// Phase 4 contamination-firewall tests. The SFT leak-prevention block
// is the most important set: it MUST be impossible for rejected or
// needs_human_review records to reach exports/sft/instruction_response.jsonl
// regardless of how the input data is crafted.
//
// Strategy: synthesize evidence + scored-runs in a temp root, run each
// exporter, assert outputs and quarantine.

import { test, expect, beforeEach, afterEach } from "bun:test";
import { mkdirSync, writeFileSync, rmSync, existsSync, readFileSync } from "node:fs";
import { resolve } from "node:path";

import { exportRag } from "../../scripts/distillation/export_rag";
import { exportSft } from "../../scripts/distillation/export_sft";
import { exportPreference } from "../../scripts/distillation/export_preference";
import { validateRagSample } from "../../auditor/schemas/distillation/rag_sample";
import { validateSftSample } from "../../auditor/schemas/distillation/sft_sample";
import { validatePreferenceSample } from "../../auditor/schemas/distillation/preference_sample";
import { EVIDENCE_SCHEMA_VERSION, type EvidenceRecord, type ModelRole } from "../../auditor/schemas/distillation/evidence_record";
import { SCORED_RUN_SCHEMA_VERSION, type ScoredRun, type ScoreCategory } from "../../auditor/schemas/distillation/scored_run";

const TMP = "/tmp/distillation_test_phase4";
const NOW = "2026-04-26T22:30:00.000Z";
const SHA = "0".repeat(64);
const PARTITION = "2026/04/27";

function setupRoot() {
  if (existsSync(TMP)) rmSync(TMP, { recursive: true, force: true });
  mkdirSync(resolve(TMP, `data/evidence/${PARTITION}`), { recursive: true });
  mkdirSync(resolve(TMP, `data/scored-runs/${PARTITION}`), { recursive: true });
}

function writeEvidence(stem: string, evs: EvidenceRecord[]) {
  const path = resolve(TMP, `data/evidence/${PARTITION}/${stem}.jsonl`);
  writeFileSync(path, evs.map(e => JSON.stringify(e)).join("\n") + "\n");
}

function writeScored(stem: string, scored: ScoredRun[]) {
  const path = resolve(TMP, `data/scored-runs/${PARTITION}/${stem}.jsonl`);
  writeFileSync(path, scored.map(s => JSON.stringify(s)).join("\n") + "\n");
}

function makeEv(opts: { run_id: string; task_id: string; source_stem: string; text?: string; role?: ModelRole; source_files?: string[] }): EvidenceRecord {
  return {
    run_id: opts.run_id,
    task_id: opts.task_id,
    timestamp: NOW,
    schema_version: EVIDENCE_SCHEMA_VERSION,
    provenance: {
      source_file: `data/_kb/${opts.source_stem}.jsonl`,
      line_offset: 0,
      sig_hash: SHA,
      recorded_at: NOW,
    },
    model_role: opts.role ?? "executor",
    text: opts.text ?? "default response text",
    source_files: opts.source_files,
  };
}

function makeScored(opts: { run_id: string; task_id: string; category: ScoreCategory; reasons: string[]; out_relpath: string }): ScoredRun {
  return {
    schema_version: SCORED_RUN_SCHEMA_VERSION,
    evidence_run_id: opts.run_id,
    evidence_task_id: opts.task_id,
    category: opts.category,
    reasons: opts.reasons,
    scored_at: NOW,
    scorer_version: "v1.0.0",
    sub_scores: {},
    provenance: {
      source_file: opts.out_relpath,
      line_offset: 0,
      sig_hash: SHA,
      recorded_at: NOW,
    },
  };
}

beforeEach(setupRoot);
afterEach(() => { if (existsSync(TMP)) rmSync(TMP, { recursive: true, force: true }); });

