@tangle-network/agent-eval 0.174.0 → 0.176.0

package/dist/opencode-sqlite-eK6HW6dr.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"opencode-sqlite-eK6HW6dr.js","names":["isRecord"],"sources":["../src/rollout/readers/claude-jsonl.ts","../src/rollout/readers/opencode-sqlite.ts"],"sourcesContent":["/**\n * Backfill reader over Claude Code project transcripts\n * (~/.claude/projects/<cwd-slug>/<sessionId>.jsonl) → canonical\n * chat-with-tools messages plus per-session token usage.\n *\n * Transcript lines consumed: type:\"user\" (string content or content blocks —\n * text + tool_result) and type:\"assistant\" (content blocks — thinking, text,\n * tool_use; message.usage carries tokens). Sidechain lines (isSidechain=true,\n * subagent threads) are separate invocations and are excluded from the main\n * transcript. Everything else (queue-operation, attachment, last-prompt…) is\n * transport metadata, not conversation.\n */\n\nimport { readdir, readFile } from 'node:fs/promises'\nimport { homedir } from 'node:os'\nimport { join } from 'node:path'\nimport type { ChatMessage, ChatToolCall } from '../schema'\n\nexport const DEFAULT_CLAUDE_PROJECTS_DIR = join(homedir(), '.claude', 'projects')\n\n/** Claude Code's project-directory slug for a working directory. */\nexport function claudeProjectSlug(cwd: string): string {\n  return cwd.replace(/[^a-zA-Z0-9-]/g, '-')\n}\n\nexport interface ClaudeTranscriptRef {\n  sessionId: string\n  path: string\n}\n\n/** Transcript files recorded for sessions launched from `cwd`. */\nexport async function findClaudeTranscripts(\n  cwd: string,\n  projectsDir: string = DEFAULT_CLAUDE_PROJECTS_DIR,\n): Promise<ClaudeTranscriptRef[]> {\n  const dir = join(projectsDir, claudeProjectSlug(cwd))\n  const names = await readdir(dir).catch(() => [])\n  return names\n    .filter((n) => n.endsWith('.jsonl'))\n    .sort()\n    .map((n) => ({ sessionId: n.replace(/\\.jsonl$/, ''), path: join(dir, n) }))\n}\n\nexport interface ClaudeUsageTotals {\n  tokensIn: number\n  tokensOut: number\n  cacheRead: number\n  cacheWrite: number\n}\n\nexport interface ClaudeTranscript {\n  messages: ChatMessage[]\n  usage: ClaudeUsageTotals\n  /** Timestamp of the first conversation line; null = empty transcript. */\n  startedAt: string | null\n  endedAt: string | null\n  model: string | null\n}\n\nconst isRecord = (v: unknown): v is Record<string, unknown> =>\n  typeof v === 'object' && v !== null && !Array.isArray(v)\n\n/**\n * One conversation line of a transcript, still in Claude Code's own shape.\n *\n * This is the single line-level parse of the format. `readClaudeTranscript`\n * projects it to canonical messages + usage; the supervision-tree reader\n * (`src/supervisor-run/claude-code-reader.ts`) projects the SAME entries to\n * spawn/settle/steer instants. Two projections, one parser — a second\n * transcript parser is how the two views silently disagree.\n */\nexport interface ClaudeEntry {\n  readonly type: 'user' | 'assistant'\n  /** ISO instant of the line; null when the line carried none. */\n  readonly timestamp: string | null\n  /** The Anthropic message body (`role`, `content`, `model`, `usage`). */\n  readonly message: Record<string, unknown>\n  /** Claude Code's structured tool result, when the line carries one. */\n  readonly toolUseResult: unknown\n  /** True on subagent threads — a separate invocation, not this transcript's turn. */\n  readonly isSidechain: boolean\n  /** Subagent id Claude Code stamps on sidechain lines; null on main-thread lines. */\n  readonly agentId: string | null\n}\n\n/** Parse transcript jsonl text into conversation lines. Non-conversation lines are dropped. */\nexport function parseClaudeEntries(raw: string): ClaudeEntry[] {\n  const out: ClaudeEntry[] = []\n  for (const line of raw.split('\\n')) {\n    if (!line.trim()) continue\n    let entry: Record<string, unknown>\n    try {\n      const parsed: unknown = JSON.parse(line)\n      if (!isRecord(parsed)) continue\n      entry = parsed\n    } catch {\n      continue\n    }\n    if (entry.type !== 'user' && entry.type !== 'assistant') continue\n    const message = entry.message\n    if (!isRecord(message)) continue\n    out.push({\n      type: entry.type,\n      timestamp: typeof entry.timestamp === 'string' ? entry.timestamp : null,\n      message,\n      toolUseResult: entry.toolUseResult,\n      isSidechain: entry.isSidechain === true,\n      agentId: typeof entry.agentId === 'string' ? entry.agentId : null,\n    })\n  }\n  return out\n}\n\nexport interface ReadClaudeTranscriptOptions {\n  /**\n   * Read the sidechain (subagent) thread instead of skipping it. Subagent\n   * transcripts under `<session>/subagents/agent-<id>.jsonl` are sidechain\n   * lines end to end, so their usage is invisible without this.\n   */\n  readonly includeSidechain?: boolean\n}\n\nfunction blockText(content: unknown): string {\n  if (typeof content === 'string') return content\n  if (!Array.isArray(content)) return ''\n  return content\n    .filter(\n      (b): b is Record<string, unknown> =>\n        isRecord(b) && b.type === 'text' && typeof b.text === 'string',\n    )\n    .map((b) => b.text as string)\n    .join('\\n')\n}\n\n/** Parse one transcript jsonl into canonical messages + usage totals. */\nexport async function readClaudeTranscript(\n  path: string,\n  options: ReadClaudeTranscriptOptions = {},\n): Promise<ClaudeTranscript> {\n  return transcriptFromEntries(parseClaudeEntries(await readFile(path, 'utf8')), options)\n}\n\n/** The messages+usage projection of already-parsed entries. */\nexport function transcriptFromEntries(\n  entries: readonly ClaudeEntry[],\n  options: ReadClaudeTranscriptOptions = {},\n): ClaudeTranscript {\n  const wantSidechain = options.includeSidechain === true\n  const messages: ChatMessage[] = []\n  const usage: ClaudeUsageTotals = { tokensIn: 0, tokensOut: 0, cacheRead: 0, cacheWrite: 0 }\n  let startedAt: string | null = null\n  let endedAt: string | null = null\n  let model: string | null = null\n  // Claude Code writes one jsonl line PER CONTENT BLOCK of an API message,\n  // repeating message.id and usage on each — merge blocks into one canonical\n  // assistant turn and count usage once per API message id.