using AyCode.Core.Serializers.Binaries;
using AyCode.Core.Tests.TestModels;
using System.Globalization;
using System.Runtime.CompilerServices;
using System.Text;

namespace AyCode.Core.Serializers.Console;

/// <summary>
/// Output formatters for the benchmark suite: per-row console table, .log file, .LLM markdown file.
/// Consumes <see cref="BenchmarkResult"/> rows produced by the benchmark loop and emits human-readable
/// + LLM-friendly outputs. Display-only helpers (per-op µs conversion, allocation KB formatting,
/// range/CV formatting, paired-aggregation stats) live here too — they are display-side concerns.
/// </summary>
internal static class Output
{
    /// <summary>
    /// Per-cell-paired aggregation of an overall comparison. Captures three different aggregation
    /// strategies so the reader can judge whether the headline delta is dominated by one large cell
    /// (arithmetic mean) or representative of typical workload (geometric mean / median).
    /// </summary>
    /// <param name="ArithMeanPct">Arithmetic mean of µs/op — magnitude-weighted; biased toward Large cell.</param>
    /// <param name="GeoMeanPct">Geometric mean of per-cell ratios — magnitude-neutral; each cell weighted equally.</param>
    /// <param name="MedianPct">Median of per-cell ratios — outlier-resistant.</param>
    /// <param name="AcAvg">Arithmetic mean AcBinary value (µs/op or bytes).</param>
    /// <param name="MpAvg">Arithmetic mean MemPack value.</param>
    /// <param name="CellCount">Number of paired cells contributing to the geo/median.</param>
    internal record OverallStats(double ArithMeanPct, double GeoMeanPct, double MedianPct, double AcAvg, double MpAvg, int CellCount);

    [MethodImpl(MethodImplOptions.AggressiveInlining)]
    internal static double ToPerOpMicros(double totalMs, int iterations) => iterations > 0 ? totalMs / iterations * 1000.0 : 0;

    // Per-row per-op µs accessors — pull batch-time + iter from BenchmarkResult and convert. Used wherever
    // averaging or comparison happens across rows with potentially different iter counts (Winners summary,
    // Overall comparison, per-cell summary row). Keeping these as methods rather than properties on
    // BenchmarkResult preserves the result-as-data-bag distinction.
    internal static double SerPerOp(BenchmarkResult r) => ToPerOpMicros(r.SerializeTimeMs, r.SerializeIterations);
    internal static double DesPerOp(BenchmarkResult r) => ToPerOpMicros(r.DeserializeTimeMs, r.DeserializeIterations);
    internal static double RtPerOp(BenchmarkResult r) => ToPerOpMicros(r.RoundTripTimeMs, r.RoundTripIterations);

    /// <summary>
    /// Converts a byte count to KB (1 KB = 1024 B). Display-only helper so allocation columns can
    /// render compact F2 KB values (e.g. <c>4.05 KB</c> instead of <c>4,144 B</c>) — header carries
    /// the unit so per-row entries stay numbers-only. CSV / raw-data outputs keep the precise byte
    /// integers untouched.
    /// </summary>
    [MethodImpl(MethodImplOptions.AggressiveInlining)]
    internal static double ToKilobytes(long bytes) => bytes / 1024.0;

    /// <summary>
    /// Computes arithmetic + geometric + median aggregation of an AcBinary-vs-MemPack comparison
    /// across paired cells (joined by <c>TestDataName</c>). Per-cell pairing is required for the
    /// geo/median variants — a cell where AcBinary or MemPack is missing is dropped from all stats.
    /// Returns null when no paired cell has a valid value.
    /// </summary>
    internal static OverallStats? ComputeOverallStats(List<BenchmarkResult> acResults, List<BenchmarkResult> mpResults, Func<BenchmarkResult, double> getValue)
    {
        if (acResults.Count == 0 || mpResults.Count == 0) return null;

        var pairs = (from ac in acResults
                     join mp in mpResults on ac.TestDataName equals mp.TestDataName
                     let acV = getValue(ac)
                     let mpV = getValue(mp)
                     where acV > 0 && mpV > 0
                     select (ac: acV, mp: mpV)).ToList();

        if (pairs.Count == 0) return null;

        var acAvg = pairs.Average(p => p.ac);
        var mpAvg = pairs.Average(p => p.mp);
        var ratios = pairs.Select(p => p.ac / p.mp).ToList();

        // Geometric mean: exp(avg(ln(ratios))) — numerically stable vs Π ratios then ^(1/N).
        var geoMean = Math.Exp(ratios.Sum(Math.Log) / ratios.Count);

        // Median (paired-ratio): for even N use the midpoint of the two middle values.
        var sorted = ratios.OrderBy(r => r).ToList();
        var median = sorted.Count % 2 == 1
            ? sorted[sorted.Count / 2]
            : (sorted[sorted.Count / 2 - 1] + sorted[sorted.Count / 2]) / 2.0;

        return new OverallStats(
            ArithMeanPct: (acAvg / mpAvg - 1) * 100,
            GeoMeanPct: (geoMean - 1) * 100,
            MedianPct: (median - 1) * 100,
            AcAvg: acAvg,
            MpAvg: mpAvg,
            CellCount: ratios.Count);
    }

    /// <summary>
    /// Formats a per-op micros value with its inter-sample range and CV-threshold marker as
    /// <c>"26.86 (24.5..29.1)"</c> or <c>"26.86 (24.5..29.1) ⚠️5.2%"</c>. Median first, range in parentheses,
    /// CV warning suffix only when CV > <see cref="Configuration.UnstableCVThreshold"/>. When min == max == median
    /// (single-sample / Debug / quick mode), collapses to bare median to avoid visual clutter.
    /// All time inputs are total-batch milliseconds; <paramref name="iterations"/> is the per-row iter
    /// count (post-adaptive-calibration).
    /// </summary>
    internal static string FormatMicrosWithRange(double medianMs, double minMs, double maxMs, double stdDevMs, int iterations, CultureInfo inv)
    {
        var med = ToPerOpMicros(medianMs, iterations);
        // No range data (single-sample fast path) — surface as bare median, identical to the prior format.
        if (minMs <= 0 && maxMs <= 0) return med.ToString("F2", inv);
        if (minMs >= medianMs && maxMs <= medianMs) return med.ToString("F2", inv);

        var min = ToPerOpMicros(minMs, iterations);
        var max = ToPerOpMicros(maxMs, iterations);
        var range = $"{med.ToString("F2", inv)} ({min.ToString("F2", inv)}..{max.ToString("F2", inv)})";

