lean_ctx/core/

benchmark.rs

use std::collections::HashMap;
use std::path::{Path, PathBuf};
use std::time::Instant;

use walkdir::WalkDir;

use crate::core::compressor;
use crate::core::deps;
use crate::core::entropy;
use crate::core::preservation;
use crate::core::signatures;
use crate::core::tokens::count_tokens;

const COST_PER_TOKEN: f64 = crate::core::stats::DEFAULT_INPUT_PRICE_PER_M / 1_000_000.0;
const MAX_FILE_SIZE: u64 = 100 * 1024;
const MAX_FILES: usize = 50;
fn cache_hit_tokens() -> usize {
    let stub = "F1=src/example.rs [unchanged, 500L, use cached context]";
    count_tokens(stub)
}

// ── Types ───────────────────────────────────────────────────

#[derive(Debug, Clone)]
pub struct ModeMeasurement {
    pub mode: String,
    pub tokens: usize,
    pub savings_pct: f64,
    pub latency_us: u64,
    pub preservation_score: f64,
}

#[derive(Debug, Clone)]
pub struct FileMeasurement {
    #[allow(dead_code)]
    pub path: String,
    pub ext: String,
    pub raw_tokens: usize,
    pub modes: Vec<ModeMeasurement>,
}

#[derive(Debug, Clone)]
pub struct LanguageStats {
    pub ext: String,
    pub count: usize,
    pub total_tokens: usize,
    pub best_mode: String,
    pub best_mode_tokens: usize,
    pub best_savings_pct: f64,
}

#[derive(Debug, Clone)]
pub struct ModeSummary {
    pub mode: String,
    pub total_compressed_tokens: usize,
    pub avg_savings_pct: f64,
    pub avg_latency_us: u64,
    pub avg_preservation: f64,
}

#[derive(Debug, Clone)]
pub struct SessionSimResult {
    pub raw_tokens: usize,
    pub lean_tokens: usize,
    pub lean_ccp_tokens: usize,
    pub raw_cost: f64,
    pub lean_cost: f64,
    pub ccp_cost: f64,
}

#[derive(Debug, Clone)]
pub struct ProjectBenchmark {
    pub root: String,
    pub files_scanned: usize,
    pub files_measured: usize,
    pub total_raw_tokens: usize,
    pub languages: Vec<LanguageStats>,
    pub mode_summaries: Vec<ModeSummary>,
    pub session_sim: SessionSimResult,
    #[allow(dead_code)]
    pub file_results: Vec<FileMeasurement>,
}

// ── Scanner ─────────────────────────────────────────────────

fn is_skipped_dir(name: &str) -> bool {
    matches!(
        name,
        "node_modules"
            | ".git"
            | "target"
            | "dist"
            | "build"
            | ".next"
            | ".nuxt"
            | "__pycache__"
            | ".cache"
            | "coverage"
            | "vendor"
            | ".svn"
            | ".hg"
    )
}

fn is_text_ext(ext: &str) -> bool {
    matches!(
        ext,
        "rs" | "ts"
            | "tsx"
            | "js"
            | "jsx"
            | "py"
            | "go"
            | "java"
            | "c"
            | "cpp"
            | "h"
            | "hpp"
            | "cs"
            | "kt"
            | "swift"
            | "rb"
            | "php"
            | "vue"
            | "svelte"
            | "html"
            | "css"
            | "scss"
            | "less"
            | "json"
            | "yaml"
            | "yml"
            | "toml"
            | "xml"
            | "md"
            | "txt"
            | "sh"
            | "bash"
            | "zsh"
            | "fish"
            | "sql"
            | "graphql"
            | "proto"
            | "ex"
            | "exs"
            | "zig"
            | "lua"
            | "r"
            | "R"
            | "dart"
            | "scala"
    )
}

fn scan_project(root: &str) -> Vec<PathBuf> {
    let mut files: Vec<(PathBuf, u64)> = Vec::new();

    for entry in WalkDir::new(root)
        .max_depth(8)
        .into_iter()
        .filter_entry(|e| {
            let name = e.file_name().to_string_lossy();
            if e.file_type().is_dir() {
                if e.depth() > 0 && name.starts_with('.') {
                    return false;
                }
                return !is_skipped_dir(&name);
            }
            true
        })
    {
        let Ok(entry) = entry else { continue };