\n  let lastAssistantApiId: string | null = null\n  let lastAssistantIndex = -1\n\n  for (const entry of entries) {\n    if (entry.isSidechain !== wantSidechain) continue\n    const message = entry.message\n    if (entry.timestamp !== null) {\n      if (startedAt === null) startedAt = entry.timestamp\n      endedAt = entry.timestamp\n    }\n\n    if (entry.type === 'user') {\n      lastAssistantApiId = null\n      lastAssistantIndex = -1\n      const content = message.content\n      if (typeof content === 'string') {\n        messages.push({ role: 'user', content })\n        continue\n      }\n      if (!Array.isArray(content)) continue\n      // A user line may interleave tool_result blocks (answers to the prior\n      // assistant tool_use) with plain text; preserve order.\n      let userText = ''\n      for (const block of content) {\n        if (!isRecord(block)) continue\n        if (block.type === 'tool_result' && typeof block.tool_use_id === 'string') {\n          messages.push({\n            role: 'tool',\n            tool_call_id: block.tool_use_id,\n            content:\n              blockText(block.content) || (typeof block.content === 'string' ? block.content : ''),\n          })\n        } else if (block.type === 'text' && typeof block.text === 'string') {\n          userText += (userText.length > 0 ? '\\n' : '') + block.text\n        }\n      }\n      if (userText.length > 0) messages.push({ role: 'user', content: userText })\n      continue\n    }\n\n    // assistant\n    if (typeof message.model === 'string') model = message.model\n    const apiId = typeof message.id === 'string' ? message.id : null\n    const continuesTurn = apiId !== null && apiId === lastAssistantApiId && lastAssistantIndex >= 0\n    const msgUsage = message.usage\n    if (isRecord(msgUsage) && !continuesTurn) {\n      usage.tokensIn += typeof msgUsage.input_tokens === 'number' ? msgUsage.input_tokens : 0\n      usage.tokensOut += typeof msgUsage.output_tokens === 'number' ? msgUsage.output_tokens : 0\n      usage.cacheRead +=\n        typeof msgUsage.cache_read_input_tokens === 'number' ? msgUsage.cache_read_input_tokens : 0\n      usage.cacheWrite +=\n        typeof msgUsage.cache_creation_input_tokens === 'number'\n          ? msgUsage.cache_creation_input_tokens\n          : 0\n    }\n    const content = message.content\n    if (!Array.isArray(content)) continue\n    let reasoning = ''\n    let text = ''\n    const toolCalls: ChatToolCall[] = []\n    for (const block of content) {\n      if (!isRecord(block)) continue\n      if (\n        block.type === 'thinking' &&\n        typeof block.thinking === 'string' &&\n        block.thinking.length > 0\n      ) {\n        reasoning += (reasoning.length > 0 ? '\\n' : '') + block.thinking\n      } else if (block.type === 'text' && typeof block.text === 'string') {\n        text += (text.length > 0 ? '\\n' : '') + block.text\n      } else if (block.type === 'tool_use' && typeof block.id === 'string') {\n        toolCalls.push({\n          id: block.id,\n          type: 'function',\n          function: {\n            name: typeof block.name === 'string' ? block.name : 'unknown',\n            arguments: JSON.stringify(block.input ?? {}),\n          },\n        })\n      }\n    }\n    if (reasoning.length === 0 && text.length === 0 && toolCalls.length === 0) continue\n    if (continuesTurn) {\n      const prev = messages[lastAssistantIndex]!\n      if (text.length > 0) prev.content = prev.content === null ? text : `${prev.content}\\n${text}`\n      if (reasoning.length > 0) {\n        prev.reasoning_content =\n          prev.reasoning_content === undefined\n            ? reasoning\n            : `${prev.reasoning_content}\\n${reasoning}`\n      }\n      if (toolCalls.length > 0) prev.tool_calls = [...(prev.tool_calls ?? []), ...toolCalls]\n      continue\n    }\n    messages.push({\n      role: 'assistant',\n      content: text.length > 0 ? text : null,\n      ...(reasoning.length > 0 ? { reasoning_content: reasoning } : {}),\n      ...(toolCalls.length > 0 ? { tool_calls: toolCalls } : {}),\n    })\n    lastAssistantApiId = apiId\n    lastAssistantIndex = messages.length - 1\n  }\n\n  return { messages, usage, startedAt, endedAt, model }\n}\n","/**\n * Read-only backfill reader over the opencode sqlite store\n * (~/.local/share/opencode/opencode.db) → canonical chat-with-tools messages.\n *\n * Schema consumed (observed, 2026-07): `session` rows carry directory /\n * parent_id / agent / model / cost / tokens_*; `message` rows carry a JSON\n * `data` blob ({role, modelID, providerID, tokens, cost, finish}); `part`\n * rows carry the actual content ({type: text|reasoning|tool|step-start|\n * step-finish|snapshot…}). Tool parts hold {callID, state:{input, output,\n * status}} — both the call and its result, which we split into an assistant\n * tool_call plus a role:\"tool\" result message.\n *\n * The store is mutable and can be corrupt (a `.corrupt-bak` sibling ships\n * next to it in the wild), so `openOpencodeDb` returns null instead of\n * throwing — callers record a gap line, never crash the backfill.