        // CV (coefficient of variation = stddev / mean) — flag rows above the unstable threshold so a
        // small inter-engine delta on a high-CV row is easy to discount as noise.
        if (medianMs > 0 && stdDevMs > 0)
        {
            var cv = stdDevMs / medianMs;
            if (cv > Configuration.UnstableCVThreshold)
            {
                var cvPct = (cv * 100).ToString("F1", inv);
                return $"{range} ⚠️{cvPct}%";
            }
        }

        return range;
    }

    /// <summary>
    /// Formats a signed percent delta with explicit sign for positive values (`+1.5%`, `-3.0%`, `0.0%`).
    /// Padded to 7 chars (e.g. ` +12.3%`, `-100.0%`) for column alignment in the Overall block.
    /// </summary>
    internal static string FormatPctSigned(double pct) => pct.ToString("+0.0;-0.0;0.0", CultureInfo.InvariantCulture).PadLeft(6) + "%";

    /// <summary>
    /// Renders one Overall row with arith / geo / median deltas + AcBinary/MemPack absolute means.
    /// Color is driven by the geometric-mean delta (magnitude-neutral signal). Skips silently when
    /// stats is null (no paired data).
    /// </summary>
    internal static void WriteOverallLine(string label, string unit, OverallStats? stats, string fmt = "F2")
    {
        if (stats == null) return;

        // Color follows geo-mean (the magnitude-neutral signal). The arith-mean column may show a
        // different sign when a single big cell dominates — that's exactly the signal we want to surface.
        System.Console.ForegroundColor = stats.GeoMeanPct <= 0 ? ConsoleColor.Green : ConsoleColor.Red;
        System.Console.WriteLine($"  {label,-12} arith {FormatPctSigned(stats.ArithMeanPct)} │ geo {FormatPctSigned(stats.GeoMeanPct)} │ median {FormatPctSigned(stats.MedianPct)}   ({stats.AcAvg.ToString(fmt, CultureInfo.InvariantCulture)} {unit} vs {stats.MpAvg.ToString(fmt, CultureInfo.InvariantCulture)} {unit}, {stats.CellCount} cells)");
        System.Console.ResetColor();
    }

    /// <summary>
    /// Same as <see cref="WriteOverallLine"/> but appends to a <see cref="StringBuilder"/> (no color).
    /// Used by the .log and .LLM file writers.
    /// </summary>
    internal static void AppendOverallLine(StringBuilder sb, string label, string unit, OverallStats? stats, string fmt = "F2")
    {
        if (stats == null) return;
        sb.AppendLine($"  {label,-12} arith {FormatPctSigned(stats.ArithMeanPct)} | geo {FormatPctSigned(stats.GeoMeanPct)} | median {FormatPctSigned(stats.MedianPct)}   ({stats.AcAvg.ToString(fmt, CultureInfo.InvariantCulture)} {unit} vs {stats.MpAvg.ToString(fmt, CultureInfo.InvariantCulture)} {unit}, {stats.CellCount} cells)");
    }

    internal static void PrintResult(BenchmarkResult result)
    {
        // Numbers-only per-row entries; the column-headers carry units (µs/op, KB/op).
        var ser = result.SerializeTimeMs > 0 ? $"{SerPerOp(result),7:F2}" : "    N/A";
        var des = result.DeserializeTimeMs > 0 ? $"{DesPerOp(result),7:F2}" : "    N/A";
        var serAlloc = result.SerializeTimeMs > 0 ? $"{ToKilobytes(result.SerializeAllocBytesPerOp),7:F2}" : "    N/A";
        var desAlloc = result.DeserializeTimeMs > 0 ? $"{ToKilobytes(result.DeserializeAllocBytesPerOp),7:F2}" : "    N/A";
        System.Console.WriteLine($"  {result.SerializerName,-40} | Size: {result.SerializedSize,8:N0} B | Ser: {ser} µs/op ({serAlloc} KB/op) | Des: {des} µs/op ({desAlloc} KB/op)");
    }

    internal static void PrintGroupedResults(List<BenchmarkResult> results, List<TestDataSet> testDataSets)
    {
        System.Console.WriteLine("\n");
        System.Console.WriteLine("╔══════════════════════════════════════════════════════════════════════════════════════════════════════╗");
        System.Console.WriteLine("║                                    GROUPED RESULTS BY TEST DATA                                      ║");
        System.Console.WriteLine("╚══════════════════════════════════════════════════════════════════════════════════════════════════════╝");

        // Print serializer options
        var optionsMap = results
            .Where(r => r.OptionsDescription != null)
            .Select(r => (r.SerializerName, r.OptionsDescription!))
            .Distinct()
            .ToList();

        if (optionsMap.Count > 0)
        {
            System.Console.WriteLine();
            System.Console.WriteLine("  Serializer Options:");
            foreach (var (name, opts) in optionsMap)
                System.Console.WriteLine($"    {name}: {opts}");
        }

        foreach (var testData in testDataSets)
        {
            // Order by per-op µs (iter-independent) — rows may have different iter counts post-calibration.
            var testResults = results.Where(r => r.TestDataName == testData.DisplayName).OrderBy(r => RtPerOp(r)).ToList();
            // Baseline switched MessagePack → MemoryPack: MemoryPack is the SOTA performance leader.
            var memPackResult = testResults.FirstOrDefault(r => (r.Engine == Configuration.EngineMemoryPack && r.IoMode == Configuration.IoByteArray));
            // Pin the comparison to AcBinary's SGen variant — apples-to-apples vs MemoryPack (also source-generated).
            // The Runtime variant is shown alongside in the table for context, not used as the headline number.
            var acBinaryResult = testResults.FirstOrDefault(r => (r.Engine == Configuration.EngineAcBinary && r.IoMode == Configuration.IoByteArray && r.DispatchMode == Configuration.ModeSGen));