        if entry.file_type().is_dir() {
            continue;
        }

        let path = entry.path().to_path_buf();
        let ext = path.extension().and_then(|e| e.to_str()).unwrap_or("");

        if !is_text_ext(ext) {
            continue;
        }

        let size = entry.metadata().map_or(0, |m| m.len());
        if size == 0 || size > MAX_FILE_SIZE {
            continue;
        }

        files.push((path, size));
    }

    files.sort_by_key(|x| std::cmp::Reverse(x.1));

    let mut selected = Vec::new();
    let mut ext_counts: HashMap<String, usize> = HashMap::new();

    for (path, _size) in &files {
        if selected.len() >= MAX_FILES {
            break;
        }
        let ext = path
            .extension()
            .and_then(|e| e.to_str())
            .unwrap_or("")
            .to_string();
        let count = ext_counts.entry(ext.clone()).or_insert(0);
        if *count < 10 {
            *count += 1;
            selected.push(path.clone());
        }
    }

    selected
}

// ── Measurement ─────────────────────────────────────────────

fn measure_mode(content: &str, ext: &str, mode: &str, raw_tokens: usize) -> ModeMeasurement {
    let start = Instant::now();

    let compressed = match mode {
        "map" => {
            let sigs = signatures::extract_signatures(content, ext);
            let dep_info = deps::extract_deps(content, ext);
            let mut parts = Vec::new();
            if !dep_info.imports.is_empty() {
                parts.push(format!("deps: {}", dep_info.imports.join(", ")));
            }
            let key_refs: Vec<&signatures::Signature> = sigs
                .iter()
                .filter(|s| s.is_exported || s.indent == 0)
                .collect();
            // Drop exports the API already lists so the benchmark measures the
            // same deduped map the renderers emit (mirrors ctx_read map, #361).
            let extra_exports = signatures::exports_not_in_signatures(&dep_info.exports, &key_refs);
            if !extra_exports.is_empty() {
                parts.push(format!("exports: {}", extra_exports.join(", ")));
            }
            if !key_refs.is_empty() {
                let key_sigs: Vec<String> = key_refs.iter().map(|s| s.to_compact()).collect();
                parts.push(key_sigs.join("\n"));
            }
            parts.join("\n")
        }
        "signatures" => {
            let sigs = signatures::extract_signatures(content, ext);
            sigs.iter()
                .map(super::signatures::Signature::to_compact)
                .collect::<Vec<_>>()
                .join("\n")
        }
        "aggressive" => compressor::aggressive_compress(content, Some(ext)),
        // Deterministic variant: benchmark numbers feed the scorecard's
        // reproducibility digest (#211) and must not depend on whether the
        // shared embedding engine happens to be loaded in this process.
        "entropy" => entropy::entropy_compress_deterministic(content).output,
        "cache_hit" => format!(
            "F1=src/file.{ext} [unchanged, {}L, use cached context]",
            content.lines().count()
        ),
        _ => content.to_string(),
    };

    let latency = start.elapsed();
    let tokens = count_tokens(&compressed);

    let savings_pct = if raw_tokens > 0 {
        (1.0 - tokens as f64 / raw_tokens as f64) * 100.0
    } else {
        0.0
    };

    let preservation_score = if mode == "cache_hit" {
        -1.0
    } else {
        preservation::measure(content, &compressed, ext).overall()
    };

    ModeMeasurement {
        mode: mode.to_string(),
        tokens,
        savings_pct,
        latency_us: latency.as_micros() as u64,
        preservation_score,
    }
}

fn measure_file(path: &Path, root: &str) -> Option<FileMeasurement> {
    let content = std::fs::read_to_string(path).ok()?;
    if content.is_empty() {
        return None;
    }

    let ext = path
        .extension()
        .and_then(|e| e.to_str())
        .unwrap_or("")
        .to_string();

    let raw_tokens = count_tokens(&content);
    if raw_tokens == 0 {
        return None;
    }

    let modes = ["map", "signatures", "aggressive", "entropy", "cache_hit"];
    let measurements: Vec<ModeMeasurement> = modes
        .iter()
        .map(|m| measure_mode(&content, &ext, m, raw_tokens))
        .collect();

    let display_path = path
        .strip_prefix(root)
        .unwrap_or(path)
        .to_string_lossy()
        .to_string();