\n */\n\nimport { createRequire } from 'node:module'\nimport { homedir } from 'node:os'\nimport { join } from 'node:path'\nimport type { DatabaseSync } from 'node:sqlite'\nimport type { ChatMessage, ChatToolCall } from '../schema'\n\nexport const DEFAULT_OPENCODE_DB = join(homedir(), '.local', 'share', 'opencode', 'opencode.db')\n\nexport interface OpencodeSessionRow {\n  id: string\n  parentId: string | null\n  directory: string\n  agent: string | null\n  /** Raw session.model JSON: {id, providerID, variant} where present. */\n  model: { id?: string; providerID?: string } | null\n  costUsd: number\n  tokensInput: number\n  tokensOutput: number\n  tokensReasoning: number\n  tokensCacheRead: number\n  tokensCacheWrite: number\n  timeCreated: number\n  timeUpdated: number\n}\n\n// `node:sqlite` is loaded through CommonJS `require`, not `import()`. esbuild\n// (bundling) and Vite (tests) both rewrite a dynamic import and strip the\n// `node:` prefix under an es20xx target, turning this builtin into a bogus\n// \"sqlite\" package lookup; composing the specifier at runtime does not reliably\n// defeat that (it still resolved through Vite's transform in some workers, so\n// the failure moved around as test files were added). A require obtained from\n// `createRequire` is not an analyzable module reference in either tool, so\n// neither can rewrite it.\n/** Open the store read-only; null = unavailable/corrupt (caller records a gap). */\nexport async function openOpencodeDb(\n  path: string = DEFAULT_OPENCODE_DB,\n): Promise<DatabaseSync | null> {\n  try {\n    // Keep Node-only module initialization inside the Node-only operation.\n    // Root imports are shared with edge consumers that do not define import.meta.url.\n    const nodeRequire = createRequire(import.meta.url)\n    const { DatabaseSync } = nodeRequire('node:sqlite') as typeof import('node:sqlite')\n    const db = new DatabaseSync(path, { readOnly: true })\n    // Probe: a corrupt store can open() fine and fail on first page read.\n    db.prepare('SELECT id FROM session LIMIT 1').get()\n    return db\n  } catch {\n    return null\n  }\n}\n\nconst isRecord = (v: unknown): v is Record<string, unknown> =>\n  typeof v === 'object' && v !== null && !Array.isArray(v)\n\nfunction parseSessionRow(row: Record<string, unknown>): OpencodeSessionRow {\n  let model: OpencodeSessionRow['model'] = null\n  if (typeof row.model === 'string' && row.model.length > 0) {\n    try {\n      const parsed: unknown = JSON.parse(row.model)\n      if (isRecord(parsed)) model = parsed as { id?: string; providerID?: string }\n    } catch {\n      model = null\n    }\n  }\n  return {\n    id: String(row.id),\n    parentId: row.parent_id === null || row.parent_id === undefined ? null : String(row.parent_id),\n    directory: String(row.directory),\n    agent: row.agent === null || row.agent === undefined ? null : String(row.agent),\n    model,\n    costUsd: Number(row.cost ?? 0),\n    tokensInput: Number(row.tokens_input ?? 0),\n    tokensOutput: Number(row.tokens_output ?? 0),\n    tokensReasoning: Number(row.tokens_reasoning ?? 0),\n    tokensCacheRead: Number(row.tokens_cache_read ?? 0),\n    tokensCacheWrite: Number(row.tokens_cache_write ?? 0),\n    timeCreated: Number(row.time_created ?? 0),\n    timeUpdated: Number(row.time_updated ?? 0),\n  }\n}\n\nconst SESSION_COLUMNS =\n  'id, parent_id, directory, agent, model, cost, tokens_input, tokens_output, tokens_reasoning, tokens_cache_read, tokens_cache_write, time_created, time_updated'\n\n/** Sessions whose cwd is `directory` (the worker-clone join key). */\nexport function findOpencodeSessionsByDirectory(\n  db: DatabaseSync,\n  directory: string,\n): OpencodeSessionRow[] {\n  const rows = db\n    .prepare(`SELECT ${SESSION_COLUMNS} FROM session WHERE directory = ? ORDER BY time_created`)\n    .all(directory) as Array<Record<string, unknown>>\n  return rows.map(parseSessionRow)\n}\n\ninterface OpencodePart {\n  type?: string\n  text?: string\n  tool?: string\n  callID?: string\n  state?: { status?: string; input?: unknown; output?: unknown }\n}\n\nfunction toolResultContent(output: unknown): string {\n  if (typeof output === 'string') return output\n  if (output === null || output === undefined) return ''\n  return JSON.stringify(output)\n}\n\n/**\n * Convert one session's message+part rows into canonical messages.\n * An opencode assistant message row spans several model steps; each step's\n * parts (reasoning → text → tool …) become one assistant message followed by\n * the role:\"tool\" results of its calls, preserving order.\n */\nexport function readOpencodeSessionMessages(db: DatabaseSync, sessionId: string): ChatMessage[] {\n  const messageRows = db\n    .prepare('SELECT id, data FROM message WHERE session_id = ? ORDER BY time_created, id')\n    .all(sessionId) as Array<{ id: string; data: string }>\n  const partsStmt = db.prepare('SELECT data FROM part WHERE message_id = ? ORDER BY id')\n\n  const messages: ChatMessage[] = []\n  for (const messageRow of messageRows) {\n    let data: Record<string, unknown>\n    try {\n      const parsed: unknown = JSON.parse(messageRow.data)\n      if (!isRecord(parsed)) continue\n      data = parsed\n    } catch {\n      continue\n    }\n    const parts: OpencodePart[] = []\n    for (const row of partsStmt.all(messageRow.id) as Array<{ data: string }>) {\n      try {\n        const parsed: unknown = JSON.parse(row.data)\n        if (isRecord(parsed)) parts.push(parsed as OpencodePart)\n      } catch {\n        // Malformed part payload: skip the part, keep the message.\n      }\n    }\n\n    if (data.role === 'user') {\n      const text = parts\n        .filter((p) => p.type === 'text' && typeof p.text === 'string')\n        .map((p) => p.text as string)\n        .join('\\n')\n      messages.push({ role: 'user', content: text })\n      continue\n    }\n    if (data.role !== 'assistant') continue\n\n    // Split the row into steps at step-start boundaries; parts before the\n    // first step-start (none observed, but tolerated) form an implicit step.