            System.Console.WriteLine($"\n┌─ {testData.DisplayName} ─".PadRight(172, '─') + "┐");
            // Header-only units; per-row entries are numbers (µs/op for time, KB/op for alloc, KB pair "ser / des" for Setup, B for Size).
            System.Console.WriteLine($"│ {"#",-4} │ {"Engine",-11} │ {"Options",-22} │ {"IO",-12} │ {"Mode",-8} │ {"Setup S/D KB",-14} │ {"Size B",-8} │ {"Ser µs/op",-10} │ {"SerAlc KB",-10} │ {"Des µs/op",-10} │ {"DesAlc KB",-10} │ {"RT µs/op",-10} │ {"RTAlc KB",-10} │");
            System.Console.WriteLine($"├{"─".PadRight(6, '─')}┼{"─".PadRight(13, '─')}┼{"─".PadRight(24, '─')}┼{"─".PadRight(14, '─')}┼{"─".PadRight(10, '─')}┼{"─".PadRight(16, '─')}┼{"─".PadRight(10, '─')}┼{"─".PadRight(12, '─')}┼{"─".PadRight(12, '─')}┼{"─".PadRight(12, '─')}┼{"─".PadRight(12, '─')}┼{"─".PadRight(12, '─')}┼{"─".PadRight(12, '─')}┤");

            var rank = 1;
            foreach (var result in testResults)
            {
                var size = $"{result.SerializedSize:N0}";
                var setup = $"{ToKilobytes(result.SetupSerializeAllocBytes):F2} / {ToKilobytes(result.SetupDeserializeAllocBytes):F2}";
                var ser = result.SerializeTimeMs > 0 ? $"{SerPerOp(result):F2}" : "N/A";
                var des = result.DeserializeTimeMs > 0 ? $"{DesPerOp(result):F2}" : "N/A";
                var rt = result.RoundTripTimeMs > 0 ? $"{RtPerOp(result):F2}" : "N/A";
                var serAlloc = result.SerializeTimeMs > 0 ? $"{ToKilobytes(result.SerializeAllocBytesPerOp):F2}" : "N/A";
                var desAlloc = result.DeserializeTimeMs > 0 ? $"{ToKilobytes(result.DeserializeAllocBytesPerOp):F2}" : "N/A";
                var rtAlloc = result.RoundTripAllocBytesPerOp > 0 ? $"{ToKilobytes(result.RoundTripAllocBytesPerOp):F2}" : "N/A";

                // Highlight MemoryPack baseline (any Byte[]) and AcBinary headline contender (Byte[] + SGen) with win/lose colors.
                // The AcBinary Byte[]+Runtime variant is shown unhighlighted — it's contextual (SGen speed-up reference), not the headline.
                var isHighlighted = (result.Engine == Configuration.EngineMemoryPack && result.IoMode == Configuration.IoByteArray)
                    || (result.Engine == Configuration.EngineAcBinary && result.IoMode == Configuration.IoByteArray && result.DispatchMode == Configuration.ModeSGen);

                var prefix = isHighlighted ? "│►" : "│ ";
                var suffix = isHighlighted ? "◄│" : " │";

                // Color logic: Green = winner (faster), Red = loser (slower)
                if (isHighlighted && memPackResult != null && acBinaryResult != null)
                {
                    var isMemPack = (result.Engine == Configuration.EngineMemoryPack && result.IoMode == Configuration.IoByteArray);
                    var memPackFaster = RtPerOp(memPackResult) < RtPerOp(acBinaryResult);

                    if (isMemPack)
                    {
                        System.Console.ForegroundColor = memPackFaster ? ConsoleColor.Green : ConsoleColor.Red;
                    }
                    else
                    {
                        System.Console.ForegroundColor = memPackFaster ? ConsoleColor.Red : ConsoleColor.Green;
                    }
                }

                System.Console.WriteLine($"{prefix}{rank++,4} │ {result.Engine,-11} │ {result.OptionsPreset,-22} │ {result.IoMode,-12} │ {result.DispatchMode,-8} │ {setup,14} │ {size,8} │ {ser,10} │ {serAlloc,10} │ {des,10} │ {desAlloc,10} │ {rt,10} │ {rtAlloc,10}{suffix}");

                if (isHighlighted)
                {
                    System.Console.ResetColor();
                }
            }

            // Footer row: AcBinary (Byte[]) vs MemoryPack (Byte[]) comparison per column
            if (memPackResult != null && acBinaryResult != null)
            {
                var sizePct = (acBinaryResult.SerializedSize / (double)memPackResult.SerializedSize - 1) * 100;
                // Per-op µs ratio (iter-independent) — Ser/Des may have different iter counts on the two rows.
                var serPct = SerPerOp(memPackResult) > 0 ? (SerPerOp(acBinaryResult) / SerPerOp(memPackResult) - 1) * 100 : 0;
                var desPct = DesPerOp(memPackResult) > 0 ? (DesPerOp(acBinaryResult) / DesPerOp(memPackResult) - 1) * 100 : 0;
                var rtPct = RtPerOp(memPackResult) > 0 ? (RtPerOp(acBinaryResult) / RtPerOp(memPackResult) - 1) * 100 : 0;

                var serAllocPct = memPackResult.SerializeAllocBytesPerOp > 0 ? (acBinaryResult.SerializeAllocBytesPerOp / (double)memPackResult.SerializeAllocBytesPerOp - 1) * 100 : 0;
                var desAllocPct = memPackResult.DeserializeAllocBytesPerOp > 0 ? (acBinaryResult.DeserializeAllocBytesPerOp / (double)memPackResult.DeserializeAllocBytesPerOp - 1) * 100 : 0;
                var rtAllocPct = memPackResult.RoundTripAllocBytesPerOp > 0 ? (acBinaryResult.RoundTripAllocBytesPerOp / (double)memPackResult.RoundTripAllocBytesPerOp - 1) * 100 : 0;

                // Footer separator: merge first 5 cols (#, Engine, Options, IO, Mode) → comparison label;
                // remaining 8 cols stay aligned (Setup S/D KB, Size, Ser µs/op, SerAlc KB, Des µs/op, DesAlc KB, RT µs/op, RTAlc KB).
                System.Console.WriteLine($"├{"─".PadRight(6, '─')}┴{"─".PadRight(13, '─')}┴{"─".PadRight(24, '─')}┴{"─".PadRight(14, '─')}┴{"─".PadRight(10, '─')}┼{"─".PadRight(16, '─')}┼{"─".PadRight(10, '─')}┼{"─".PadRight(12, '─')}┼{"─".PadRight(12, '─')}┼{"─".PadRight(12, '─')}┼{"─".PadRight(12, '─')}┼{"─".PadRight(12, '─')}┼{"─".PadRight(12, '─')}┤");
                // Merged label cell width = 4 + 11 + 22 + 12 + 8 + 4*3 (dropped separators) = 69
                System.Console.Write($"│ {"► AcBinary (Byte[]) vs MemoryPack (Byte[])",-69} │ ");

