool,
    state_analysis_enabled: bool,
    worker_threads: usize,
}

#[derive(Serialize)]
struct BenchmarkResults {
    benchmark_version: String,
    timestamp: String,
    scanner_version: String,
    scanner_config: ScannerConfig,
    ground_truth_hash: String,
    corpus_size: usize,
    cases_run: usize,
    cases_skipped: usize,
    outcomes: Vec<CaseOutcome>,
    aggregate_file_level: Metrics,
    aggregate_rule_level: Metrics,
    by_language: BTreeMap<String, Metrics>,
    by_vuln_class: BTreeMap<String, Metrics>,
    by_confidence: BTreeMap<String, Metrics>,
}

// ── Scanning ─────────────────────────────────────────────────────────

fn scan_corpus_file(corpus_root: &Path, relative_path: &str) -> Vec<Diag> {
    // `cve_corpus/*` cases live in a sibling of `corpus/`, see
    // `tests/benchmark/cve_corpus/`.
    let source = if relative_path.starts_with("cve_corpus/") {
        corpus_root
            .parent()
            .expect("corpus_root has a parent")
            .join(relative_path)
    } else {
        corpus_root.join(relative_path)
    };
    assert!(
        source.exists(),
        "Corpus path not found: {}",
        source.display()
    );

    let tmp = tempfile::TempDir::with_prefix("nyx_bench_").expect("tempdir");

    if source.is_dir() {
        copy_dir_recursive(&source, tmp.path());
    } else {
        let dest = tmp.path().join(source.file_name().unwrap());
        std::fs::copy(&source, &dest).expect("copy corpus file");
    }

    let cfg = test_config(AnalysisMode::Full);
    let mut diags =
        nyx_scanner::scan_no_index(tmp.path(), &cfg).expect("scan_no_index should succeed");

    // Normalize paths to filename only.
    for d in &mut diags {
        if let Some(fname) = Path::new(&d.path).file_name() {
            d.path = fname.to_string_lossy().to_string();
        }
    }

    // Sort deterministically.
    diags.sort_by(|a, b| {
        a.path
            .cmp(&b.path)
            .then(a.line.cmp(&b.line))
            .then(a.id.cmp(&b.id))
            .then(a.col.cmp(&b.col))
    });

    diags
}

fn copy_dir_recursive(src: &Path, dst: &Path) {
    for entry in std::fs::read_dir(src).expect("read corpus dir") {
        let entry = entry.expect("dir entry");
        let dest_path = dst.join(entry.file_name());
        if entry.file_type().expect("file type").is_dir() {
            std::fs::create_dir_all(&dest_path).expect("create subdir");
            copy_dir_recursive(&entry.path(), &dest_path);
        } else {
            std::fs::copy(entry.path(), &dest_path).expect("copy file");
        }
    }
}

// ── Scoring helpers ──────────────────────────────────────────────────

fn is_security(d: &Diag) -> bool {
    d.category == FindingCategory::Security
}

fn rule_matches(finding_id: &str, expected_id: &str) -> bool {
    if finding_id == expected_id {
        return true;
    }
    // Substring fallback: the expected id is contained in the finding id.
    finding_id.contains(expected_id)
}

fn score_file_level(case: &Case, diags: &[Diag]) -> Outcome {
    let has_security = diags.iter().any(is_security);
    match (case.is_vulnerable, has_security) {
        (true, true) => Outcome::TP,
        (true, false) => Outcome::FN,
        (false, true) => Outcome::FP,
        (false, false) => Outcome::TN,
    }
}

fn score_rule_level(case: &Case, diags: &[Diag]) -> (Outcome, Vec<String>, Vec<String>) {
    let security_diags: Vec<&Diag> = diags.iter().filter(|d| is_security(d)).collect();

    // Check forbidden rules.
    if case.is_vulnerable {
        for d in &security_diags {
            for forbidden in &case.forbidden_rule_ids {
                if rule_matches(&d.id, forbidden) {
                    // Wrong-reason detection counts as FP.
                    let matched = vec![];
                    let unexpected = security_diags.iter().map(|d| d.id.clone()).collect();
                    return (Outcome::FP, matched, unexpected);
                }
            }
        }
    }

    if !case.is_vulnerable {
        if security_diags.is_empty() {
            return (Outcome::TN, vec![], vec![]);
        } else {
            let unexpected = security_diags.iter().map(|d| d.id.clone()).collect();
            return (Outcome::FP, vec![], unexpected);
        }
    }