    Some(FileMeasurement {
        path: display_path,
        ext,
        raw_tokens,
        modes: measurements,
    })
}

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

/// Structural modes (map, signatures) only produce meaningful output for
/// programming languages with parseable code structures. For data/markup
/// files, these modes return near-empty output that would be misleading
/// as "compression". Only `aggressive` and `entropy` are honest for these.
fn is_mode_applicable_for_ext(mode: &str, ext: &str, tokens: usize) -> bool {
    if tokens == 0 {
        return false;
    }
    let is_structural_mode = matches!(mode, "map" | "signatures");
    if !is_structural_mode {
        return true;
    }
    // Extensions with actual code structures (functions, classes, types)
    let code_exts = [
        "rs", "ts", "tsx", "js", "jsx", "py", "go", "java", "kt", "c", "cpp", "h", "hpp", "cs",
        "rb", "swift", "scala", "zig", "lua", "php", "dart", "ex", "exs", "elm", "hs", "ml",
        "svelte", "vue", "sh", "bash", "zsh",
    ];
    code_exts.contains(&ext)
}

fn aggregate_languages(files: &[FileMeasurement]) -> Vec<LanguageStats> {
    struct LangAccum {
        count: usize,
        total_tokens: usize,
        mode_tokens: HashMap<String, usize>,
    }

    let mut map: HashMap<String, LangAccum> = HashMap::new();
    for f in files {
        let entry = map.entry(f.ext.clone()).or_insert_with(|| LangAccum {
            count: 0,
            total_tokens: 0,
            mode_tokens: HashMap::new(),
        });
        entry.count += 1;
        entry.total_tokens += f.raw_tokens;
        for m in &f.modes {
            *entry.mode_tokens.entry(m.mode.clone()).or_insert(0) += m.tokens;
        }
    }

    let mut stats: Vec<LanguageStats> = map
        .into_iter()
        .map(|(ext, acc)| {
            let (best_mode, best_tokens) = acc
                .mode_tokens
                .iter()
                .filter(|(m, _)| m.as_str() != "cache_hit")
                .filter(|(m, t)| is_mode_applicable_for_ext(m, &ext, **t))
                .min_by_key(|(_, t)| **t)
                .map_or_else(
                    || ("full".to_string(), acc.total_tokens),
                    |(m, t)| (m.clone(), *t),
                );

            let savings = if acc.total_tokens > 0 {
                (1.0 - best_tokens as f64 / acc.total_tokens as f64) * 100.0
            } else {
                0.0
            };

            LanguageStats {
                ext,
                count: acc.count,
                total_tokens: acc.total_tokens,
                best_mode,
                best_mode_tokens: best_tokens,
                best_savings_pct: savings,
            }
        })
        .collect();
    stats.sort_by_key(|x| std::cmp::Reverse(x.total_tokens));
    stats
}

fn aggregate_modes(files: &[FileMeasurement]) -> Vec<ModeSummary> {
    let mode_names = ["map", "signatures", "aggressive", "entropy", "cache_hit"];
    let mut summaries = Vec::new();

    for mode_name in &mode_names {
        let mut total_tokens = 0usize;
        let mut total_savings = 0.0f64;
        let mut total_latency = 0u64;
        let mut total_preservation = 0.0f64;
        let mut preservation_count = 0usize;
        let mut count = 0usize;

        for f in files {
            if let Some(m) = f.modes.iter().find(|m| m.mode == *mode_name) {
                total_tokens += m.tokens;
                total_savings += m.savings_pct;
                total_latency += m.latency_us;
                if m.preservation_score >= 0.0 {
                    total_preservation += m.preservation_score;
                    preservation_count += 1;
                }
                count += 1;
            }
        }

        if count == 0 {
            continue;
        }

        summaries.push(ModeSummary {
            mode: mode_name.to_string(),
            total_compressed_tokens: total_tokens,
            avg_savings_pct: total_savings / count as f64,
            avg_latency_us: total_latency / count as u64,
            avg_preservation: if preservation_count > 0 {
                total_preservation / preservation_count as f64
            } else {
                -1.0
            },
        });
    }