\n    const steps: OpencodePart[][] = []\n    let current: OpencodePart[] = []\n    for (const part of parts) {\n      if (part.type === 'step-start') {\n        if (current.length > 0) steps.push(current)\n        current = []\n        continue\n      }\n      if (part.type === 'step-finish' || part.type === 'snapshot' || part.type === 'patch') continue\n      current.push(part)\n    }\n    if (current.length > 0) steps.push(current)\n\n    for (const step of steps) {\n      const reasoning = step\n        .filter((p) => p.type === 'reasoning' && typeof p.text === 'string' && p.text.length > 0)\n        .map((p) => p.text as string)\n        .join('\\n')\n      const text = step\n        .filter((p) => p.type === 'text' && typeof p.text === 'string')\n        .map((p) => p.text as string)\n        .join('\\n')\n      const toolParts = step.filter((p) => p.type === 'tool' && typeof p.callID === 'string')\n      const toolCalls: ChatToolCall[] = toolParts.map((p) => ({\n        id: p.callID as string,\n        type: 'function',\n        function: {\n          name: p.tool ?? 'unknown',\n          arguments: JSON.stringify(p.state?.input ?? {}),\n        },\n      }))\n      if (reasoning.length === 0 && text.length === 0 && toolCalls.length === 0) continue\n      messages.push({\n        role: 'assistant',\n        content: text.length > 0 ? text : null,\n        ...(reasoning.length > 0 ? { reasoning_content: reasoning } : {}),\n        ...(toolCalls.length > 0 ? { tool_calls: toolCalls } : {}),\n      })\n      for (const p of toolParts) {\n        messages.push({\n          role: 'tool',\n          tool_call_id: p.callID as string,\n          name: p.tool ?? 'unknown',\n          content: toolResultContent(p.state?.output),\n        })\n      }\n    }\n  }\n  return messages\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAkBA,MAAa,8BAA8B,KAAK,QAAQ,GAAG,WAAW,UAAU;;AAGhF,SAAgB,kBAAkB,KAAqB;CACrD,OAAO,IAAI,QAAQ,kBAAkB,GAAG;AAC1C;;AAQA,eAAsB,sBACpB,KACA,cAAsB,6BACU;CAChC,MAAM,MAAM,KAAK,aAAa,kBAAkB,GAAG,CAAC;CAEpD,QAAO,MADa,QAAQ,GAAG,CAAC,CAAC,YAAY,CAAC,CAAC,EAAA,CAE5C,QAAQ,MAAM,EAAE,SAAS,QAAQ,CAAC,CAAC,CACnC,KAAK,CAAC,CACN,KAAK,OAAO;EAAE,WAAW,EAAE,QAAQ,YAAY,EAAE;EAAG,MAAM,KAAK,KAAK,CAAC;CAAE,EAAE;AAC9E;AAkBA,MAAMA,cAAY,MAChB,OAAO,MAAM,YAAY,MAAM,QAAQ,CAAC,MAAM,QAAQ,CAAC;;AA0BzD,SAAgB,mBAAmB,KAA4B;CAC7D,MAAM,MAAqB,CAAC;CAC5B,KAAK,MAAM,QAAQ,IAAI,MAAM,IAAI,GAAG;EAClC,IAAI,CAAC,KAAK,KAAK,GAAG;EAClB,IAAI;EACJ,IAAI;GACF,MAAM,SAAkB,KAAK,MAAM,IAAI;GACvC,IAAI,CAACA,WAAS,MAAM,GAAG;GACvB,QAAQ;EACV,QAAQ;GACN;EACF;EACA,IAAI,MAAM,SAAS,UAAU,MAAM,SAAS,aAAa;EACzD,MAAM,UAAU,MAAM;EACtB,IAAI,CAACA,WAAS,OAAO,GAAG;EACxB,IAAI,KAAK;GACP,MAAM,MAAM;GACZ,WAAW,OAAO,MAAM,cAAc,WAAW,MAAM,YAAY;GACnE;GACA,eAAe,MAAM;GACrB,aAAa,MAAM,gBAAgB;GACnC,SAAS,OAAO,MAAM,YAAY,WAAW,MAAM,UAAU;EAC/D,CAAC;CACH;CACA,OAAO;AACT;AAWA,SAAS,UAAU,SAA0B;CAC3C,IAAI,OAAO,YAAY,UAAU,OAAO;CACxC,IAAI,CAAC,MAAM,QAAQ,OAAO,GAAG,OAAO;CACpC,OAAO,QACJ,QACE,MACCA,WAAS,CAAC,KAAK,EAAE,SAAS,UAAU,OAAO,EAAE,SAAS,QAC1D,CAAC,CACA,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;AACd;;AAGA,eAAsB,qBACpB,MACA,UAAuC,CAAC,GACb;CAC3B,OAAO,sBAAsB,mBAAmB,MAAM,SAAS,MAAM,MAAM,CAAC,GAAG,OAAO;AACxF;;AAGA,SAAgB,sBACd,SACA,UAAuC,CAAC,GACtB;CAClB,MAAM,gBAAgB,QAAQ,qBAAqB;CACnD,MAAM,WAA0B,CAAC;CACjC,MAAM,QAA2B;EAAE,UAAU;EAAG,WAAW;EAAG,WAAW;EAAG,YAAY;CAAE;CAC1F,IAAI,YAA2B;CAC/B,IAAI,UAAyB;CAC7B,IAAI,QAAuB;CAI3B,IAAI,qBAAoC;CACxC,IAAI,qBAAqB;CAEzB,KAAK,MAAM,SAAS,SAAS;EAC3B,IAAI,MAAM,gBAAgB,eAAe;EACzC,MAAM,UAAU,MAAM;EACtB,IAAI,MAAM,cAAc,MAAM;GAC5B,IAAI,cAAc,MAAM,YAAY,MAAM;GAC1C,UAAU,MAAM;EAClB;EAEA,IAAI,MAAM,SAAS,QAAQ;GACzB,qBAAqB;GACrB,qBAAqB;GACrB,MAAM,UAAU,QAAQ;GACxB,IAAI,OAAO,YAAY,UAAU;IAC/B,SAAS,KAAK;KAAE,MAAM;KAAQ;IAAQ,CAAC;IACvC;GACF;GACA,IAAI,CAAC,MAAM,QAAQ,OAAO,GAAG;GAG7B,IAAI,WAAW;GACf,KAAK,MAAM,SAAS,SAAS;IAC3B,IAAI,CAACA,WAAS,KAAK,GAAG;IACtB,IAAI,MAAM,SAAS,iBAAiB,OAAO,MAAM,gBAAgB,UAC/D,SAAS,KAAK;KACZ,MAAM;KACN,cAAc,MAAM;KACpB,SACE,UAAU,MAAM,OAAO,MAAM,OAAO,MAAM,YAAY,WAAW,MAAM,UAAU;IACrF,CAAC;SACI,IAAI,MAAM,SAAS,UAAU,OAAO,MAAM,SAAS,UACxD,aAAa,SAAS,SAAS,IAAI,OAAO,MAAM,MAAM;GAE1D;GACA,IAAI,SAAS,SAAS,GAAG,SAAS,KAAK;IAAE,MAAM;IAAQ,SAAS;GAAS,CAAC;GAC1E;EACF;EAGA,IAAI,OAAO,QAAQ,UAAU,UAAU,QAAQ,QAAQ;EACvD,MAAM,QAAQ,OAAO,QAAQ,OAAO,WAAW,QAAQ,KAAK;EAC5D,MAAM,gBAAgB,UAAU,QAAQ,UAAU,sBAAsB,sBAAsB;EAC9F,MAAM,WAAW,QAAQ;EACzB,IAAIA,WAAS,QAAQ,KAAK,CAAC,eAAe;GACxC,MAAM,YAAY,OAAO,SAAS,iBAAiB,WAAW,SAAS,eAAe;GACtF,MAAM,aAAa,OAAO,SAAS,kBAAkB,WAAW,SAAS,gBAAgB;GACzF,MAAM,aACJ,OAAO,SAAS,4BAA4B,WAAW,SAAS,0BAA0B;GAC5F,MAAM,cACJ,OAAO,SAAS,gCAAgC,WAC5C,SAAS,8BACT;EACR;EACA,MAAM,UAAU,QAAQ;EACxB,IAAI,CAAC,MAAM,QAAQ,OAAO,GAAG;EAC7B,IAAI,YAAY;EAChB,IAAI,OAAO;EACX,MAAM,YAA4B,CAAC;EACnC,KAAK,MAAM,SAAS,SAAS;GAC3B,IAAI,CAACA,WAAS,KAAK,GAAG;GACtB,IACE,MAAM,SAAS,cACf,OAAO,MAAM,aAAa,YAC1B,MAAM,SAAS,SAAS,GAExB,cAAc,UAAU,SAAS,IAAI,OAAO,MAAM,MAAM;QACnD,IAAI,MAAM,SAAS,UAAU,OAAO,MAAM,SAAS,UACxD,SAAS,KAAK,SAAS,IAAI,OAAO,MAAM,MAAM;QACzC,IAAI,MAAM,SAAS,cAAc,OAAO,MAAM,OAAO,UAC1D,UAAU,KAAK;IACb,IAAI,MAAM;IACV,MAAM;IACN,UAAU;KACR,MAAM,OAAO,MAAM,SAAS,WAAW,MAAM,OAAO;KACpD,WAAW,KAAK,UAAU,MAAM,SAAS,CAAC,CAAC;IAC7C;GACF,CAAC;EAEL;EACA,IAAI,UAAU,WAAW,KAAK,KAAK,WAAW,KAAK,UAAU,WAAW,GAAG;EAC3E,IAAI,eAAe;GACjB,MAAM,OAAO,SAAS;GACtB,IAAI,KAAK,SAAS,GAAG,KAAK,UAAU,KAAK,YAAY,OAAO,OAAO,GAAG,KAAK,QAAQ,IAAI;GACvF,IAAI,UAAU,SAAS,GACrB,KAAK,oBACH,KAAK,sBAAsB,KAAA,IACvB,YACA,GAAG,KAAK,kBAAkB,IAAI;GAEtC,IAAI,UAAU,SAAS,GAAG,KAAK,aAAa,CAAC,GAAI,KAAK,cAAc,CAAC,GAAI,GAAG,SAAS;GACrF;EACF;EACA,SAAS,KAAK;GACZ,MAAM;GACN,SAAS,KAAK,SAAS,IAAI,OAAO;GAClC,GAAI,UAAU,SAAS,IAAI,EAAE,mBAAmB,UAAU,IAAI,CAAC;GAC/D,GAAI,UAAU,SAAS,IAAI,EAAE,YAAY,UAAU,IAAI,CAAC;EAC1D,CAAC;EACD,qBAAqB;EACrB,qBAAqB,SAAS,SAAS;CACzC;CAEA,OAAO;EAAE;EAAU;EAAO;EAAW;EAAS;CAAM;AACtD;;;;;;;;;;;;;;;;;;;AC9OA,MAAa,sBAAsB,KAAK,QAAQ,GAAG,UAAU,SAAS,YAAY,aAAa;;AA4B/F,eAAsB,eACpB,OAAe,qBACe;CAC9B,IAAI;EAIF,MAAM,EAAE,iBADY,cAAc,OAAO,KAAK,GACX,CAAC,CAAC,aAAa;EAClD,MAAM,KAAK,IAAI,aAAa,MAAM,EAAE,UAAU,KAAK,CAAC;EAEpD,GAAG,QAAQ,gCAAgC,CAAC,CAAC,IAAI;EACjD,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;AAEA,MAAM,YAAY,MAChB,OAAO,MAAM,YAAY,MAAM,QAAQ,CAAC,MAAM,QAAQ,CAAC;AAEzD,SAAS,gBAAgB,KAAkD;CACzE,IAAI,QAAqC;CACzC,IAAI,OAAO,IAAI,UAAU,YAAY,IAAI,MAAM,SAAS,GACtD,IAAI;EACF,MAAM,SAAkB,KAAK,MAAM,IAAI,KAAK;EAC5C,IAAI,SAAS,MAAM,GAAG,QAAQ;CAChC,QAAQ;EACN,QAAQ;CACV;CAEF,OAAO;EACL,IAAI,OAAO,IAAI,EAAE;EACjB,UAAU,IAAI,cAAc,QAAQ,IAAI,cAAc,KAAA,IAAY,OAAO,OAAO,IAAI,SAAS;EAC7F,WAAW,OAAO,IAAI,SAAS;EAC/B,OAAO,IAAI,UAAU,QAAQ,IAAI,UAAU,KAAA,IAAY,OAAO,OAAO,IAAI,KAAK;EAC9E;EACA,SAAS,OAAO,IAAI,QAAQ,CAAC;EAC7B,aAAa,OAAO,IAAI,gBAAgB,CAAC;EACzC,cAAc,OAAO,IAAI,iBAAiB,CAAC;EAC3C,iBAAiB,OAAO,IAAI,oBAAoB,CAAC;EACjD,iBAAiB,OAAO,IAAI,qBAAqB,CAAC;EAClD,kBAAkB,OAAO,IAAI,sBAAsB,CAAC;EACpD,aAAa,OAAO,IAAI,gBAAgB,CAAC;EACzC,aAAa,OAAO,IAAI,gBAAgB,CAAC;CAC3C;AACF;AAEA,MAAM,kBACJ;;AAGF,SAAgB,gCACd,IACA,WACsB;CAItB,OAHa,GACV,QAAQ,UAAU,gBAAgB,wDAAwD,CAAC,CAC3F,IAAI,SACG,CAAC,CAAC,IAAI,eAAe;AACjC;AAUA,SAAS,kBAAkB,QAAyB;CAClD,IAAI,OAAO,WAAW,UAAU,OAAO;CACvC,IAAI,WAAW,QAAQ,WAAW,KAAA,GAAW,OAAO;CACpD,OAAO,KAAK,UAAU,MAAM;AAC9B;;;;;;;AAQA,SAAgB,4BAA4B,IAAkB,WAAkC;CAC9F,MAAM,cAAc,GACjB,QAAQ,6EAA6E,CAAC,CACtF,IAAI,SAAS;CAChB,MAAM,YAAY,GAAG,QAAQ,wDAAwD;CAErF,MAAM,WAA0B,CAAC;CACjC,KAAK,MAAM,cAAc,aAAa;EACpC,IAAI;EACJ,IAAI;GACF,MAAM,SAAkB,KAAK,MAAM,WAAW,IAAI;GAClD,IAAI,CAAC,SAAS,MAAM,GAAG;GACvB,OAAO;EACT,QAAQ;GACN;EACF;EACA,MAAM,QAAwB,CAAC;EAC/B,KAAK,MAAM,OAAO,UAAU,IAAI,WAAW,EAAE,GAC3C,IAAI;GACF,MAAM,SAAkB,KAAK,MAAM,IAAI,IAAI;GAC3C,IAAI,SAAS,MAAM,GAAG,MAAM,KAAK,MAAsB;EACzD,QAAQ,CAER;EAGF,IAAI,KAAK,SAAS,QAAQ;GACxB,MAAM,OAAO,MACV,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,SAAS,QAAQ,CAAC,CAC9D,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,SAAS,KAAK;IAAE,MAAM;IAAQ,SAAS;GAAK,CAAC;GAC7C;EACF;EACA,IAAI,KAAK,SAAS,aAAa;EAI/B,MAAM,QAA0B,CAAC;EACjC,IAAI,UAA0B,CAAC;EAC/B,KAAK,MAAM,QAAQ,OAAO;GACxB,IAAI,KAAK,SAAS,cAAc;IAC9B,IAAI,QAAQ,SAAS,GAAG,MAAM,KAAK,OAAO;IAC1C,UAAU,CAAC;IACX;GACF;GACA,IAAI,KAAK,SAAS,iBAAiB,KAAK,SAAS,cAAc,KAAK,SAAS,SAAS;GACtF,QAAQ,KAAK,IAAI;EACnB;EACA,IAAI,QAAQ,SAAS,GAAG,MAAM,KAAK,OAAO;EAE1C,KAAK,MAAM,QAAQ,OAAO;GACxB,MAAM,YAAY,KACf,QAAQ,MAAM,EAAE,SAAS,eAAe,OAAO,EAAE,SAAS,YAAY,EAAE,KAAK,SAAS,CAAC,CAAC,CACxF,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,MAAM,OAAO,KACV,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,SAAS,QAAQ,CAAC,CAC9D,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,MAAM,YAAY,KAAK,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,WAAW,QAAQ;GACtF,MAAM,YAA4B,UAAU,KAAK,OAAO;IACtD,IAAI,EAAE;IACN,MAAM;IACN,UAAU;KACR,MAAM,EAAE,QAAQ;KAChB,WAAW,KAAK,UAAU,EAAE,OAAO,SAAS,CAAC,CAAC;IAChD;GACF,EAAE;GACF,IAAI,UAAU,WAAW,KAAK,KAAK,WAAW,KAAK,UAAU,WAAW,GAAG;GAC3E,SAAS,KAAK;IACZ,MAAM;IACN,SAAS,KAAK,SAAS,IAAI,OAAO;IAClC,GAAI,UAAU,SAAS,IAAI,EAAE,mBAAmB,UAAU,IAAI,CAAC;IAC/D,GAAI,UAAU,SAAS,IAAI,EAAE,YAAY,UAAU,IAAI,CAAC;GAC1D,CAAC;GACD,KAAK,MAAM,KAAK,WACd,SAAS,KAAK;IACZ,MAAM;IACN,cAAc,EAAE;IAChB,MAAM,EAAE,QAAQ;IAChB,SAAS,kBAAkB,EAAE,OAAO,MAAM;GAC5C,CAAC;EAEL;CACF;CACA,OAAO;AACT"}

package/dist/power-preflight-CFXm0Vjo.js

@@ -1,502 +0,0 @@
-import { g as pairedRiskDifferenceScore, h as pairedRiskDifferenceExact, p as pairedBinaryScale } from "./paired-arms-D4aeIHUy.js";
-import { a as pairedSignTest, i as pairedDeltaTieFraction, r as pairedBootstrap } from "./paired-tests-C8iCsioC.js";
-//#region src/paired-delta-test.ts
-/** Smallest all-positive sample that can clear a one-sided exact sign test. */
-function minimumPairsForPairedDeltaTest(confidence = .95) {
-	if (!Number.isFinite(confidence) || confidence <= 0 || confidence >= 1) throw new Error(`minimumPairsForPairedDeltaTest: confidence must be in (0,1), got ${confidence}`);
-	const oneSidedAlpha = (1 - confidence) / 2;
-	return Math.ceil(Math.log2(1 / oneSidedAlpha));
-}
-/**
-* Tests whether a paired candidate-minus-baseline delta clears a threshold.
-*
-* At 20 or more pairs, the percentile bootstrap lower bound carries the
-* decision. Below that point the interval is descriptive only, so the function
-* switches to a pre-registered one-sided exact sign test. The exact path is
-* deliberately conservative: it requires both a point estimate above the
-* threshold and enough consistently positive paired differences.
-*
-* ## A zero-width interval is never significant
-*
-* When every paired delta is identical the resample distribution is a point
-* mass and the interval collapses: `[0, 0]` when all pairs tie, `[g, g]` on n
-* identical deltas of g. Neither says the effect is certain — both say the
-* sample carries no information about how far the estimate could be wrong, and
-* `low > threshold` then answers on the point estimate alone. It fails in both
-* directions: `[0, 0]` clears every NEGATIVE threshold, which is how a
-* tie-dominated pass/fail comparison laundered a regression into a
-* noninferiority pass, and `[g, g]` clears every threshold below g with no
-* spread behind it. Under a bounded asymmetric null whose true mean paired
-* delta is exactly 0 — 2 % of pairs dropping by 1.0, the rest gaining 0.0204 —
-* every sample that misses the drop is exactly that shape, and deciding on
-* `low > 0` promoted 65.65 % of samples at n = 20 against a nominal 5 %.
-*
-* So `indeterminate` is reported and `significant` is false whenever the
-* interval has zero width, on BOTH paths: at small n the exact sign test is a
-* test of the MEDIAN and a zero-spread sample is precisely where it stops
-* saying anything about the mean the caller is thresholding.
-*
-* `threshold` may be negative — that is a noninferiority margin, and it is the
-* regime the zero-width hole is worst in. For a two-point (pass/fail) outcome
-* the percentile bootstrap is not a valid interval at a nonzero margin at all;
-* use {@link decidePairedPromotion}, which routes those to Tango's score
-* interval, rather than thresholding this function's bootstrap directly.