    // Positive case: check expected + alternative rule matches.
    let all_acceptable: Vec<&str> = case
        .expected_rule_ids
        .iter()
        .chain(case.allowed_alternative_rule_ids.iter())
        .map(|s| s.as_str())
        .collect();

    let mut matched = Vec::new();
    let mut unexpected = Vec::new();

    for d in &security_diags {
        let is_expected = all_acceptable.iter().any(|exp| rule_matches(&d.id, exp));
        if is_expected {
            matched.push(d.id.clone());
        } else {
            unexpected.push(d.id.clone());
        }
    }

    if matched.is_empty() {
        (Outcome::FN, matched, unexpected)
    } else {
        (Outcome::TP, matched, unexpected)
    }
}

fn score_location_level(
    case: &Case,
    diags: &[Diag],
    matched_rule_ids: &[String],
) -> Option<Outcome> {
    let expected_sinks = case.expected_sink_lines.as_ref()?;
    if expected_sinks.is_empty() {
        return None;
    }

    if !case.is_vulnerable {
        let has_security = diags.iter().any(is_security);
        return Some(if has_security {
            Outcome::FP
        } else {
            Outcome::TN
        });
    }

    // Check if any matched finding has a line within tolerance of expected sinks.
    let all_acceptable: Vec<&str> = case
        .expected_rule_ids
        .iter()
        .chain(case.allowed_alternative_rule_ids.iter())
        .map(|s| s.as_str())
        .collect();

    let security_diags: Vec<&Diag> = diags.iter().filter(|d| is_security(d)).collect();

    if matched_rule_ids.is_empty() {
        return Some(Outcome::FN);
    }

    for d in &security_diags {
        let is_expected = all_acceptable.iter().any(|exp| rule_matches(&d.id, exp));
        if !is_expected {
            continue;
        }
        let primary_ok = expected_sinks.iter().any(|r| {
            let lo = r[0].saturating_sub(2);
            let hi = r[1] + 2;
            d.line >= lo && d.line <= hi
        });
        if !primary_ok {
            continue;
        }
        // Optional: require at least one flow_step to fall within the
        // caller's call-site range. Only active when the fixture opts in.
        if let Some(call_ranges) = case.expected_call_site_lines.as_ref()
            && !call_ranges.is_empty()
        {
            let steps = d
                .evidence
                .as_ref()
                .map(|e| e.flow_steps.as_slice())
                .unwrap_or(&[]);
            let call_ok = steps.iter().any(|s| {
                call_ranges.iter().any(|r| {
                    let lo = r[0].saturating_sub(2);
                    let hi = r[1] + 2;
                    (s.line as usize) >= lo && (s.line as usize) <= hi
                })
            });
            if !call_ok {
                continue;
            }
        }
        return Some(Outcome::TP);
    }

    // Rule matched but location didn't.
    Some(Outcome::FN)
}

// ── Filtering ────────────────────────────────────────────────────────

fn should_run(case: &Case) -> bool {
    if case.disabled {
        return false;
    }

    if let Ok(lang) = std::env::var("NYX_BENCH_LANG")
        && case.language != lang
    {
        return false;
    }
    if let Ok(class) = std::env::var("NYX_BENCH_CLASS")
        && case.vuln_class != class
    {
        return false;
    }
    if let Ok(id) = std::env::var("NYX_BENCH_CASE")
        && case.case_id != id
    {
        return false;
    }
    if std::env::var("NYX_BENCH_POSITIVE_ONLY").is_ok() && !case.is_vulnerable {
        return false;
    }
    if std::env::var("NYX_BENCH_NEGATIVE_ONLY").is_ok() && case.is_vulnerable {
        return false;
    }
    if let Ok(tag) = std::env::var("NYX_BENCH_TAG")
        && !case.tags.iter().any(|t| t == &tag)
    {
        return false;
    }

    true
}

// ── Aggregation ──────────────────────────────────────────────────────

fn aggregate(outcomes: &[CaseOutcome], level: &str) -> Metrics {
    let (mut tp, mut fp, mut fn_, mut tn) = (0, 0, 0, 0);
    for o in outcomes {
        let outcome = match level {
            "file" => &o.outcome_file_level,
            "rule" => &o.outcome_rule_level,
            _ => &o.outcome_file_level,
        };
        match outcome {
            Outcome::TP => tp += 1,
            Outcome::FP => fp += 1,
            Outcome::FN => fn_ += 1,
            Outcome::TN => tn += 1,
        }
    }
    Metrics::compute(tp, fp, fn_, tn)
}

