@tangle-network/agent-eval 0.133.0 → 0.133.2

package/dist/ledger-core-CPZfcrC2.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"ledger-core-CPZfcrC2.js","names":[],"sources":["../src/ledger-core/atomic-file-lock.ts","../src/ledger-core/canonical.ts","../src/ledger-core/journal-file.ts","../src/ledger-core/trusted-head.ts","../src/ledger-core/journal.ts"],"sourcesContent":["/** Crash-safe filesystem lock shared by campaign persistence and run exclusion. */\n\nimport { randomUUID } from 'node:crypto'\nimport {\n  closeSync,\n  constants,\n  existsSync,\n  fsyncSync,\n  linkSync,\n  openSync,\n  readFileSync,\n  renameSync,\n  unlinkSync,\n  writeFileSync,\n} from 'node:fs'\nimport { hostname } from 'node:os'\n\nexport interface AtomicFileLockOwner {\n  readonly pid: number\n  readonly host: string\n  readonly nonce: string\n}\n\nexport interface AtomicFileLock {\n  readonly owner: AtomicFileLockOwner\n  release(): void\n}\n\nexport type AtomicFileLockUnavailable =\n  | { readonly acquired: false; readonly reason: 'held'; readonly holder: AtomicFileLockOwner }\n  | { readonly acquired: false; readonly reason: 'recovery' }\n\nexport type AtomicFileLockAcquisition =\n  | { readonly acquired: true; readonly lock: AtomicFileLock }\n  | AtomicFileLockUnavailable\n\nexport interface AtomicFileLockOptions {\n  readonly lockPath: string\n  readonly pid?: number\n}\n\nexport class AtomicFileLockError extends Error {\n  override readonly name = 'AtomicFileLockError'\n}\n\ntype OwnerState =\n  | { readonly state: 'missing' }\n  | { readonly state: 'stale'; readonly owner: AtomicFileLockOwner }\n  | { readonly state: 'held'; readonly owner: AtomicFileLockOwner }\n\n/** Report whether another lock path prevents acquisition without modifying it. */\nexport function probeAtomicFileLock(\n  options: AtomicFileLockOptions,\n): AtomicFileLockUnavailable | null {\n  if (existsSync(recoveryPath(options.lockPath))) return { acquired: false, reason: 'recovery' }\n  const state = ownerState(options.lockPath)\n  if (state.state === 'held') {\n    return { acquired: false, reason: 'held', holder: state.owner }\n  }\n  return null\n}\n\n/**\n * Try to acquire a complete, uniquely-owned lock inode.\n *\n * A hard link publishes fully-written owner metadata atomically. Stale-owner\n * removal is serialized so one reclaimer cannot delete a new owner's lock.\n */\nexport function tryAcquireAtomicFileLock(\n  options: AtomicFileLockOptions,\n): AtomicFileLockAcquisition {\n  const pid = options.pid ?? process.pid\n  if (!Number.isSafeInteger(pid) || pid <= 0) {\n    throw new AtomicFileLockError('atomic file lock pid must be a positive integer')\n  }\n  const owner: AtomicFileLockOwner = { pid, host: hostname(), nonce: randomUUID() }\n  for (let attempt = 0; attempt < 8; attempt += 1) {\n    if (existsSync(recoveryPath(options.lockPath))) {\n      return { acquired: false, reason: 'recovery' }\n    }\n    if (tryLinkOwner(options.lockPath, owner, `acquire.${attempt}`)) {\n      return acquired(options.lockPath, owner)\n    }\n\n    const holder = ownerState(options.lockPath)\n    if (holder.state === 'missing') continue\n    if (holder.state === 'held') {\n      return { acquired: false, reason: 'held', holder: holder.owner }\n    }\n\n    const reclaimPath = recoveryPath(options.lockPath)\n    if (!tryLinkOwner(reclaimPath, owner, `reclaim.${attempt}`)) {\n      return { acquired: false, reason: 'recovery' }\n    }\n    try {\n      const current = ownerState(options.lockPath)\n      if (current.state === 'held') {\n        return { acquired: false, reason: 'held', holder: current.owner }\n      }\n      if (current.state === 'stale') {\n        const tombstone = `${options.lockPath}.stale.${owner.nonce}.${attempt}`\n        try {\n          renameSync(options.lockPath, tombstone)\n          unlinkSync(tombstone)\n        } catch (error) {\n          if (!isMissing(error)) throw error\n        }\n      }\n      if (tryLinkOwner(options.lockPath, owner, `recovered.