// ─── RAG export ─────────────────────────────────────────────────────

test("RAG: accepted + partial flow through; rejected quarantined", () => {
  writeEvidence("scrum_reviews", [
    makeEv({ run_id: "ra1", task_id: "t1", source_stem: "scrum_reviews", text: "good review" }),
    makeEv({ run_id: "ra2", task_id: "t2", source_stem: "scrum_reviews", text: "ok review" }),
    makeEv({ run_id: "ra3", task_id: "t3", source_stem: "scrum_reviews", text: "bad review" }),
  ]);
  writeScored("scrum_reviews", [
    makeScored({ run_id: "ra1", task_id: "t1", category: "accepted", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
    makeScored({ run_id: "ra2", task_id: "t2", category: "partially_accepted", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
    makeScored({ run_id: "ra3", task_id: "t3", category: "rejected", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
  ]);
});

test("RAG: needs_human_review excluded by default, included with flag", async () => {
  writeEvidence("scrum_reviews", [
    makeEv({ run_id: "rh1", task_id: "t1", source_stem: "scrum_reviews", text: "default skip" }),
  ]);
  writeScored("scrum_reviews", [
    makeScored({ run_id: "rh1", task_id: "t1", category: "needs_human_review", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
  ]);

  const r1 = await exportRag({ root: TMP, recorded_at: NOW, include_review: false });
  expect(r1.records_exported).toBe(0);
  expect(r1.records_quarantined).toBe(1);

  // Reset for include_review run.
  if (existsSync(resolve(TMP, "exports"))) rmSync(resolve(TMP, "exports"), { recursive: true });
  const r2 = await exportRag({ root: TMP, recorded_at: NOW, include_review: true });
  expect(r2.records_exported).toBe(1);
  expect(r2.records_quarantined).toBe(0);
});

test("RAG: every output row validates against RagSample schema", async () => {
  writeEvidence("scrum_reviews", [
    makeEv({ run_id: "rv1", task_id: "t1", source_stem: "scrum_reviews", text: "review content" }),
  ]);
  writeScored("scrum_reviews", [
    makeScored({ run_id: "rv1", task_id: "t1", category: "accepted", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
  ]);
  await exportRag({ root: TMP, recorded_at: NOW });
  const path = resolve(TMP, "exports/rag/playbooks.jsonl");
  expect(existsSync(path)).toBe(true);
  const rows = readFileSync(path, "utf8").trim().split("\n").map(l => JSON.parse(l));
  for (const row of rows) {
    const v = validateRagSample(row);
    expect(v.valid).toBe(true);
  }
});

test("RAG: empty content quarantined", async () => {
  writeEvidence("scrum_reviews", [makeEv({ run_id: "re1", task_id: "t1", source_stem: "scrum_reviews", text: "" })]);
  writeScored("scrum_reviews", [makeScored({ run_id: "re1", task_id: "t1", category: "accepted", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` })]);
  const r = await exportRag({ root: TMP, recorded_at: NOW });
  expect(r.records_exported).toBe(0);
  expect(r.records_quarantined).toBe(1);
  const qPath = resolve(TMP, "exports/quarantine/rag.jsonl");
  expect(existsSync(qPath)).toBe(true);
  expect(readFileSync(qPath, "utf8")).toContain("empty_content");
});

// ─── SFT export — THE CONTAMINATION FIREWALL ────────────────────────

test("SFT: rejected NEVER ships (spec non-negotiable)", async () => {
  writeEvidence("scrum_reviews", [
    makeEv({ run_id: "sf1", task_id: "t1", source_stem: "scrum_reviews", text: "rejected output that should NOT train" }),
  ]);
  writeScored("scrum_reviews", [
    makeScored({ run_id: "sf1", task_id: "t1", category: "rejected", reasons: ["bad"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
  ]);
  const r = await exportSft({ root: TMP, recorded_at: NOW });
  expect(r.records_exported).toBe(0);
  expect(r.records_quarantined).toBe(1);
  const qPath = resolve(TMP, "exports/quarantine/sft.jsonl");
  expect(readFileSync(qPath, "utf8")).toContain("unsafe_sft_category");
});