-*/
-function pairedDeltaTest(before, after, options = {}) {
-	const threshold = options.threshold ?? 0;
-	if (!Number.isFinite(threshold)) throw new Error(`pairedDeltaTest: threshold must be finite, got ${threshold}`);
-	const exactMinimum = minimumPairsForPairedDeltaTest(options.confidence ?? .95);
-	const requestedMinimum = options.minPairs ?? exactMinimum;
-	if (!Number.isInteger(requestedMinimum) || requestedMinimum < 1) throw new Error(`pairedDeltaTest: minPairs must be a positive integer, got ${requestedMinimum}`);
-	const minimumPairs = Math.max(requestedMinimum, exactMinimum);
-	const bootstrap = pairedBootstrap(before, after, options);
-	const sufficient = bootstrap.n >= minimumPairs;
-	const indeterminate = bootstrap.n > 0 && (!Number.isFinite(bootstrap.low) || !Number.isFinite(bootstrap.high) || bootstrap.low === bootstrap.high);
-	if (bootstrap.gateEligible) return {
-		bootstrap,
-		method: "bootstrap-ci",
-		pValue: null,
-		minimumPairs,
-		sufficient,
-		indeterminate,
-		significant: sufficient && !indeterminate && bootstrap.low > threshold
-	};
-	const exact = pairedSignTest(before.map((value, index) => after[index] - value - threshold), "greater");
-	const estimate = options.statistic === "mean" ? bootstrap.mean : bootstrap.median;
-	return {
-		bootstrap,
-		method: "exact-sign",
-		pValue: exact.pValue,
-		minimumPairs,
-		sufficient,
-		indeterminate,
-		significant: sufficient && !indeterminate && estimate > threshold && exact.pValue <= (1 - bootstrap.confidence) / 2
-	};
-}
-//#endregion
-//#region src/paired-promotion-decision.ts
-/**
-* @module
-* ONE rule for "does this paired interval clear a promotion threshold".
-*
-* The rule below was derived on `HeldOutGate` (#479) after the same estimator
-* bug shipped twice. It then turned out that a SECOND gate — the composable
-* `heldOutGate`, plus everything else routed through `heldoutSignificance` —
-* still carried the original defect, because the rule had been written into one
-* gate's method body rather than into a shared function. Two copies of a
-* statistical rule is how a defect survives in one of them, so there is now
-* exactly one copy and both gates call it.
-*
-* Three things the rule does that a bare `pairedBootstrap(...).low > threshold`
-* does not:
-*
-* 1. **Two-point (pass/fail) outcomes decide on Tango's SCORE interval.** On a
-*    pass/fail eval the paired delta vector is dominated by ties, so the
-*    bootstrap of the mean is a resample of a lattice with three atoms and its
-*    percentile interval is not valid at a nonzero margin. The score interval
-*    (`pairedRiskDifferenceScore`) re-estimates the nuisance loss rate under
-*    each hypothesised margin instead of fixing it at the observed value, which
-*    is the only construction that stays a confidence interval as the margin
-*    moves off zero — the regime every noninferiority threshold lives in.
-*    Measured on the composable gate before this change, at a true risk
-*    difference sitting exactly on the production caller's -0.05 margin and a
-*    nominal 5 %: 14.60 % false promotion at n = 40 and 10.10 % at n = 76.
-* 2. **McNemar's exact test holds a VETO at every non-negative threshold.**
-*    Redundant with the interval by construction and kept anyway, so that
-*    swapping the estimator for one without that duality cannot silently
-*    reintroduce "promotes what the exact test refuses". Witness: n = 6, b = 5,
-*    c = 0 — no exact argument reaches alpha = 0.05 with 5 discordant pairs
-*    (two-sided floor 2/2^5 = 0.0625), whatever an interval says. A NEGATIVE
-*    threshold is a noninferiority question, which McNemar's test of "no
-*    difference" is not the right test for, so the veto does not apply there.
-* 3. **A ZERO-WIDTH interval is refused, wherever it sits.** At [0, 0] it
-*    cannot tell a gain from a regression and clears every negative threshold.
-*    Away from zero it fails the opposite way: n identical positive deltas give
-*    [g, g], which clears threshold 0 on no spread at all. Both are an absence
-*    of evidence. Measured on the composable gate before this change, under a
-*    bounded asymmetric null whose true mean paired delta is exactly 0: 88.50 %
-*    false promotion at n = 6 and 65.65 % at n = 20 against a nominal 5 %.
-*
-* Orthogonal to the small-sample switch inside {@link pairedDeltaTest}: that
-* picks the TEST from the sample size (bootstrap CI at n >= 20, pre-registered
-* exact sign test below it), this picks the ESTIMATOR from the outcome's shape.
-* Both are needed — an exact sign test applied to a tie-pinned median is still
-* blind, and a mean bootstrap CI at n = 6 is still not a valid test.
-*/
-/**
-* Which estimator {@link decidePairedPromotion} would use on this data, and the
-* shape facts behind it — for callers that must report the shape on a path
-* where no interval is computed at all (an early rejection, or zero pairs).
-* Cheap: no bootstrap, no interval.
-*/
-function pairedDecisionShape(before, after, statistic = "mean") {
-	const tieFraction = before.length === 0 ? null : pairedDeltaTieFraction(before, after);
-	if (statistic === "median") return {
-		statistic: "median_bootstrap",
-		binaryScale: null,
-		tieFraction
-	};
-	const binaryScale = pairedBinaryScale(before, after);
-	if (binaryScale !== null) return {
-		statistic: "paired_risk_difference",
-		binaryScale,
-		tieFraction
-	};
-	return {
-		statistic: "mean_bootstrap",
-		binaryScale: null,
-		tieFraction
-	};
-}
-/**
-* Decide whether a paired candidate-minus-baseline delta clears a promotion
-* threshold. `before` is the baseline arm, `after` the candidate arm, paired by
-* position. Throws on unequal lengths.
-*/
-function decidePairedPromotion(before, after, options = {}) {
-	if (before.length !== after.length) throw new Error(`decidePairedPromotion: unequal sample sizes (${before.length} vs ${after.length})`);
-	const threshold = options.threshold ?? 0;
-	if (!Number.isFinite(threshold)) throw new Error(`decidePairedPromotion: threshold must be finite, got ${threshold}`);
-	const confidence = options.confidence ?? .95;
-	const exactMinimum = minimumPairsForPairedDeltaTest(confidence);
-	const requestedMinimum = options.minPairs ?? exactMinimum;
-	if (!Number.isInteger(requestedMinimum) || requestedMinimum < 1) throw new Error(`decidePairedPromotion: minPairs must be a positive integer, got ${requestedMinimum}`);
-	const minimumPairs = Math.max(requestedMinimum, exactMinimum);
-	const n = before.length;
-	const sufficient = n >= minimumPairs;
-	const { binaryScale, tieFraction } = pairedDecisionShape(before, after, options.statistic);
-	let core;
-	if (binaryScale !== null) {
-		const unitControl = before.map((v) => v / binaryScale);
-		const unitTreatment = after.map((v) => v / binaryScale);
-		const exact = pairedRiskDifferenceExact(unitControl, unitTreatment, confidence);
-		const score = pairedRiskDifferenceScore(unitControl, unitTreatment, confidence);
-		const low = score.lower * binaryScale;
-		core = {
-			statistic: "paired_risk_difference",
-			method: "score-interval",
-			delta: score.riskDifference * binaryScale,
-			low,
-			high: score.upper * binaryScale,
-			bootstrap: null,
-			mcnemar: {
-				b: exact.b,
-				c: exact.c,
-				nDiscordant: exact.nDiscordant,
-				pValue: exact.pValue
-			},
-			pValue: null,
-			clearsThreshold: low > threshold,
-			label: "success-rate",
-			methodDetail: ""
-		};
-	} else {
-		const bootstrapStatistic = options.statistic === "median" ? "median" : "mean";
-		const test = pairedDeltaTest(before, after, {
-			confidence,
-			resamples: options.resamples,
-			statistic: bootstrapStatistic,
-			seed: options.seed,
-			threshold,
-			minPairs: options.minPairs
-		});
-		const ci = test.bootstrap;
-		core = {
-			statistic: bootstrapStatistic === "mean" ? "mean_bootstrap" : "median_bootstrap",
-			method: test.method,
-			delta: bootstrapStatistic === "mean" ? ci.mean : ci.median,
-			low: ci.low,
-			high: ci.high,
-			bootstrap: ci,
-			mcnemar: null,
-			pValue: test.pValue,
-			clearsThreshold: test.significant,
-			label: bootstrapStatistic,
-			methodDetail: test.method === "exact-sign" ? ` Below ${test.minimumPairs} pairs the interval is descriptive only; the decision is the exact one-sided sign test, p=${fmt(test.pValue ?? 1)}.` : ""
-		};
-	}
-	const indeterminate = !Number.isFinite(core.low) || !Number.isFinite(core.high) || core.low === core.high;
-	const indeterminateCause = !indeterminate ? "" : tieFraction === 1 ? "every paired delta is an exact tie" : core.mcnemar !== null && core.mcnemar.nDiscordant === 0 ? "every pair is concordant (0 discordant pairs)" : `the ${core.label} CI collapsed to a point at ${fmt(core.low)}`;
-	const exactTestVetoes = core.mcnemar !== null && threshold >= 0 && !(core.mcnemar.pValue < 1 - confidence);
-	return {
-		n,
-		threshold,
-		confidence,
-		binaryScale,
-		tieFraction,
-		minimumPairs,
-		sufficient,
-		indeterminate,
-		indeterminateCause,
-		exactTestVetoes,
-		promote: sufficient && !indeterminate && core.clearsThreshold && !exactTestVetoes,
-		...core
-	};
-}
-function fmt(x) {
-	return x.toFixed(4);
-}
-//#endregion
-//#region src/campaign/gates/statistical-heldout.ts
-/**
-* Statistical held-out promotion machinery — the trustworthy core the
-* point-estimate `heldout-delta` gate lacked.