                // Setup S/D KB (n/a for Byte[] vs Byte[] — neither pre-allocates)
                System.Console.Write($"{"—",14}");
                System.Console.Write(" │ ");

                // Size
                System.Console.ForegroundColor = sizePct <= 0 ? ConsoleColor.Green : ConsoleColor.Red;
                System.Console.Write($"{sizePct,+7:+0;-0}%");
                System.Console.ResetColor();
                System.Console.Write(" │ ");

                // Serialize
                System.Console.ForegroundColor = serPct <= 0 ? ConsoleColor.Green : ConsoleColor.Red;
                System.Console.Write($"{serPct,+9:+0;-0}%");
                System.Console.ResetColor();
                System.Console.Write(" │ ");

                // Serialize Alloc
                System.Console.ForegroundColor = serAllocPct <= 0 ? ConsoleColor.Green : ConsoleColor.Red;
                System.Console.Write($"{serAllocPct,+9:+0;-0}%");
                System.Console.ResetColor();
                System.Console.Write(" │ ");

                // Deserialize
                System.Console.ForegroundColor = desPct <= 0 ? ConsoleColor.Green : ConsoleColor.Red;
                System.Console.Write($"{desPct,+9:+0;-0}%");
                System.Console.ResetColor();
                System.Console.Write(" │ ");

                // Deserialize Alloc
                System.Console.ForegroundColor = desAllocPct <= 0 ? ConsoleColor.Green : ConsoleColor.Red;
                System.Console.Write($"{desAllocPct,+9:+0;-0}%");
                System.Console.ResetColor();
                System.Console.Write(" │ ");

                // Round-trip
                System.Console.ForegroundColor = rtPct <= 0 ? ConsoleColor.Green : ConsoleColor.Red;
                System.Console.Write($"{rtPct,+9:+0;-0}%");
                System.Console.ResetColor();
                System.Console.Write(" │ ");

                // Round-trip Alloc
                System.Console.ForegroundColor = rtAllocPct <= 0 ? ConsoleColor.Green : ConsoleColor.Red;
                System.Console.Write($"{rtAllocPct,+9:+0;-0}%");
                System.Console.ResetColor();
                System.Console.WriteLine(" │");
            }

            // Closing line: merged on left (─ between cols 1-5), ┴ on the right (cols 6-13 boundary, 8 unmerged cells).
            System.Console.WriteLine($"└{"─".PadRight(6, '─')}─{"─".PadRight(13, '─')}─{"─".PadRight(24, '─')}─{"─".PadRight(14, '─')}─{"─".PadRight(10, '─')}┴{"─".PadRight(16, '─')}┴{"─".PadRight(10, '─')}┴{"─".PadRight(12, '─')}┴{"─".PadRight(12, '─')}┴{"─".PadRight(12, '─')}┴{"─".PadRight(12, '─')}┴{"─".PadRight(12, '─')}┴{"─".PadRight(12, '─')}┘");
        }

        // Summary: Best serializer for each category
        System.Console.WriteLine("\n");
        System.Console.WriteLine("╔══════════════════════════════════════════════════════════════════════════════════════════════════════╗");
        System.Console.WriteLine("║                                         SUMMARY: WINNERS                                             ║");
        System.Console.WriteLine("╚══════════════════════════════════════════════════════════════════════════════════════════════════════╝");

        System.Console.WriteLine($"\n{"Category",-20} │ {"Winner",-40} │ {"Avg Value",-18}");
        System.Console.WriteLine($"{"─".PadRight(20, '─')}─┼─{"─".PadRight(40, '─')}─┼─{"─".PadRight(18, '─')}");

        // Fastest Serialize — round-trip-only serializers (NamedPipe etc.) excluded:
        // their Serialize() captures the full round-trip and isn't comparable to a pure Ser metric.
        var fastestSer = results.Where(r => r.SerializeTimeMs > 0 && !r.IsRoundTripOnly)
            .GroupBy(r => r.SerializerName)
            .Select(g => new { Name = g.Key, AvgPerOp = g.Average(r => SerPerOp(r)) })
            .OrderBy(x => x.AvgPerOp)
            .FirstOrDefault();

        if (fastestSer != null)
            System.Console.WriteLine($"{"Fastest Serialize",-20} │ {fastestSer.Name,-40} │ {fastestSer.AvgPerOp,12:F2} µs/op");

        // Fastest Deserialize — round-trip-only serializers excluded (their Deserialize() is a no-op).
        var fastestDes = results.Where(r => r.DeserializeTimeMs > 0 && !r.IsRoundTripOnly)
            .GroupBy(r => r.SerializerName)
            .Select(g => new { Name = g.Key, AvgPerOp = g.Average(r => DesPerOp(r)) })
            .OrderBy(x => x.AvgPerOp)
            .FirstOrDefault();

        if (fastestDes != null)
            System.Console.WriteLine($"{"Fastest Deserialize",-20} │ {fastestDes.Name,-40} │ {fastestDes.AvgPerOp,12:F2} µs/op");

        // Smallest Size
        var smallestSize = results
            .GroupBy(r => r.SerializerName)
            .Select(g => new { Name = g.Key, AvgSize = g.Average(r => r.SerializedSize) })
            .OrderBy(x => x.AvgSize)
            .FirstOrDefault();

        if (smallestSize != null)
            System.Console.WriteLine($"{"Smallest Size",-20} │ {smallestSize.Name,-40} │ {smallestSize.AvgSize,15:F0} B");

        // Fastest Round-trip — iter-independent per-op average.
        var fastestRt = results.Where(r => r.RoundTripTimeMs > 0)
            .GroupBy(r => r.SerializerName)
            .Select(g => new { Name = g.Key, AvgPerOp = g.Average(r => RtPerOp(r)) })
            .OrderBy(x => x.AvgPerOp)
            .FirstOrDefault();

        if (fastestRt != null)
            System.Console.WriteLine($"{"Fastest Round-trip",-20} │ {fastestRt.Name,-40} │ {fastestRt.AvgPerOp,12:F2} µs/op");