test("SFT: needs_human_review NEVER ships (spec non-negotiable)", async () => {
  writeEvidence("scrum_reviews", [
    makeEv({ run_id: "sh1", task_id: "t1", source_stem: "scrum_reviews", text: "hum text" }),
  ]);
  writeScored("scrum_reviews", [
    makeScored({ run_id: "sh1", task_id: "t1", category: "needs_human_review", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
  ]);
  const r = await exportSft({ root: TMP, recorded_at: NOW });
  expect(r.records_exported).toBe(0);
  expect(r.records_quarantined).toBe(1);
});

test("SFT: partially_accepted excluded by default; included with --include-partial", async () => {
  writeEvidence("scrum_reviews", [makeEv({ run_id: "sp1", task_id: "t1", source_stem: "scrum_reviews", text: "partial output" })]);
  writeScored("scrum_reviews", [makeScored({ run_id: "sp1", task_id: "t1", category: "partially_accepted", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` })]);

  const r1 = await exportSft({ root: TMP, recorded_at: NOW });
  expect(r1.records_exported).toBe(0);
  expect(r1.records_quarantined).toBe(1);

  // Reset for include_partial.
  if (existsSync(resolve(TMP, "exports"))) rmSync(resolve(TMP, "exports"), { recursive: true });
  const r2 = await exportSft({ root: TMP, recorded_at: NOW, include_partial: true });
  expect(r2.records_exported).toBe(1);
});

test("SFT: extraction-class records (no instruction→response shape) quarantined", async () => {
  writeEvidence("distilled_facts", [
    makeEv({ run_id: "sx1", task_id: "t1", source_stem: "distilled_facts", text: "extracted fact", role: "extractor" }),
  ]);
  writeScored("distilled_facts", [
    // Force category=accepted to prove it's the role-shape gate that catches it, not the category gate.
    makeScored({ run_id: "sx1", task_id: "t1", category: "accepted", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/distilled_facts.jsonl` }),
  ]);
  const r = await exportSft({ root: TMP, recorded_at: NOW });
  expect(r.records_exported).toBe(0);
  expect(r.records_quarantined).toBe(1);
});

test("SFT: every output row validates against SftSample (provenance + non-empty + quality_score)", async () => {
  writeEvidence("scrum_reviews", [
    makeEv({ run_id: "sv1", task_id: "t1", source_stem: "scrum_reviews", text: "real instruction response text" }),
  ]);
  writeScored("scrum_reviews", [
    makeScored({ run_id: "sv1", task_id: "t1", category: "accepted", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
  ]);
  await exportSft({ root: TMP, recorded_at: NOW });
  const rows = readFileSync(resolve(TMP, "exports/sft/instruction_response.jsonl"), "utf8").trim().split("\n").map(l => JSON.parse(l));
  expect(rows.length).toBe(1);
  for (const row of rows) {
    const v = validateSftSample(row);
    expect(v.valid).toBe(true);
    expect(row.quality_score).toBe("accepted");  // never partial here
    expect(row.provenance.sig_hash).toMatch(/^[0-9a-f]{64}$/);
  }
});

test("SFT: idempotent — second run produces 0 new exports", async () => {
  writeEvidence("scrum_reviews", [makeEv({ run_id: "si1", task_id: "t1", source_stem: "scrum_reviews", text: "idem" })]);
  writeScored("scrum_reviews", [makeScored({ run_id: "si1", task_id: "t1", category: "accepted", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` })]);
  await exportSft({ root: TMP, recorded_at: NOW });
  const r2 = await exportSft({ root: TMP, recorded_at: NOW });
  expect(r2.records_exported).toBe(0);
});