    summaries
}

// ── Session Simulation ──────────────────────────────────────

/// Honest CCP resume size (GL #573): measure what `ctx_session load` actually
/// emits for the latest real session instead of asserting a constant. The
/// model value is the documented fallback for machines without session history.
fn measured_ccp_resume_tokens() -> usize {
    const RESUME_CCP_MODEL_TOKENS: usize = 400;
    crate::core::session::SessionState::load_latest().map_or(RESUME_CCP_MODEL_TOKENS, |s| {
        crate::core::tokens::count_tokens(&format!("Session loaded.\n{}", s.format_compact()))
    })
}

fn simulate_session(files: &[FileMeasurement]) -> SessionSimResult {
    if files.is_empty() {
        return SessionSimResult {
            raw_tokens: 0,
            lean_tokens: 0,
            lean_ccp_tokens: 0,
            raw_cost: 0.0,
            lean_cost: 0.0,
            ccp_cost: 0.0,
        };
    }

    let file_count = files.len().min(15);
    let selected = &files[..file_count];

    let first_read_raw: usize = selected.iter().map(|f| f.raw_tokens).sum();

    let first_read_lean: usize = selected
        .iter()
        .enumerate()
        .map(|(i, f)| {
            let mode = if i % 3 == 0 { "aggressive" } else { "map" };
            f.modes
                .iter()
                .find(|m| m.mode == mode)
                .map_or(f.raw_tokens, |m| m.tokens)
        })
        .sum();

    let cache_reread_count = 10usize.min(file_count);
    let cache_raw: usize = selected[..cache_reread_count]
        .iter()
        .map(|f| f.raw_tokens)
        .sum();
    let cache_lean: usize = cache_reread_count * cache_hit_tokens();

    let shell_count = 8usize;
    let shell_raw = shell_count * 500;
    let shell_lean = shell_count * 200;

    let resume_raw: usize = selected.iter().map(|f| f.raw_tokens).sum();
    let resume_lean: usize = selected
        .iter()
        .map(|f| {
            f.modes
                .iter()
                .find(|m| m.mode == "map")
                .map_or(f.raw_tokens, |m| m.tokens)
        })
        .sum();
    let resume_ccp = measured_ccp_resume_tokens();

    let raw_total = first_read_raw + cache_raw + shell_raw + resume_raw;
    let lean_total = first_read_lean + cache_lean + shell_lean + resume_lean;
    let ccp_total = first_read_lean + cache_lean + shell_lean + resume_ccp;

    SessionSimResult {
        raw_tokens: raw_total,
        lean_tokens: lean_total,
        lean_ccp_tokens: ccp_total,
        raw_cost: raw_total as f64 * COST_PER_TOKEN,
        lean_cost: lean_total as f64 * COST_PER_TOKEN,
        ccp_cost: ccp_total as f64 * COST_PER_TOKEN,
    }
}

// ── Public API ──────────────────────────────────────────────

pub fn run_project_benchmark(path: &str) -> ProjectBenchmark {
    let root = if path.is_empty() { "." } else { path };
    let scanned = scan_project(root);
    let files_scanned = scanned.len();

    let file_results: Vec<FileMeasurement> = scanned
        .iter()
        .filter_map(|p| measure_file(p, root))
        .collect();

    let total_raw_tokens: usize = file_results.iter().map(|f| f.raw_tokens).sum();
    let languages = aggregate_languages(&file_results);
    let mode_summaries = aggregate_modes(&file_results);
    let session_sim = simulate_session(&file_results);

    ProjectBenchmark {
        root: root.to_string(),
        files_scanned,
        files_measured: file_results.len(),
        total_raw_tokens,
        languages,
        mode_summaries,
        session_sim,
        file_results,
    }
}