fn aggregate_by_key(
    outcomes: &[CaseOutcome],
    key_fn: impl Fn(&CaseOutcome) -> &str,
) -> BTreeMap<String, Metrics> {
    let mut groups: BTreeMap<String, Vec<&CaseOutcome>> = BTreeMap::new();
    for o in outcomes {
        groups.entry(key_fn(o).to_string()).or_default().push(o);
    }
    groups
        .into_iter()
        .map(|(k, cases)| {
            let (mut tp, mut fp, mut fn_, mut tn) = (0, 0, 0, 0);
            for o in &cases {
                match &o.outcome_rule_level {
                    Outcome::TP => tp += 1,
                    Outcome::FP => fp += 1,
                    Outcome::FN => fn_ += 1,
                    Outcome::TN => tn += 1,
                }
            }
            (k, Metrics::compute(tp, fp, fn_, tn))
        })
        .collect()
}

// ── Printing ─────────────────────────────────────────────────────────

fn print_case_table(outcomes: &[CaseOutcome]) {
    println!(
        "\n{:<25} {:<40} {:<6} {:<6} {:<6} {:<4} {:<4}",
        "CASE_ID", "FILE", "FILE", "RULE", "LOC", "SEC", "OTH"
    );
    println!("{}", "-".repeat(100));
    for o in outcomes {
        let loc = match &o.outcome_location_level {
            Some(out) => format!("{}", out),
            None => "-".to_string(),
        };
        println!(
            "{:<25} {:<40} {:<6} {:<6} {:<6} {:<4} {:<4}",
            o.case_id,
            truncate(&o.file, 39),
            o.outcome_file_level,
            o.outcome_rule_level,
            loc,
            o.security_finding_count,
            o.non_security_finding_count,
        );
    }
}

fn print_metrics_table(label: &str, metrics: &Metrics) {
    println!(
        "  {:<20} TP={:<4} FP={:<4} FN={:<4} TN={:<4}  P={:.3} R={:.3} F1={:.3}",
        label,
        metrics.tp,
        metrics.fp,
        metrics.fn_,
        metrics.tn,
        metrics.precision,
        metrics.recall,
        metrics.f1
    );
}

fn print_map_table(title: &str, map: &BTreeMap<String, Metrics>) {
    println!("\n  {}:", title);
    for (k, m) in map {
        print_metrics_table(k, m);
    }
}

fn truncate(s: &str, max: usize) -> String {
    if s.len() <= max {
        s.to_string()
    } else {
        format!("...{}", &s[s.len() - max + 3..])
    }
}

// ── Main test ────────────────────────────────────────────────────────

#[test]
#[ignore]
fn benchmark_evaluation() {
    let benchmark_dir = PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("tests/benchmark");
    let corpus_root = benchmark_dir.join("corpus");
    let gt_path = benchmark_dir.join("ground_truth.json");
    let results_dir = benchmark_dir.join("results");

    // Load ground truth.
    let gt_bytes = std::fs::read(&gt_path).expect("read ground_truth.json");
    let gt: GroundTruth = serde_json::from_slice(&gt_bytes).expect("parse ground_truth.json");

    // Compute ground truth hash for provenance.
    let gt_hash = format!("sha256:{}", sha256_hex(&gt_bytes));

    // Filter cases.
    let mut cases_skipped = 0usize;
    let cases_to_run: Vec<&Case> = gt
        .cases
        .iter()
        .filter(|c| {
            if !should_run(c) {
                cases_skipped += 1;
                false
            } else {
                true
            }
        })
        .collect();

    println!("\n=== Nyx Benchmark Evaluation ===");
    println!(
        "Corpus: {} total, {} to run, {} skipped",
        gt.cases.len(),
        cases_to_run.len(),
        cases_skipped
    );

    // Run each case.
    let mut outcomes: Vec<CaseOutcome> = Vec::with_capacity(cases_to_run.len());

    for case in &cases_to_run {
        let diags = scan_corpus_file(&corpus_root, &case.file);

        let security_count = diags.iter().filter(|d| is_security(d)).count();
        let non_security_count = diags.len() - security_count;
        let all_finding_ids: Vec<String> = diags.iter().map(|d| d.id.clone()).collect();

        let file_outcome = score_file_level(case, &diags);
        let (rule_outcome, matched, unexpected) = score_rule_level(case, &diags);
        let loc_outcome = score_location_level(case, &diags, &matched);

        outcomes.push(CaseOutcome {
            case_id: case.case_id.clone(),
            file: case.file.clone(),
            language: case.language.clone(),
            vuln_class: case.vuln_class.clone(),
            is_vulnerable: case.is_vulnerable,
            outcome_file_level: file_outcome,
            outcome_rule_level: rule_outcome,
            outcome_location_level: loc_outcome,
            matched_rule_ids: matched,
            unexpected_rule_ids: unexpected,
            all_finding_ids,
            security_finding_count: security_count,
            non_security_finding_count: non_security_count,
            diags,
        });
    }