// ─── Preference export — pairing logic ──────────────────────────────

test("Preference: same task_id, accepted vs rejected → exports a pair", async () => {
  writeEvidence("scrum_reviews", [
    makeEv({ run_id: "pa1", task_id: "task-X", source_stem: "scrum_reviews", text: "good chosen output" }),
    makeEv({ run_id: "pr1", task_id: "task-X", source_stem: "scrum_reviews", text: "bad rejected output" }),
  ]);
  writeScored("scrum_reviews", [
    makeScored({ run_id: "pa1", task_id: "task-X", category: "accepted", reasons: ["good"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
    makeScored({ run_id: "pr1", task_id: "task-X", category: "rejected", reasons: ["bad"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
  ]);
  const r = await exportPreference({ root: TMP, recorded_at: NOW });
  expect(r.pairs_exported).toBe(1);
  const path = resolve(TMP, "exports/preference/chosen_rejected.jsonl");
  const rows = readFileSync(path, "utf8").trim().split("\n").map(l => JSON.parse(l));
  expect(rows.length).toBe(1);
  expect(rows[0].chosen).toContain("good");
  expect(rows[0].rejected).toContain("bad");
  expect(rows[0].chosen_run_id).not.toBe(rows[0].rejected_run_id);
  const v = validatePreferenceSample(rows[0]);
  expect(v.valid).toBe(true);
});

test("Preference: different task_ids never pair (no fabrication)", async () => {
  writeEvidence("scrum_reviews", [
    makeEv({ run_id: "pd1", task_id: "task-A", source_stem: "scrum_reviews", text: "A good" }),
    makeEv({ run_id: "pd2", task_id: "task-B", source_stem: "scrum_reviews", text: "B bad" }),
  ]);
  writeScored("scrum_reviews", [
    makeScored({ run_id: "pd1", task_id: "task-A", category: "accepted", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
    makeScored({ run_id: "pd2", task_id: "task-B", category: "rejected", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
  ]);
  const r = await exportPreference({ root: TMP, recorded_at: NOW });
  expect(r.pairs_exported).toBe(0);  // no shared task_id
  expect(r.insufficient_pair_task_ids).toBe(2);
});

test("Preference: identical text in chosen and rejected quarantined", async () => {
  writeEvidence("scrum_reviews", [
    makeEv({ run_id: "pi1", task_id: "task-X", source_stem: "scrum_reviews", text: "identical text" }),
    makeEv({ run_id: "pi2", task_id: "task-X", source_stem: "scrum_reviews", text: "identical text" }),
  ]);
  writeScored("scrum_reviews", [
    makeScored({ run_id: "pi1", task_id: "task-X", category: "accepted", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
    makeScored({ run_id: "pi2", task_id: "task-X", category: "rejected", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
  ]);
  const r = await exportPreference({ root: TMP, recorded_at: NOW });
  expect(r.pairs_exported).toBe(0);
  expect(r.records_quarantined).toBeGreaterThan(0);
  const qPath = resolve(TMP, "exports/quarantine/preference.jsonl");
  expect(readFileSync(qPath, "utf8")).toContain("identical");
});

test("Preference: accepted vs partially_accepted is a softer fallback pair", async () => {
  writeEvidence("scrum_reviews", [
    makeEv({ run_id: "ps1", task_id: "task-X", source_stem: "scrum_reviews", text: "best output" }),
    makeEv({ run_id: "ps2", task_id: "task-X", source_stem: "scrum_reviews", text: "ok output" }),
  ]);
  writeScored("scrum_reviews", [
    makeScored({ run_id: "ps1", task_id: "task-X", category: "accepted", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
    makeScored({ run_id: "ps2", task_id: "task-X", category: "partially_accepted", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` }),
  ]);
  const r = await exportPreference({ root: TMP, recorded_at: NOW });
  expect(r.pairs_exported).toBe(1);
});

// ─── Quarantine populated when expected ─────────────────────────────

test("Quarantine: every export creates exports/quarantine/<exporter>.jsonl when needed", async () => {
  // SFT with a forbidden category should populate quarantine
  writeEvidence("scrum_reviews", [makeEv({ run_id: "q1", task_id: "t1", source_stem: "scrum_reviews", text: "x" })]);
  writeScored("scrum_reviews", [makeScored({ run_id: "q1", task_id: "t1", category: "rejected", reasons: ["x"], out_relpath: `data/scored-runs/${PARTITION}/scrum_reviews.jsonl` })]);
  await exportSft({ root: TMP, recorded_at: NOW });
  expect(existsSync(resolve(TMP, "exports/quarantine/sft.jsonl"))).toBe(true);
});