// ── Report: Terminal ────────────────────────────────────────

pub fn format_terminal(b: &ProjectBenchmark) -> String {
    let mut out = Vec::new();
    let sep = "\u{2550}".repeat(66);

    out.push(sep.clone());
    out.push(format!("  lean-ctx Benchmark — {}", b.root));
    out.push(sep.clone());

    let lang_summary: Vec<String> = b
        .languages
        .iter()
        .take(5)
        .map(|l| format!("{} {}", l.count, l.ext))
        .collect();
    out.push(format!(
        "  Scanned: {} files ({})",
        b.files_measured,
        lang_summary.join(", ")
    ));
    out.push(format!(
        "  Total raw tokens: {}",
        format_num(b.total_raw_tokens)
    ));
    out.push(String::new());

    out.push("  Compression by Language:".to_string());
    out.push(format!(
        "  {:<10} {:>6} {:>10} {:>10} {:>10} {:>10}",
        "Lang", "Files", "Raw Tok", "Best Mode", "Compressed", "Savings"
    ));
    out.push(format!("  {}", "\u{2500}".repeat(62)));
    for l in &b.languages {
        out.push(format!(
            "  {:<10} {:>6} {:>10} {:>10} {:>10} {:>9.1}%",
            l.ext,
            l.count,
            format_num(l.total_tokens),
            l.best_mode,
            format_num(l.best_mode_tokens),
            l.best_savings_pct,
        ));
    }
    out.push(String::new());

    out.push("  Mode Performance:".to_string());
    out.push(format!(
        "  {:<14} {:>10} {:>10} {:>10} {:>10}",
        "Mode", "Tokens", "Savings", "Latency", "Quality"
    ));
    out.push(format!("  {}", "\u{2500}".repeat(58)));

    for m in &b.mode_summaries {
        let qual = if m.avg_preservation < 0.0 {
            "N/A".to_string()
        } else {
            format!("{:.1}%", m.avg_preservation * 100.0)
        };
        let latency = if m.avg_latency_us > 1000 {
            format!("{:.1}ms", m.avg_latency_us as f64 / 1000.0)
        } else {
            format!("{}μs", m.avg_latency_us)
        };
        out.push(format!(
            "  {:<14} {:>10} {:>9.1}% {:>10} {:>10}",
            m.mode,
            format_num(m.total_compressed_tokens),
            m.avg_savings_pct,
            latency,
            qual,
        ));
    }

    out.push(String::new());
    out.push("  Session Simulation (30-min coding):".to_string());
    out.push(format!(
        "  {:<24} {:>10} {:>10} {:>10}",
        "Approach", "Tokens", "Cost", "Savings"
    ));
    out.push(format!("  {}", "\u{2500}".repeat(58)));

    let s = &b.session_sim;
    out.push(format!(
        "  {:<24} {:>10} {:>10} {:>10}",
        "Raw (no compression)",
        format_num(s.raw_tokens),
        format!("${:.3}", s.raw_cost),
        "\u{2014}",
    ));

    let lean_pct = if s.raw_tokens > 0 {
        (1.0 - s.lean_tokens as f64 / s.raw_tokens as f64) * 100.0
    } else {
        0.0
    };
    out.push(format!(
        "  {:<24} {:>10} {:>10} {:>9.1}%",
        "lean-ctx (no CCP)",
        format_num(s.lean_tokens),
        format!("${:.3}", s.lean_cost),
        lean_pct,
    ));

    let ccp_pct = if s.raw_tokens > 0 {
        (1.0 - s.lean_ccp_tokens as f64 / s.raw_tokens as f64) * 100.0
    } else {
        0.0
    };
    out.push(format!(
        "  {:<24} {:>10} {:>10} {:>9.1}%",
        "lean-ctx + CCP",
        format_num(s.lean_ccp_tokens),
        format!("${:.3}", s.ccp_cost),
        ccp_pct,
    ));

    out.push(sep.clone());
    out.join("\n")
}

// ── Report: Markdown ────────────────────────────────────────

pub fn format_markdown(b: &ProjectBenchmark) -> String {
    let mut out = Vec::new();

    out.push("# lean-ctx Benchmark Report".to_string());
    out.push(String::new());
    out.push(format!("**Project:** `{}`", b.root));
    out.push(format!("**Files measured:** {}", b.files_measured));
    out.push(format!(
        "**Total raw tokens:** {}",
        format_num(b.total_raw_tokens)
    ));
    out.push(String::new());