        // Overall AcBinary (SGen) vs MemoryPack comparison.
        var memPackSerResults = results.Where(r => (r.Engine == Configuration.EngineMemoryPack && r.IoMode == Configuration.IoByteArray) && r.SerializeTimeMs > 0).ToList();
        var memPackDesResults = results.Where(r => (r.Engine == Configuration.EngineMemoryPack && r.IoMode == Configuration.IoByteArray) && r.DeserializeTimeMs > 0).ToList();
        var memPackRtResults = results.Where(r => (r.Engine == Configuration.EngineMemoryPack && r.IoMode == Configuration.IoByteArray) && r.RoundTripTimeMs > 0).ToList();

        var acBinarySerResults = results.Where(r => (r.Engine == Configuration.EngineAcBinary && r.IoMode == Configuration.IoByteArray && r.DispatchMode == Configuration.ModeSGen) && r.SerializeTimeMs > 0).ToList();
        var acBinaryDesResults = results.Where(r => (r.Engine == Configuration.EngineAcBinary && r.IoMode == Configuration.IoByteArray && r.DispatchMode == Configuration.ModeSGen) && r.DeserializeTimeMs > 0).ToList();
        var acBinaryRtResults = results.Where(r => (r.Engine == Configuration.EngineAcBinary && r.IoMode == Configuration.IoByteArray && r.DispatchMode == Configuration.ModeSGen) && r.RoundTripTimeMs > 0).ToList();

        // Skip comparison if no data available
        if (memPackRtResults.Count == 0 || acBinaryRtResults.Count == 0)
        {
            System.Console.WriteLine();
            System.Console.WriteLine("── AcBinary (Byte[], SGen) vs MemoryPack (Byte[]) (Overall) ──");
            System.Console.WriteLine("  (Comparison requires both serialize and deserialize data)");
            return;
        }

        // All averages are over per-op µs (iter-independent). Three aggregations per metric.
        var sizeAcResults = results.Where(r => (r.Engine == Configuration.EngineAcBinary && r.IoMode == Configuration.IoByteArray && r.DispatchMode == Configuration.ModeSGen)).ToList();
        var sizeMpResults = results.Where(r => (r.Engine == Configuration.EngineMemoryPack && r.IoMode == Configuration.IoByteArray)).ToList();

        var serStats = ComputeOverallStats(acBinarySerResults, memPackSerResults, SerPerOp);
        var desStats = ComputeOverallStats(acBinaryDesResults, memPackDesResults, DesPerOp);
        var rtStats = ComputeOverallStats(acBinaryRtResults, memPackRtResults, RtPerOp);
        var sizeStats = ComputeOverallStats(sizeAcResults, sizeMpResults, r => r.SerializedSize);
        var serAllocStats = ComputeOverallStats(acBinarySerResults, memPackSerResults, r => r.SerializeAllocBytesPerOp);
        var desAllocStats = ComputeOverallStats(acBinaryDesResults, memPackDesResults, r => r.DeserializeAllocBytesPerOp);

        System.Console.WriteLine();
        System.Console.WriteLine("── AcBinary (Byte[], SGen) vs MemoryPack (Byte[]) (Overall) ──");

        WriteOverallLine("Serialize",   "µs/op", serStats);
        WriteOverallLine("Deserialize", "µs/op", desStats);
        WriteOverallLine("Round-trip",  "µs/op", rtStats);
        WriteOverallLine("Size",        "B",     sizeStats, "F0");
        WriteOverallLine("Ser Alloc",   "B/op",  serAllocStats, "F0");
        WriteOverallLine("Des Alloc",   "B/op",  desAllocStats, "F0");
    }

    internal static void SaveResults(List<BenchmarkResult> results, List<TestDataSet> testDataSets)
    {
        Directory.CreateDirectory(Configuration.ResultsDirectory);

        var timestamp = DateTime.Now.ToString("yyyy-MM-dd_HH-mm-ss");
        var baseFileName = $"Console.FullBenchmark_{Configuration.BuildConfiguration}_{timestamp}";
        var logFilePath = Path.Combine(Configuration.ResultsDirectory, $"{baseFileName}.log");
        var outputFilePath = Path.Combine(Configuration.ResultsDirectory, $"{baseFileName}.output");

        // Save binary output to separate .output file
        var largeTestData = testDataSets.FirstOrDefault(t => t.Name.StartsWith("Large"));
        if (largeTestData != null)
        {
            var outputSb = new StringBuilder();
            outputSb.AppendLine("╔══════════════════════════════════════════════════════════════════════════════════════════════════════╗");
            outputSb.AppendLine("║                              SERIALIZED BINARY OUTPUT                                               ║");
            outputSb.AppendLine($"║  Generated: {DateTime.Now:yyyy-MM-dd HH:mm:ss}".PadRight(100) + "║");
            outputSb.AppendLine("╚══════════════════════════════════════════════════════════════════════════════════════════════════════╝");
            outputSb.AppendLine();

            outputSb.AppendLine("=== SERIALIZED BYTES: Large (5x5x5x10) - AcBinary (Default) ===");
            var serializedBytes = AcBinarySerializer.Serialize(largeTestData.Order, AcBinarySerializerOptions.Default);
            outputSb.AppendLine($"Size: {serializedBytes.Length:N0} bytes");
            outputSb.AppendLine();
            outputSb.AppendLine("Hex dump:");
            outputSb.AppendLine(FormatHexDump(serializedBytes));

            File.WriteAllText(outputFilePath, outputSb.ToString(), Configuration.Utf8NoBom);
            System.Console.WriteLine($"✓ Binary output saved to: {outputFilePath}");
        }

        // Save benchmark results to .log file
        var sb = new StringBuilder();
        sb.AppendLine("╔══════════════════════════════════════════════════════════════════════════════════════════════════════╗");
        sb.AppendLine("║                              SERIALIZER BENCHMARK RESULTS                                            ║");
        sb.AppendLine($"║  Generated: {DateTime.Now:yyyy-MM-dd HH:mm:ss}".PadRight(100) + "║");
        sb.AppendLine($"║  Build: {Configuration.BuildConfiguration}".PadRight(100) + "║");
        sb.AppendLine($"║  Charset: {Configuration.GetCurrentCharsetName()}".PadRight(100) + "║");
        sb.AppendLine($"║  Iterations: per-cell adaptive (~{Configuration.TargetSampleMs} ms target)".PadRight(100) + "║");
        sb.AppendLine($"║  Samples: {Configuration.BenchmarkSamples} (median) + 1 pilot discarded".PadRight(100) + "║");
        sb.AppendLine($"║  Test Type: {testDataSets.FirstOrDefault()?.TypeName ?? "unknown"}".PadRight(100) + "║");
        sb.AppendLine("╚══════════════════════════════════════════════════════════════════════════════════════════════════════╝");
        sb.AppendLine();