-*
-* The shipped false positive it prevents: a winner re-scored against the
-* baseline on the holdout read run-to-run model NOISE (e.g. 91 vs 95) as a
-* "+4 lift" and shipped, because the gate compared point estimates with no
-* confidence interval. Here we pair candidate vs baseline holdout observations
-* and bootstrap a CI on the paired delta — a candidate ships only when the CI
-* lower bound clears the effect-size threshold (the gain is real at the
-* confidence level, not noise), and is blocked when a critical dimension
-* (e.g. `hallucination_free` for a legal agent) significantly regresses even if
-* the net composite rose (anti-Goodhart).
-*
-* Two traps this module is built around (both produce a NEW false positive if
-* gotten wrong):
-*   1. PAIRING GRANULARITY — pairs by FULL `cellId` (`scenario:rep`), never by
-*      `scenarioId` (which averages reps away and destroys the within-pair
-*      variance reduction that makes a paired bootstrap tighter than unpaired).
-*      One paired observation per cell ⇒ reps multiply n.
-*   2. SCALE — a judge may emit composites/dimensions on [0,1] or 0-100. The
-*      threshold + tolerance are interpreted in the judge's NATIVE scale; the
-*      per-dimension tolerance auto-scales off the observed baseline magnitudes
-*      so `-0.10` on [0,1] doesn't silently become a no-op on a 0-100 dimension.
-*/
-/** Tie fraction at/above which a gate annotates its verdict with the tie share.
-*  Tie-domination of the median bites structurally at >= 0.5 (the median is then
-*  0 by construction); 0.4 is a softer warn threshold that flags a run APPROACHING
-*  that regime, so an operator sees it before the median goes fully blind. */
-const TIE_WARN_FRACTION = .4;
-/**
-* Pair candidate vs baseline holdout observations by FULL cellId. `select`
-* pulls the scalar from a cell's judge reports (composite, or a named
-* dimension); a cell contributes the mean of `select` across its judges. Cells
-* whose scenario is not in `scenarioIds`, or where `select` is undefined for
-* every judge on either side, are skipped on BOTH sides so the arrays stay
-* paired. Throws when the two maps disagree on which holdout cells exist — a
-* load-bearing invariant: the baseline + winner holdout campaigns run the same
-* scenarios with the same seed base, so their cellIds MUST align; a mismatch
-* means a silent pairing bug, not a soft fallback.
-*/
-function pairHoldout(candidate, baseline, scenarioIds, select) {
-	const cellValue = (byCell, cellId) => {
-		const scores = byCell.get(cellId);
-		if (!scores) return void 0;
-		const vals = [];
-		for (const s of Object.values(scores)) {
-			if (s.failed === true) throw new Error(`pairHoldout: cell '${cellId}' contains a failed judge score`);
-			const v = select(s);
-			if (typeof v === "number" && !Number.isFinite(v)) throw new Error(`pairHoldout: cell '${cellId}' contains a non-finite selected score`);
-			if (typeof v === "number") vals.push(v);
-		}
-		if (vals.length === 0) return void 0;
-		return vals.reduce((a, b) => a + b, 0) / vals.length;
-	};
-	const inScope = (cellId) => scenarioIds.has(cellId.split(":")[0] ?? "");
-	const candCells = [...candidate.keys()].filter(inScope).sort();
-	const baseCells = [...baseline.keys()].filter(inScope).sort();
-	if (candCells.length !== baseCells.length || candCells.some((c, i) => c !== baseCells[i])) throw new Error(`pairHoldout: candidate/baseline holdout cells do not align — candidate=[${candCells.join(",")}] baseline=[${baseCells.join(",")}]. Both holdout campaigns must run the same scenarios with the same seed base.`);
-	const before = [];
-	const after = [];
-	const cellIds = [];
-	for (const cellId of candCells) {
-		const b = cellValue(baseline, cellId);
-		const a = cellValue(candidate, cellId);
-		if (b === void 0 && a === void 0) continue;
-		if (b === void 0 || a === void 0) throw new Error(`pairHoldout: cell '${cellId}' has a selected score on only one arm`);
-		before.push(b);
-		after.push(a);
-		cellIds.push(cellId);
-	}
-	return {
-		before,
-		after,
-		cellIds
-	};
-}
-/**
-* Significance of the held-out composite lift: ship only when the lower bound
-* of the interval the outcome's shape ADMITS exceeds `deltaThreshold` (default
-* 0 ⇒ "confidently positive"). Interpret `deltaThreshold` in the judge's native
-* scale.
-*
-* The decision is delegated whole to {@link decidePairedPromotion}, the one
-* copy of the rule (`src/paired-promotion-decision.ts`), which `HeldOutGate`
-* also calls. That module's header carries the measurements; the short version
-* is three guards a bare `bootstrap.low > threshold` does not have:
-*
-*   - a two-point (pass/fail) outcome decides on Tango's SCORE interval, the
-*     only paired-binary construction that stays valid at a nonzero margin;
-*   - McNemar's exact test VETOES at any non-negative threshold;
-*   - a ZERO-WIDTH interval is refused rather than promoted, in either
-*     direction — [0,0] clears every negative threshold and [g,g] clears every
-*     threshold below g, and both are an absence of evidence, not a result.
-*
-* Measured on this function before those guards landed, at a nominal 5 %:
-* 14.60 % false promotion at n = 40 on a paired-binary noninferiority boundary,
-* and 88.50 % at n = 6 under a bounded asymmetric null whose true mean paired
-* delta is exactly 0.
-*
-* At small n, where the percentile bootstrap is descriptive only, a
-* pre-registered exact sign test still carries the bootstrap path.