    out.push("## Compression by Language".to_string());
    out.push(String::new());
    out.push("| Language | Files | Raw Tokens | Best Mode | Compressed | Savings |".to_string());
    out.push("|----------|------:|-----------:|-----------|----------:|--------:|".to_string());
    for l in &b.languages {
        out.push(format!(
            "| {} | {} | {} | {} | {} | {:.1}% |",
            l.ext,
            l.count,
            format_num(l.total_tokens),
            l.best_mode,
            format_num(l.best_mode_tokens),
            l.best_savings_pct,
        ));
    }
    out.push(String::new());

    out.push("## Mode Performance".to_string());
    out.push(String::new());
    out.push("| Mode | Tokens | Savings | Latency | Quality |".to_string());
    out.push("|------|-------:|--------:|--------:|--------:|".to_string());
    for m in &b.mode_summaries {
        let qual = if m.avg_preservation < 0.0 {
            "N/A".to_string()
        } else {
            format!("{:.1}%", m.avg_preservation * 100.0)
        };
        let latency = if m.avg_latency_us > 1000 {
            format!("{:.1}ms", m.avg_latency_us as f64 / 1000.0)
        } else {
            format!("{}μs", m.avg_latency_us)
        };
        out.push(format!(
            "| {} | {} | {:.1}% | {} | {} |",
            m.mode,
            format_num(m.total_compressed_tokens),
            m.avg_savings_pct,
            latency,
            qual
        ));
    }
    out.push(String::new());

    out.push("## Session Simulation (30-min coding)".to_string());
    out.push(String::new());
    out.push("| Approach | Tokens | Cost | Savings |".to_string());
    out.push("|----------|-------:|-----:|--------:|".to_string());

    let s = &b.session_sim;
    out.push(format!(
        "| Raw (no compression) | {} | ${:.3} | — |",
        format_num(s.raw_tokens),
        s.raw_cost
    ));

    let lean_pct = if s.raw_tokens > 0 {
        (1.0 - s.lean_tokens as f64 / s.raw_tokens as f64) * 100.0
    } else {
        0.0
    };
    out.push(format!(
        "| lean-ctx (no CCP) | {} | ${:.3} | {:.1}% |",
        format_num(s.lean_tokens),
        s.lean_cost,
        lean_pct
    ));

    let ccp_pct = if s.raw_tokens > 0 {
        (1.0 - s.lean_ccp_tokens as f64 / s.raw_tokens as f64) * 100.0
    } else {
        0.0
    };
    out.push(format!(
        "| lean-ctx + CCP | {} | ${:.3} | {:.1}% |",
        format_num(s.lean_ccp_tokens),
        s.ccp_cost,
        ccp_pct
    ));

    out.push(String::new());
    out.push(format!(
        "*Generated by lean-ctx benchmark v{} — https://leanctx.com*",
        env!("CARGO_PKG_VERSION")
    ));

    out.join("\n")
}

// ── Report: JSON ────────────────────────────────────────────

pub fn format_json(b: &ProjectBenchmark) -> String {
    let modes: Vec<serde_json::Value> = b.mode_summaries.iter().map(|m| {
        serde_json::json!({
            "mode": m.mode,
            "total_compressed_tokens": m.total_compressed_tokens,
            "avg_savings_pct": round2(m.avg_savings_pct),
            "avg_latency_us": m.avg_latency_us,
            "avg_preservation": if m.avg_preservation < 0.0 { serde_json::Value::Null } else { serde_json::json!(round2(m.avg_preservation * 100.0)) },
        })
    }).collect();

    let languages: Vec<serde_json::Value> = b
        .languages
        .iter()
        .map(|l| {
            serde_json::json!({
                "ext": l.ext,
                "count": l.count,
                "total_tokens": l.total_tokens,
                "best_mode": l.best_mode,
                "best_mode_tokens": l.best_mode_tokens,
                "best_savings_pct": round2(l.best_savings_pct),
            })
        })
        .collect();