    // Sort outcomes by case_id for deterministic output.
    outcomes.sort_by(|a, b| a.case_id.cmp(&b.case_id));

    // Print per-case table.
    print_case_table(&outcomes);

    // Compute aggregates.
    let agg_file = aggregate(&outcomes, "file");
    let agg_rule = aggregate(&outcomes, "rule");
    let by_language = aggregate_by_key(&outcomes, |o| &o.language);
    let by_class = aggregate_by_key(&outcomes, |o| &o.vuln_class);

    // Score at each confidence threshold.
    let by_confidence = score_by_confidence(&outcomes, &cases_to_run);

    // Print summary.
    println!("\n=== Aggregate Metrics ===");
    print_metrics_table("File-level", &agg_file);
    print_metrics_table("Rule-level", &agg_rule);
    print_map_table("By language (rule-level)", &by_language);
    print_map_table("By vuln class (rule-level)", &by_class);
    print_map_table("By confidence threshold (rule-level)", &by_confidence);

    // Write results JSON.
    std::fs::create_dir_all(&results_dir).ok();
    let results = BenchmarkResults {
        benchmark_version: "1.0".to_string(),
        timestamp: chrono_now(),
        scanner_version: env!("CARGO_PKG_VERSION").to_string(),
        scanner_config: ScannerConfig {
            analysis_mode: "Full".to_string(),
            taint_enabled: true,
            ast_patterns_enabled: true,
            state_analysis_enabled: true,
            worker_threads: 1,
        },
        ground_truth_hash: gt_hash,
        corpus_size: gt.cases.len(),
        cases_run: outcomes.len(),
        cases_skipped,
        outcomes,
        aggregate_file_level: agg_file,
        aggregate_rule_level: agg_rule,
        by_language,
        by_vuln_class: by_class,
        by_confidence,
    };

    let results_path = results_dir.join("latest.json");
    let json = serde_json::to_string_pretty(&results).expect("serialize results");
    std::fs::write(&results_path, &json).expect("write results/latest.json");

    println!("\nResults written to: {}", results_path.display());
    println!("=== Benchmark complete ===\n");

    // ── Regression thresholds (current baseline minus ~5pp) ────────
    // Baseline (2026-04-24, post per-return-path PathFact landing):
    // Rule-level P=0.947 R=0.994 F1=0.970 on the 316-case corpus
    // (TP=179 FP=10 FN=1 TN=125).  Adds 8 cases from the path-sanitiser
    // FP cluster (rs-safe-014, rs-safe-016, CVE-2018-20997 +
    // CVE-2022-36113 + CVE-2024-24576 patched/vulnerable pairs).  Rust
    // language F1 stayed at 1.000.
    //
    // Floors sit ~5pp below that baseline: a single-case flip is ~0.3pp
    // on this corpus, so 5pp is generous enough to absorb honest
    // FP↔TN trades while still catching a real regression in a
    // vulnerability class.  When you land a durable, measurable
    // improvement, tighten these floors, do not relax them to paper
    // over a regression.
    let rule = &results.aggregate_rule_level;
    assert!(
        rule.precision >= 0.897,
        "Rule-level precision {:.3} fell below threshold 0.897 (baseline 0.947)",
        rule.precision,
    );
    assert!(
        rule.recall >= 0.944,
        "Rule-level recall {:.3} fell below threshold 0.944 (baseline 0.994)",
        rule.recall,
    );
    assert!(
        rule.f1 >= 0.920,
        "Rule-level F1 {:.3} fell below threshold 0.920 (baseline 0.970)",
        rule.f1,
    );