        // Serializer options summary
        var optionsMap = results
            .Where(r => r.OptionsDescription != null)
            .Select(r => (r.SerializerName, r.OptionsDescription!))
            .Distinct()
            .ToList();
        if (optionsMap.Count > 0)
        {
            sb.AppendLine("=== SERIALIZER OPTIONS ===");
            foreach (var (name, opts) in optionsMap)
                sb.AppendLine($"  {name}: {opts}");
            sb.AppendLine();
        }

        // CSV-like data for easy import — keeps raw byte integers (no KB rounding) so external tools can compute precisely.
        sb.AppendLine("=== RAW DATA (CSV) ===");
        sb.AppendLine("TestData,Engine,IO,Mode,Options,Size,SerializeMicrosPerOp,DeserializeMicrosPerOp,RoundTripMicrosPerOp,SerializeAllocBytesPerOp,DeserializeAllocBytesPerOp,RoundTripAllocBytesPerOp,SetupSerializeAllocBytes,SetupDeserializeAllocBytes");

        foreach (var testData in testDataSets)
        {
            var testResults = results.Where(r => r.TestDataName == testData.DisplayName).ToList();
            foreach (var result in testResults)
            {
                sb.AppendLine($"{result.TestDataName},{result.Engine},{result.IoMode},{result.DispatchMode},{result.OptionsPreset},{result.SerializedSize},{SerPerOp(result):F2},{DesPerOp(result):F2},{RtPerOp(result):F2},{result.SerializeAllocBytesPerOp},{result.DeserializeAllocBytesPerOp},{result.RoundTripAllocBytesPerOp},{result.SetupSerializeAllocBytes},{result.SetupDeserializeAllocBytes}");
            }
        }
        sb.AppendLine();

        // Formatted results
        sb.AppendLine("=== FORMATTED RESULTS BY TEST DATA ===");
        sb.AppendLine("(►) = Highlighted: MemoryPack (Byte[]) (baseline) and AcBinary (Byte[])");
        sb.AppendLine();

        foreach (var testData in testDataSets)
        {
            // Order by per-op µs (iter-independent) — rows may have different iter counts post-calibration.
            var testResults = results.Where(r => r.TestDataName == testData.DisplayName).OrderBy(r => RtPerOp(r)).ToList();
            var memPackResult = testResults.FirstOrDefault(r => (r.Engine == Configuration.EngineMemoryPack && r.IoMode == Configuration.IoByteArray));
            var acBinaryResult = testResults.FirstOrDefault(r => (r.Engine == Configuration.EngineAcBinary && r.IoMode == Configuration.IoByteArray && r.DispatchMode == Configuration.ModeSGen));

            sb.AppendLine();
            sb.AppendLine($"--- {testData.DisplayName} ---");
            sb.AppendLine($"{"#",-4} {"Serializer",-42} {"Size B",-12} {"Setup S/D KB",-14} {"Ser µs/op",-12} {"Des µs/op",-12} {"RT µs/op",-12} {"SerAlc KB",-11} {"DesAlc KB",-11}");
            sb.AppendLine(new string('-', 140));

            var rank = 1;
            foreach (var result in testResults)
            {
                var isHighlighted = ((result.Engine == Configuration.EngineMemoryPack || result.Engine == Configuration.EngineAcBinary) && result.IoMode == Configuration.IoByteArray);
                var prefix = isHighlighted ? "► " : "  ";

                var size = $"{result.SerializedSize:N0}";
                var setup = $"{ToKilobytes(result.SetupSerializeAllocBytes):F2} / {ToKilobytes(result.SetupDeserializeAllocBytes):F2}";
                var ser = result.SerializeTimeMs > 0 ? $"{SerPerOp(result):F2}" : "N/A";
                var des = result.DeserializeTimeMs > 0 ? $"{DesPerOp(result):F2}" : "N/A";
                var rt = result.RoundTripTimeMs > 0 ? $"{RtPerOp(result):F2}" : "N/A";
                var serAlloc = result.SerializeTimeMs > 0 ? $"{ToKilobytes(result.SerializeAllocBytesPerOp):F2}" : "N/A";
                var desAlloc = result.DeserializeTimeMs > 0 ? $"{ToKilobytes(result.DeserializeAllocBytesPerOp):F2}" : "N/A";

                sb.AppendLine($"{rank++,2} {prefix}{result.SerializerName,-40} {size,-12} {setup,-14} {ser,-12} {des,-12} {rt,-12} {serAlloc,-11} {desAlloc,-11}");
            }

            // Summary row for this test data (vs MemoryPack)
            if (memPackResult != null && acBinaryResult != null)
            {
                var sizePct = (acBinaryResult.SerializedSize / (double)memPackResult.SerializedSize - 1) * 100;
                var serPct = SerPerOp(memPackResult) > 0 ? (SerPerOp(acBinaryResult) / SerPerOp(memPackResult) - 1) * 100 : 0;
                var desPct = DesPerOp(memPackResult) > 0 ? (DesPerOp(acBinaryResult) / DesPerOp(memPackResult) - 1) * 100 : 0;
                var rtPct = RtPerOp(memPackResult) > 0 ? (RtPerOp(acBinaryResult) / RtPerOp(memPackResult) - 1) * 100 : 0;

                sb.AppendLine($"   AcBinary (Byte[]) vs MemoryPack (Byte[]): Size {sizePct:+0;-0}% │ Ser {serPct:+0;-0}% │ Des {desPct:+0;-0}% │ RT {rtPct:+0;-0}%");
            }
        }