    let file_details: Vec<serde_json::Value> = b
        .file_results
        .iter()
        .map(|f| {
            let file_modes: Vec<serde_json::Value> = f
                .modes
                .iter()
                .map(|m| {
                    serde_json::json!({
                        "mode": m.mode,
                        "tokens": m.tokens,
                        "savings_pct": round2(m.savings_pct),
                        "latency_us": m.latency_us,
                        "preservation": if m.preservation_score < 0.0 {
                            serde_json::Value::Null
                        } else {
                            serde_json::json!(round2(m.preservation_score * 100.0))
                        },
                    })
                })
                .collect();
            serde_json::json!({
                "path": f.path,
                "ext": f.ext,
                "raw_tokens": f.raw_tokens,
                "modes": file_modes,
            })
        })
        .collect();

    let s = &b.session_sim;
    let report = serde_json::json!({
        "version": env!("CARGO_PKG_VERSION"),
        "root": b.root,
        "files_scanned": b.files_scanned,
        "files_measured": b.files_measured,
        "total_raw_tokens": b.total_raw_tokens,
        "languages": languages,
        "mode_summaries": modes,
        "files": file_details,
        "session_simulation": {
            "raw_tokens": s.raw_tokens,
            "lean_tokens": s.lean_tokens,
            "lean_ccp_tokens": s.lean_ccp_tokens,
            "raw_cost_usd": round2(s.raw_cost),
            "lean_cost_usd": round2(s.lean_cost),
            "ccp_cost_usd": round2(s.ccp_cost),
        },
    });

    serde_json::to_string_pretty(&report).unwrap_or_else(|_| "{}".to_string())
}

// ── Helpers ─────────────────────────────────────────────────

fn format_num(n: usize) -> String {
    if n >= 1_000_000 {
        format!("{:.1}M", n as f64 / 1_000_000.0)
    } else if n >= 1_000 {
        format!("{:.1}K", n as f64 / 1_000.0)
    } else {
        format!("{n}")
    }
}

fn round2(v: f64) -> f64 {
    (v * 100.0).round() / 100.0
}

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

    fn mock_file(path: &str, ext: &str, raw: usize, modes: Vec<(&str, usize)>) -> FileMeasurement {
        FileMeasurement {
            path: path.to_string(),
            ext: ext.to_string(),
            raw_tokens: raw,
            modes: modes
                .into_iter()
                .map(|(mode, tokens)| ModeMeasurement {
                    mode: mode.to_string(),
                    tokens,
                    savings_pct: if raw > 0 {
                        (1.0 - tokens as f64 / raw as f64) * 100.0
                    } else {
                        0.0
                    },
                    latency_us: 100,
                    preservation_score: 0.85,
                })
                .collect(),
        }
    }

    #[test]
    fn aggregate_languages_computes_best_mode() {
        let files = vec![
            mock_file(
                "a.rs",
                "rs",
                1000,
                vec![("map", 400), ("signatures", 200), ("aggressive", 300)],
            ),
            mock_file(
                "b.rs",
                "rs",
                800,
                vec![("map", 300), ("signatures", 150), ("aggressive", 250)],
            ),
            mock_file(
                "c.py",
                "py",
                600,
                vec![("map", 100), ("signatures", 250), ("aggressive", 200)],
            ),
        ];

        let langs = aggregate_languages(&files);
        assert_eq!(langs.len(), 2);

        let rs = langs.iter().find(|l| l.ext == "rs").unwrap();
        assert_eq!(rs.count, 2);
        assert_eq!(rs.total_tokens, 1800);
        assert_eq!(rs.best_mode, "signatures");
        assert_eq!(rs.best_mode_tokens, 350);
        assert!(rs.best_savings_pct > 80.0);

        let py = langs.iter().find(|l| l.ext == "py").unwrap();
        assert_eq!(py.best_mode, "map");
        assert_eq!(py.best_mode_tokens, 100);
    }

    #[test]
    fn aggregate_modes_averages() {
        let files = vec![
            mock_file("a.rs", "rs", 1000, vec![("map", 400), ("aggressive", 300)]),
            mock_file("b.rs", "rs", 500, vec![("map", 200), ("aggressive", 100)]),
        ];

        let modes = aggregate_modes(&files);
        let map = modes.iter().find(|m| m.mode == "map").unwrap();
        assert_eq!(map.total_compressed_tokens, 600);
        assert!(map.avg_savings_pct > 50.0);
    }

    #[test]
    fn session_sim_empty_files() {
        let result = simulate_session(&[]);
        assert_eq!(result.raw_tokens, 0);
        assert_eq!(result.lean_tokens, 0);
        assert!((result.raw_cost).abs() < f64::EPSILON);
    }