    // ── Per-class floors ────────────────────────────────────────────
    // DATA_EXFIL: 13 TP fixtures across 8 languages.  Baseline at the
    // 0.5.x → next-minor ship is P=1.000 R=1.000 F1=1.000 with 6 paired
    // safe fixtures (sensitivity-gate, sanitizer-wrap) holding FP=0 on
    // the data_exfil-class noise budget.  Floor at 0.85 absorbs a one-
    // case regression (~0.077 on 13 cases) while still catching a
    // structural break.  When you land a durable improvement, tighten
    // this floor; do not relax it to paper over a regression.
    if let Some(de) = results.by_vuln_class.get("data_exfil") {
        assert!(
            de.f1 >= 0.85,
            "data_exfil rule-level F1 {:.3} fell below threshold 0.85 (baseline 1.000)",
            de.f1,
        );
        assert!(
            de.recall >= 0.85,
            "data_exfil rule-level recall {:.3} fell below threshold 0.85 (baseline 1.000)",
            de.recall,
        );
        assert!(
            de.precision >= 0.85,
            "data_exfil rule-level precision {:.3} fell below threshold 0.85 (baseline 1.000)",
            de.precision,
        );
    } else {
        panic!("data_exfil class missing from by_vuln_class breakdown");
    }
}

// ── Confidence-threshold scoring ─────────────────────────────────────

fn score_by_confidence(outcomes: &[CaseOutcome], cases: &[&Case]) -> BTreeMap<String, Metrics> {
    let mut result = BTreeMap::new();
    for (label, min_conf) in [
        (">=Low", Confidence::Low),
        (">=Medium", Confidence::Medium),
        (">=High", Confidence::High),
    ] {
        let (mut tp, mut fp, mut fn_, mut tn) = (0, 0, 0, 0);
        for (outcome, case) in outcomes.iter().zip(cases.iter()) {
            let filtered: Vec<&Diag> = outcome
                .diags
                .iter()
                .filter(|d| d.confidence.is_some_and(|c| c >= min_conf))
                .collect();
            let (rule_outcome, _, _) = score_rule_level_with_diags(case, &filtered);
            match rule_outcome {
                Outcome::TP => tp += 1,
                Outcome::FP => fp += 1,
                Outcome::FN => fn_ += 1,
                Outcome::TN => tn += 1,
            }
        }
        result.insert(label.to_string(), Metrics::compute(tp, fp, fn_, tn));
    }
    result
}

fn score_rule_level_with_diags(
    case: &Case,
    security_diags: &[&Diag],
) -> (Outcome, Vec<String>, Vec<String>) {
    if case.is_vulnerable {
        for d in security_diags {
            if !is_security(d) {
                continue;
            }
            for forbidden in &case.forbidden_rule_ids {
                if rule_matches(&d.id, forbidden) {
                    let unexpected = security_diags
                        .iter()
                        .filter(|d| is_security(d))
                        .map(|d| d.id.clone())
                        .collect();
                    return (Outcome::FP, vec![], unexpected);
                }
            }
        }
    }

    let sec_diags: Vec<&&Diag> = security_diags.iter().filter(|d| is_security(d)).collect();

    if !case.is_vulnerable {
        if sec_diags.is_empty() {
            return (Outcome::TN, vec![], vec![]);
        } else {
            let unexpected = sec_diags.iter().map(|d| d.id.clone()).collect();
            return (Outcome::FP, vec![], unexpected);
        }
    }

    let all_acceptable: Vec<&str> = case
        .expected_rule_ids
        .iter()
        .chain(case.allowed_alternative_rule_ids.iter())
        .map(|s| s.as_str())
        .collect();

    let mut matched = Vec::new();
    let mut unexpected = Vec::new();
    for d in &sec_diags {
        let is_expected = all_acceptable.iter().any(|exp| rule_matches(&d.id, exp));
        if is_expected {
            matched.push(d.id.clone());
        } else {
            unexpected.push(d.id.clone());
        }
    }

    if matched.is_empty() {
        (Outcome::FN, matched, unexpected)
    } else {
        (Outcome::TP, matched, unexpected)
    }
}

// ── Utilities ────────────────────────────────────────────────────────

fn sha256_hex(data: &[u8]) -> String {
    use std::io::Write;
    // Simple SHA-256 via command, avoids adding a crypto dependency.
    let mut child = std::process::Command::new("shasum")
        .args(["-a", "256"])
        .stdin(std::process::Stdio::piped())
        .stdout(std::process::Stdio::piped())
        .spawn()
        .expect("shasum");
    child.stdin.as_mut().unwrap().write_all(data).unwrap();
    let out = child.wait_with_output().unwrap();
    let s = String::from_utf8(out.stdout).unwrap();
    s.split_whitespace().next().unwrap_or("unknown").to_string()
}

fn chrono_now() -> String {
    // ISO 8601 timestamp without chrono dependency.
    let out = std::process::Command::new("date")
        .args(["-u", "+%Y-%m-%dT%H:%M:%SZ"])
        .output()
        .expect("date");
    String::from_utf8(out.stdout).unwrap().trim().to_string()
}