-*/
-function heldoutSignificance(paired, opts = {}) {
-	const deltaThreshold = opts.deltaThreshold ?? 0;
-	const confidence = opts.confidence ?? .95;
-	const resamples = opts.resamples ?? 2e3;
-	const seed = opts.seed ?? 1337;
-	const statistic = opts.statistic ?? "mean";
-	const decision = decidePairedPromotion(paired.before, paired.after, {
-		confidence,
-		resamples,
-		statistic,
-		seed,
-		threshold: deltaThreshold,
-		minPairs: opts.minProductiveRuns
-	});
-	const bootstrap = decision.bootstrap ?? pairedBootstrap(paired.before, paired.after, {
-		confidence,
-		resamples,
-		statistic,
-		seed
-	});
-	const medianBootstrap = statistic === "median" ? bootstrap : pairedBootstrap(paired.before, paired.after, {
-		confidence,
-		resamples,
-		statistic: "median",
-		seed
-	});
-	const n = paired.before.length;
-	let ties = 0;
-	for (let i = 0; i < n; i += 1) {
-		const after = paired.after[i] ?? 0;
-		const before = paired.before[i] ?? 0;
-		if (Math.abs(after - before) < 1e-9) ties += 1;
-	}
-	const tieFraction = n === 0 ? 0 : ties / n;
-	return {
-		paired,
-		bootstrap,
-		medianBootstrap,
-		decision,
-		decisionStatistic: decision.statistic,
-		mcnemar: decision.mcnemar,
-		tieFraction,
-		n,
-		minimumRequired: decision.minimumPairs,
-		decisionMethod: decision.method,
-		pValue: decision.pValue,
-		significant: decision.promote,
-		fewRuns: !decision.sufficient
-	};
-}
-/** Detect the native scale of a set of scores: 0-100 when any magnitude clears
-*  1.5, else [0,1]. Used to auto-scale the regression tolerance so a default
-*  expressed for [0,1] is not silently a no-op on a 0-100 dimension. */
-function detectScale(values) {
-	return values.some((v) => Math.abs(v) > 1.5) ? 100 : 1;
-}
-/** Per-critical-dimension regression guard. For each dimension, pair the
-*  candidate vs baseline values by full cellId and bootstrap the paired delta;
-*  a dimension is "regressed" when the CI lower bound < −tolerance (conservative
-*  — blocks if the credible worst case exceeds tolerance, which is the right
-*  posture for safety dimensions like `hallucination_free`). When `tolerance`
-*  is omitted it auto-scales: 0.05 on [0,1], 5 on 0-100.
-*
-*  The interval comes from {@link decidePairedPromotion}, so a pass/fail
-*  dimension is judged on Tango's score interval rather than a percentile
-*  bootstrap of the mean — `tolerance` is a NONZERO margin, and the bootstrap
-*  is not a valid interval at one. That matters most here because this guard
-*  fails OPEN by construction: `tolerance` is positive, so an interval pinned at
-*  [0,0] never satisfies `low < −tolerance` and a real regression on a safety
-*  dimension would be reported as `regressed: false`. On the median it fails the
-*  same way for the same reason — when most pairs tie, which is automatic for a
-*  pass/fail dimension on {0,1} and on the 0-100 encoding `detectScale` exists
-*  to support, the median CI collapses to [0,0]. Pass `statistic: 'median'` to
-*  restore the pre-0.134 behaviour. */
-function dimensionRegressions(candidate, baseline, scenarioIds, criticalDimensions, opts = {}) {
-	const out = [];
-	for (const dim of criticalDimensions) {
-		const paired = pairHoldout(candidate, baseline, scenarioIds, (s) => s.dimensions[dim]);
-		if (paired.before.length === 0) continue;
-		const tolerance = opts.tolerance ?? .05 * detectScale([...paired.before, ...paired.after]);
-		const bootstrapStatistic = opts.statistic ?? "mean";
-		const shared = {
-			confidence: opts.confidence ?? .95,
-			resamples: opts.resamples ?? 2e3,
-			statistic: bootstrapStatistic,
-			seed: opts.seed ?? 1337
-		};
-		const guard = decidePairedPromotion(paired.before, paired.after, shared);
-		const regression = decidePairedPromotion(paired.after, paired.before, {
-			...shared,
-			threshold: tolerance
-		});
-		const bootstrap = guard.bootstrap ?? pairedBootstrap(paired.before, paired.after, shared);
-		out.push({
-			dimension: dim,
-			bootstrap,
-			bootstrapStatistic,
-			ci: {
-				low: guard.low,
-				high: guard.high
-			},
-			decisionStatistic: guard.statistic,
-			mcnemar: guard.mcnemar,
-			indeterminate: guard.indeterminate,
-			regressed: bootstrap.low < -tolerance || regression.promote,
-			tolerance,
-			n: paired.before.length
-		});
-	}
-	return out;
-}
-//#endregion
-//#region src/campaign/gates/power-preflight.ts
-/** Two-sided z for the common confidence levels; interpolation is overkill here. */
-function zFor(confidence) {
-	if (confidence >= .99) return 2.576;
-	if (confidence >= .95) return 1.96;
-	if (confidence >= .9) return 1.645;
-	return 1.282;
-}
-/** Estimate the minimum detectable lift a paired-holdout improvement run can
-*  ship at a given budget, from the baseline holdout composites — call it BEFORE
-*  spending a search to learn whether the effect you are hunting is even
-*  observable at this holdout size and worker variance. */
-function powerPreflight(opts) {
-	const composites = opts.baselineComposites.filter((v) => Number.isFinite(v));
-	if (composites.length < 3) throw new Error(`powerPreflight: need >= 3 finite baseline composites to estimate variance, got ${composites.length}`);
-	const deltaThreshold = opts.deltaThreshold ?? .05;
-	const confidence = opts.confidence ?? .95;
-	const n = opts.pairedN ?? composites.length;
-	if (n < 2) throw new Error(`powerPreflight: pairedN must be >= 2, got ${n}`);
-	const mean = composites.reduce((a, b) => a + b, 0) / composites.length;
-	const variance = composites.reduce((a, b) => a + (b - mean) * (b - mean), 0) / (composites.length - 1);
-	const sd = Math.sqrt(variance);
-	const z = zFor(confidence);
-	const mde = deltaThreshold + z * Math.SQRT2 * sd / Math.sqrt(n);
-	const scaleAssumed = composites.every((v) => v >= -.001 && v <= 1.5);
-	const headroom = Math.max(0, 1 - mean);
-	const underpowered = scaleAssumed && mde > headroom;
-	const sharedChannelCaveat = opts.sharedScorerChannel ? "Holdout and gate share one scoring channel: raising n/reps reduces only idiosyncratic noise — systematic judge bias remains and this MDE is a lower bound. Full debiasing needs an independent second scoring channel (different judge/benchmark family)." : void 0;
-	const recommendation = underpowered ? `UNDERPOWERED: minimum detectable lift ${mde.toFixed(3)} exceeds the ${headroom.toFixed(3)} headroom above the baseline (${mean.toFixed(3)}) — no achievable effect can ship at this budget. Raise paired n (scenarios x reps) to ~${Math.ceil((z * Math.SQRT2 * sd / Math.max(headroom - deltaThreshold, .01)) ** 2)} or reduce worker variance before searching.` : `Minimum detectable lift at n=${n}: ${mde.toFixed(3)} (baseline sd ${sd.toFixed(3)}). Effects smaller than this cannot clear the gate; budget the search for effects you believe exceed it.`;
-	return {
-		n,
-		sd,
-		mde,
-		baselineMean: mean,
-		headroom,
-		underpowered,
-		scaleAssumed,
-		deltaThreshold,
-		confidence,
-		...sharedChannelCaveat ? { sharedChannelCaveat } : {},
-		recommendation: sharedChannelCaveat ? `${recommendation} ${sharedChannelCaveat}` : recommendation
-	};
-}
-//#endregion
-export { heldoutSignificance as a, pairedDecisionShape as c, dimensionRegressions as i, minimumPairsForPairedDeltaTest as l, TIE_WARN_FRACTION as n, pairHoldout as o, detectScale as r, decidePairedPromotion as s, powerPreflight as t, pairedDeltaTest as u };
-
-//# sourceMappingURL=power-preflight-CFXm0Vjo.js.map