    #[test]
    fn session_sim_basic() {
        let files: Vec<FileMeasurement> = (0..5)
            .map(|i| {
                mock_file(
                    &format!("file_{i}.rs"),
                    "rs",
                    2000,
                    vec![
                        ("map", 800),
                        ("aggressive", 600),
                        ("cache_hit", cache_hit_tokens()),
                    ],
                )
            })
            .collect();
        let result = simulate_session(&files);
        assert!(result.raw_tokens > 0);
        assert!(result.lean_tokens < result.raw_tokens);
        assert!(
            result.lean_ccp_tokens < result.lean_tokens,
            "CCP resume ({}) should beat map-based resume ({}) with enough files",
            result.lean_ccp_tokens,
            result.lean_tokens
        );
    }

    #[test]
    fn format_json_includes_files_and_language_savings() {
        let files = vec![mock_file(
            "src/main.rs",
            "rs",
            500,
            vec![("map", 200), ("signatures", 100), ("cache_hit", 13)],
        )];
        let bench = ProjectBenchmark {
            root: ".".to_string(),
            files_scanned: 1,
            files_measured: 1,
            total_raw_tokens: 500,
            languages: aggregate_languages(&files),
            mode_summaries: aggregate_modes(&files),
            session_sim: simulate_session(&files),
            file_results: files,
        };

        let json_str = format_json(&bench);
        let parsed: serde_json::Value = serde_json::from_str(&json_str).unwrap();

        assert!(parsed["files"].is_array());
        assert_eq!(parsed["files"].as_array().unwrap().len(), 1);
        assert_eq!(parsed["files"][0]["path"], "src/main.rs");
        assert!(parsed["files"][0]["modes"].is_array());

        assert!(parsed["languages"][0]["best_mode"].is_string());
        assert!(parsed["languages"][0]["best_savings_pct"].is_number());
    }

    #[test]
    fn format_markdown_contains_language_savings() {
        let files = vec![mock_file(
            "lib.rs",
            "rs",
            1000,
            vec![("map", 300), ("signatures", 200)],
        )];
        let bench = ProjectBenchmark {
            root: ".".to_string(),
            files_scanned: 1,
            files_measured: 1,
            total_raw_tokens: 1000,
            languages: aggregate_languages(&files),
            mode_summaries: aggregate_modes(&files),
            session_sim: simulate_session(&files),
            file_results: files,
        };

        let md = format_markdown(&bench);
        assert!(md.contains("Compression by Language"));
        assert!(md.contains("Best Mode"));
        assert!(md.contains("Savings"));
    }

    #[test]
    fn format_terminal_contains_language_section() {
        let files = vec![mock_file(
            "app.py",
            "py",
            800,
            vec![("map", 200), ("aggressive", 300)],
        )];
        let bench = ProjectBenchmark {
            root: ".".to_string(),
            files_scanned: 1,
            files_measured: 1,
            total_raw_tokens: 800,
            languages: aggregate_languages(&files),
            mode_summaries: aggregate_modes(&files),
            session_sim: simulate_session(&files),
            file_results: files,
        };

        let out = format_terminal(&bench);
        assert!(out.contains("Compression by Language"));
        assert!(out.contains("py"));
        assert!(out.contains("Best Mode"));
    }

    #[test]
    fn run_project_benchmark_on_current_crate() {
        let bench = run_project_benchmark("src");
        assert!(bench.files_measured > 0);
        assert!(bench.total_raw_tokens > 0);
        assert!(!bench.languages.is_empty());
        assert!(!bench.mode_summaries.is_empty());

        for lang in &bench.languages {
            assert!(!lang.best_mode.is_empty());
            assert!(lang.best_savings_pct >= 0.0);
        }

        let json = format_json(&bench);
        let parsed: serde_json::Value = serde_json::from_str(&json).unwrap();
        assert!(!parsed["files"].as_array().unwrap().is_empty());

        let md = format_markdown(&bench);
        assert!(md.contains("lean-ctx Benchmark Report"));

        let term = format_terminal(&bench);
        assert!(term.contains("Session Simulation"));
    }
}