        // Summary comparison (vs MemoryPack)
        sb.AppendLine();
        sb.AppendLine("=== AcBinary (Byte[], SGen) vs MemoryPack (Byte[]) (Overall) ===");

        var memPackSerResults2 = results.Where(r => (r.Engine == Configuration.EngineMemoryPack && r.IoMode == Configuration.IoByteArray) && r.SerializeTimeMs > 0).ToList();
        var memPackDesResults2 = results.Where(r => (r.Engine == Configuration.EngineMemoryPack && r.IoMode == Configuration.IoByteArray) && r.DeserializeTimeMs > 0).ToList();
        var memPackRtResults2 = results.Where(r => (r.Engine == Configuration.EngineMemoryPack && r.IoMode == Configuration.IoByteArray) && r.RoundTripTimeMs > 0).ToList();

        var acBinarySerResults2 = results.Where(r => (r.Engine == Configuration.EngineAcBinary && r.IoMode == Configuration.IoByteArray && r.DispatchMode == Configuration.ModeSGen) && r.SerializeTimeMs > 0).ToList();
        var acBinaryDesResults2 = results.Where(r => (r.Engine == Configuration.EngineAcBinary && r.IoMode == Configuration.IoByteArray && r.DispatchMode == Configuration.ModeSGen) && r.DeserializeTimeMs > 0).ToList();
        var acBinaryRtResults2 = results.Where(r => (r.Engine == Configuration.EngineAcBinary && r.IoMode == Configuration.IoByteArray && r.DispatchMode == Configuration.ModeSGen) && r.RoundTripTimeMs > 0).ToList();

        // Skip comparison block if either side has no Byte[] data
        if (memPackRtResults2.Count == 0 || acBinaryRtResults2.Count == 0)
        {
            sb.AppendLine("  (Comparison requires both serialize and deserialize data)");
            File.WriteAllText(logFilePath, sb.ToString(), Configuration.Utf8NoBom);
            System.Console.WriteLine($"✓ Results saved to: {logFilePath}");

            var llmFilePathEarly = Path.Combine(Configuration.ResultsDirectory, $"{baseFileName}.LLM");
            SaveLlmResults(llmFilePathEarly, results, testDataSets);
            return;
        }

        var sizeAcResults2 = results.Where(r => (r.Engine == Configuration.EngineAcBinary && r.IoMode == Configuration.IoByteArray && r.DispatchMode == Configuration.ModeSGen)).ToList();
        var sizeMpResults2 = results.Where(r => (r.Engine == Configuration.EngineMemoryPack && r.IoMode == Configuration.IoByteArray)).ToList();

        AppendOverallLine(sb, "Serialize",   "µs/op", ComputeOverallStats(acBinarySerResults2, memPackSerResults2, SerPerOp));
        AppendOverallLine(sb, "Ser Alloc",   "B/op",  ComputeOverallStats(acBinarySerResults2, memPackSerResults2, r => r.SerializeAllocBytesPerOp), "F0");
        AppendOverallLine(sb, "Deserialize", "µs/op", ComputeOverallStats(acBinaryDesResults2, memPackDesResults2, DesPerOp));
        AppendOverallLine(sb, "Des Alloc",   "B/op",  ComputeOverallStats(acBinaryDesResults2, memPackDesResults2, r => r.DeserializeAllocBytesPerOp), "F0");
        AppendOverallLine(sb, "Round-trip",  "µs/op", ComputeOverallStats(acBinaryRtResults2, memPackRtResults2, RtPerOp));
        AppendOverallLine(sb, "Size",        "B",     ComputeOverallStats(sizeAcResults2, sizeMpResults2, r => r.SerializedSize), "F0");

        File.WriteAllText(logFilePath, sb.ToString(), Configuration.Utf8NoBom);
        System.Console.WriteLine($"✓ Results saved to: {logFilePath}");

        // Save LLM-optimized results
        var llmFilePath = Path.Combine(Configuration.ResultsDirectory, $"{baseFileName}.LLM");
        SaveLlmResults(llmFilePath, results, testDataSets);
    }

    internal static void SaveLlmResults(string filePath, List<BenchmarkResult> results, List<TestDataSet> testDataSets)
    {
        var sb = new StringBuilder();
        var testTypeName = testDataSets.FirstOrDefault()?.TypeName ?? "unknown";
        sb.AppendLine($"# AcBinary Benchmark {Configuration.BuildConfiguration} {DateTime.Now:yyyy-MM-dd HH:mm:ss}");
        sb.AppendLine($"Charset: {Configuration.GetCurrentCharsetName()} | Iterations: per-cell adaptive (target ~{Configuration.TargetSampleMs} ms/sample) | Warmup: {Configuration.WarmupIterations} per phase (Ser/Des isolated) | Samples: {Configuration.BenchmarkSamples} (median) + 1 pilot discarded | .NET: {Environment.Version} | TestType: {testTypeName} | UnstableCV threshold: {Configuration.UnstableCVThreshold * 100:F0}%");
        sb.AppendLine("Baseline: MemoryPack (Byte[]) (SOTA reference) | Verified: round-trip correctness checked once per cell before warmup");

        // Options summary
        var optionsMap = results
            .Where(r => r.OptionsDescription != null)
            .Select(r => (r.SerializerName, r.OptionsDescription!))
            .Distinct()
            .ToList();
        if (optionsMap.Count > 0)
        {
            sb.AppendLine();
            sb.AppendLine("## Options");
            sb.AppendLine();
            foreach (var (name, opts) in optionsMap)
                sb.AppendLine($"- **{name}**: {opts}");
        }

        sb.AppendLine();
        sb.AppendLine("## Results");
        sb.AppendLine();
        sb.AppendLine("TestData | Engine | IO | Mode | Options | Size(B) | Ser(µs/op) | Deser(µs/op) | RT(µs/op) | SerAlloc(KB/op) | DesAlloc(KB/op) | RTAlloc(KB/op) | Setup S/D(KB) | Iter Ser/Des");
        sb.AppendLine("---|---|---|---|---|---|---|---|---|---|---|---|---|---");

        foreach (var testData in testDataSets)
        {
            var testResults = results
                .Where(r => r.TestDataName == testData.DisplayName)
                .OrderBy(RtPerOp)
                .ToList();

