/** Benchmark result rendering: markdown + JSON, stamped and comparable. */
import { PromptVariant } from "../src/summarizer";
import { StructuralReport } from "./metrics";

export interface FixtureResult {
  fixture: string;
  asked: number;
  correctAfterCompaction: number;
  correctAfterConsolidation?: number;
  reductionPercent: number;
  latency: { mean: number; p50: number; max: number };
  structural: StructuralReport;
  cacheDeterministic: boolean;
}

export interface BenchRun {
  date: string;
  modelId: string;
  /** Summarizer prompt variant benched (see bench/run.ts's --variant flag). Omitted/undefined means "full", the default. */
  variant?: PromptVariant;
  settings: Record<string, unknown>;
  fixtures: FixtureResult[];
}

export function renderMarkdown(run: BenchRun): string {
  const lines: string[] = [
    `# context-compressor benchmark — ${run.date}`,
    "",
    `Model: \`${run.modelId}\``,
    "",
    ...(run.variant !== undefined && run.variant !== "full"
      ? [`Variant: \`${run.variant}\``, ""]
      : []),
    `Settings: \`${JSON.stringify(run.settings)}\``,
    "",
    "| Fixture | Recall (compacted) | Recall (consolidated) | Reduction | Chunk latency mean/p50/max (ms) | Structure | Cache |",
    "|---|---|---|---|---|---|---|",
  ];
  for (const f of run.fixtures) {
    const consolidated =
      f.correctAfterConsolidation !== undefined
        ? `${f.correctAfterConsolidation}/${f.asked}`
        : "—";
    lines.push(
      `| ${f.fixture} | ${f.correctAfterCompaction}/${f.asked} | ${consolidated} | ${f.reductionPercent}% | ${f.latency.mean}/${f.latency.p50}/${f.latency.max} | ${f.structural.ok ? "✅" : "❌"} | ${f.cacheDeterministic ? "✅" : "❌"} |`,
    );
  }
  lines.push(
    "",
    "Recall = seeded facts answered correctly from the compacted view alone. " +
      "Results are comparable only within one model; see bench/README.md.",
  );
  return lines.join("\n");
}

/**
 * e.g. "2026-08-20-qwen-qwen3.8-27b" — filesystem-safe. Non-"full" variants
 * get a suffix (e.g. "-compact") so runs against a different prompt variant
 * never overwrite the default run's result files.
 */
export function resultFileBase(run: BenchRun): string {
  const model = run.modelId.replace(/[^a-z0-9.-]+/gi, "-");
  const variantSuffix =
    run.variant !== undefined && run.variant !== "full" ? `-${run.variant}` : "";
  return `${run.date}-${model}${variantSuffix}`;
}

/** Benchmark result rendering: markdown + JSON, stamped and comparable. */
import { PromptVariant } from "../src/summarizer";
import { StructuralReport } from "./metrics";

export interface FixtureResult {
  fixture: string;
  asked: number;
  correctAfterCompaction: number;
  correctAfterConsolidation?: number;
  reductionPercent: number;
  latency: { mean: number; p50: number; max: number };
  structural: StructuralReport;
  cacheDeterministic: boolean;
}

export interface BenchRun {
  date: string;
  modelId: string;
  /** Summarizer prompt variant benched (see bench/run.ts's --variant flag). Omitted/undefined means "full", the default. */
  variant?: PromptVariant;
  settings: Record<string, unknown>;
  fixtures: FixtureResult[];
}

export function renderMarkdown(run: BenchRun): string {
  const lines: string[] = [
    `# context-compressor benchmark — ${run.date}`,
    "",
    `Model: \`${run.modelId}\``,
    "",
    ...(run.variant !== undefined && run.variant !== "full"
      ? [`Variant: \`${run.variant}\``, ""]
      : []),
    `Settings: \`${JSON.stringify(run.settings)}\``,
    "",
    "| Fixture | Recall (compacted) | Recall (consolidated) | Reduction | Chunk latency mean/p50/max (ms) | Structure | Cache |",
    "|---|---|---|---|---|---|---|",
  ];
  for (const f of run.fixtures) {
    const consolidated =
      f.correctAfterConsolidation !== undefined
        ? `${f.correctAfterConsolidation}/${f.asked}`
        : "—";
    lines.push(
      `| ${f.fixture} | ${f.correctAfterCompaction}/${f.asked} | ${consolidated} | ${f.reductionPercent}% | ${f.latency.mean}/${f.latency.p50}/${f.latency.max} | ${f.structural.ok ? "✅" : "❌"} | ${f.cacheDeterministic ? "✅" : "❌"} |`,
    );
  }
  lines.push(
    "",
    "Recall = seeded facts answered correctly from the compacted view alone. " +
      "Results are comparable only within one model; see bench/README.md.",
  );
  return lines.join("\n");
}

/**
 * e.g. "2026-08-20-qwen-qwen3.8-27b" — filesystem-safe. Non-"full" variants
 * get a suffix (e.g. "-compact") so runs against a different prompt variant
 * never overwrite the default run's result files.
 */
export function resultFileBase(run: BenchRun): string {
  const model = run.modelId.replace(/[^a-z0-9.-]+/gi, "-");
  const variantSuffix =
    run.variant !== undefined && run.variant !== "full" ? `-${run.variant}` : "";
  return `${run.date}-${model}${variantSuffix}`;
}