            foreach (var r in testResults)
            {
                var inv = CultureInfo.InvariantCulture;
                var ser = r.SerializeTimeMs > 0 ? FormatMicrosWithRange(r.SerializeTimeMs, r.SerializeTimeMinMs, r.SerializeTimeMaxMs, r.SerializeTimeStdDevMs, r.SerializeIterations, inv) : "-";
                var des = r.DeserializeTimeMs > 0 ? FormatMicrosWithRange(r.DeserializeTimeMs, r.DeserializeTimeMinMs, r.DeserializeTimeMaxMs, r.DeserializeTimeStdDevMs, r.DeserializeIterations, inv) : "-";
                var rt = r.RoundTripTimeMs > 0
                    ? (r.IsRoundTripOnly
                        ? FormatMicrosWithRange(r.RoundTripTimeMs, r.RoundTripTimeMinMs, r.RoundTripTimeMaxMs, r.RoundTripTimeStdDevMs, r.RoundTripIterations, inv)
                        : RtPerOp(r).ToString("F2", inv))
                    : "-";

                var serAlloc = r.SerializeTimeMs > 0 ? ToKilobytes(r.SerializeAllocBytesPerOp).ToString("F2", inv) : "-";
                var desAlloc = r.DeserializeTimeMs > 0 ? ToKilobytes(r.DeserializeAllocBytesPerOp).ToString("F2", inv) : "-";
                var rtAlloc = r.RoundTripAllocBytesPerOp > 0 ? ToKilobytes(r.RoundTripAllocBytesPerOp).ToString("F2", inv) : "-";
                var setupAlloc = $"{ToKilobytes(r.SetupSerializeAllocBytes).ToString("F2", inv)} / {ToKilobytes(r.SetupDeserializeAllocBytes).ToString("F2", inv)}";

                var iterCol = r.IsRoundTripOnly
                    ? r.RoundTripIterations.ToString(inv)
                    : $"{(r.SerializeIterations > 0 ? r.SerializeIterations.ToString(inv) : "-")} / {(r.DeserializeIterations > 0 ? r.DeserializeIterations.ToString(inv) : "-")}";
                sb.AppendLine($"{r.TestDataName} | {r.Engine} | {r.IoMode} | {r.DispatchMode} | {r.OptionsPreset} | {r.SerializedSize} | {ser} | {des} | {rt} | {serAlloc} | {desAlloc} | {rtAlloc} | {setupAlloc} | {iterCol}");
            }
        }

        // Overall AcBinary (SGen, Byte[]) vs MemoryPack (Byte[]) comparison
        var memPackByteArrayResults = results.Where(r => r.Engine == Configuration.EngineMemoryPack && r.IoMode == Configuration.IoByteArray).ToList();
        var acBinarySGenByteArrayResults = results.Where(r => r.Engine == Configuration.EngineAcBinary && r.IoMode == Configuration.IoByteArray && r.DispatchMode == Configuration.ModeSGen).ToList();
        var memPackSerResultsLlm = memPackByteArrayResults.Where(r => r.SerializeTimeMs > 0).ToList();
        var memPackDesResultsLlm = memPackByteArrayResults.Where(r => r.DeserializeTimeMs > 0).ToList();
        var memPackRtResultsLlm = memPackByteArrayResults.Where(r => r.RoundTripTimeMs > 0).ToList();
        var acBinarySerResultsLlm = acBinarySGenByteArrayResults.Where(r => r.SerializeTimeMs > 0).ToList();
        var acBinaryDesResultsLlm = acBinarySGenByteArrayResults.Where(r => r.DeserializeTimeMs > 0).ToList();
        var acBinaryRtResultsLlm = acBinarySGenByteArrayResults.Where(r => r.RoundTripTimeMs > 0).ToList();

        if (memPackRtResultsLlm.Count > 0 && acBinaryRtResultsLlm.Count > 0)
        {
            sb.AppendLine();
            sb.AppendLine("## Overall: AcBinary (Byte[], SGen) vs MemoryPack (Byte[])");
            sb.AppendLine();
            sb.AppendLine("Three aggregations of per-cell results: **arith** = arithmetic mean of µs/op (magnitude-weighted, Large cell dominates); **geo** = geometric mean of per-cell ratios (each cell weighted equally); **median** = median of per-cell ratios (outlier-resistant). Negative % = AcBinary faster/smaller; positive % = MemPack faster/smaller. The geo/median variants surface when a single big cell skews the arithmetic mean.");
            sb.AppendLine();
            sb.AppendLine("```");
            AppendOverallLine(sb, "Serialize",   "µs/op", ComputeOverallStats(acBinarySerResultsLlm, memPackSerResultsLlm, SerPerOp));
            AppendOverallLine(sb, "Ser Alloc",   "B/op",  ComputeOverallStats(acBinarySerResultsLlm, memPackSerResultsLlm, r => r.SerializeAllocBytesPerOp), "F0");
            AppendOverallLine(sb, "Deserialize", "µs/op", ComputeOverallStats(acBinaryDesResultsLlm, memPackDesResultsLlm, DesPerOp));
            AppendOverallLine(sb, "Des Alloc",   "B/op",  ComputeOverallStats(acBinaryDesResultsLlm, memPackDesResultsLlm, r => r.DeserializeAllocBytesPerOp), "F0");
            AppendOverallLine(sb, "Round-trip",  "µs/op", ComputeOverallStats(acBinaryRtResultsLlm, memPackRtResultsLlm, RtPerOp));
            AppendOverallLine(sb, "Size",        "B",     ComputeOverallStats(acBinarySGenByteArrayResults, memPackByteArrayResults, r => r.SerializedSize), "F0");
            sb.AppendLine("```");
        }

        File.WriteAllText(filePath, sb.ToString(), Configuration.Utf8NoBom);
        System.Console.WriteLine($"✓ LLM results saved to: {filePath}");
    }

    /// <summary>
    /// Formats byte array as hex dump with offset, hex values, and ASCII representation.
    /// </summary>
    internal static string FormatHexDump(byte[] bytes, int bytesPerLine = 16)
    {
        var sb = new StringBuilder();
        for (var i = 0; i < bytes.Length; i += bytesPerLine)
        {
            // Offset
            sb.Append($"{i:X8}  ");

            // Hex bytes
            for (var j = 0; j < bytesPerLine; j++)
            {
                if (i + j < bytes.Length)
                    sb.Append($"{bytes[i + j]:X2} ");
                else
                    sb.Append("   ");

                if (j == 7) sb.Append(' '); // Extra space in middle
            }

            sb.Append(" |");

            // ASCII representation
            for (var j = 0; j < bytesPerLine && i + j < bytes.Length; j++)
            {
                var b = bytes[i + j];
                sb.Append(b is >= 32 and < 127 ? (char)b : '.');
            }

            sb.AppendLine("|");
        }
        return sb.ToString();
    }
}