${attempt}`)) {\n        return acquired(options.lockPath, owner)\n      }\n    } finally {\n      releaseOwnedPath(reclaimPath, owner)\n    }\n  }\n\n  throw new AtomicFileLockError(`could not acquire atomic file lock ${options.lockPath}`)\n}\n\nfunction acquired(lockPath: string, owner: AtomicFileLockOwner): AtomicFileLockAcquisition {\n  return {\n    acquired: true,\n    lock: {\n      owner,\n      release: () => releaseOwnedPath(lockPath, owner),\n    },\n  }\n}\n\nfunction tryLinkOwner(lockPath: string, owner: AtomicFileLockOwner, suffix: string): boolean {\n  const ownerPath = `${lockPath}.${owner.pid}.${owner.nonce}.${suffix}.owner`\n  const descriptor = openSync(\n    ownerPath,\n    constants.O_CREAT | constants.O_EXCL | constants.O_WRONLY,\n    0o600,\n  )\n  try {\n    writeFileSync(descriptor, `${canonicalOwner(owner)}\\n`, 'utf8')\n    fsyncSync(descriptor)\n  } catch (error) {\n    closeSync(descriptor)\n    unlinkIfExists(ownerPath)\n    throw error\n  }\n  closeSync(descriptor)\n\n  try {\n    linkSync(ownerPath, lockPath)\n    return true\n  } catch (error) {\n    if (!isAlreadyExists(error)) throw error\n    return false\n  } finally {\n    unlinkIfExists(ownerPath)\n  }\n}\n\nfunction ownerState(lockPath: string): OwnerState {\n  const owner = readOwner(lockPath)\n  if (!owner) return { state: 'missing' }\n  if (owner.host !== hostname()) return { state: 'held', owner }\n  try {\n    process.kill(owner.pid, 0)\n    return { state: 'held', owner }\n  } catch (error) {\n    if (isNoSuchProcess(error)) return { state: 'stale', owner }\n    return { state: 'held', owner }\n  }\n}\n\nfunction readOwner(lockPath: string): AtomicFileLockOwner | undefined {\n  let contents: string\n  try {\n    contents = readFileSync(lockPath, 'utf8').trim()\n  } catch (error) {\n    if (isMissing(error)) return undefined\n    throw error\n  }\n\n  let value: unknown\n  try {\n    value = JSON.parse(contents)\n  } catch (error) {\n    throw invalidOwner(lockPath, error)\n  }\n  if (!isOwner(value)) throw invalidOwner(lockPath)\n  return value\n}\n\nfunction isOwner(value: unknown): value is AtomicFileLockOwner {\n  if (typeof value !== 'object' || value === null || Array.isArray(value)) return false\n  const record = value as Record<string, unknown>\n  return (\n    Object.keys(record).length === 3 &&\n    Number.isSafeInteger(record.pid) &&\n    Number(record.pid) > 0 &&\n    typeof record.host === 'string' &&\n    record.host.trim().length > 0 &&\n    typeof record.nonce === 'string' &&\n    record.nonce.trim().length > 0\n  )\n}\n\nfunction invalidOwner(lockPath: string, cause?: unknown): AtomicFileLockError {\n  const error = new AtomicFileLockError(\n    `atomic file lock has an invalid owner (${lockPath}); refusing unsafe recovery`,\n  )\n  if (cause !== undefined) Object.defineProperty(error, 'cause', { value: cause })\n  return error\n}\n\nfunction releaseOwnedPath(lockPath: string, owner: AtomicFileLockOwner): void {\n  const current = readOwner(lockPath)\n  if (!current) return\n  if (canonicalOwner(current) !== canonicalOwner(owner)) {\n    throw new AtomicFileLockError(`atomic file lock owner changed before release (${lockPath})`)\n  }\n  try {\n    unlinkSync(lockPath)\n  } catch (error) {\n    if (!isMissing(error)) throw error\n  }\n}\n\nfunction canonicalOwner(owner: AtomicFileLockOwner): string {\n  return JSON.stringify({ host: owner.host, nonce: owner.nonce, pid: owner.pid })\n}\n\nfunction recoveryPath(lockPath: string): string {\n  return `${lockPath}.reclaim`\n}\n\nfunction isAlreadyExists(error: unknown): boolean {\n  return error instanceof Error && 'code' in error && error.code === 'EEXIST'\n}\n\nfunction isMissing(error: unknown): boolean {\n  return error instanceof Error && 'code' in error && error.code === 'ENOENT'\n}\n\nfunction isNoSuchProcess(error: unknown): boolean {\n  return error instanceof Error && 'code' in error && error.code === 'ESRCH'\n}\n\nfunction unlinkIfExists(path: string): void {\n  try {\n    unlinkSync(path)\n  } catch (error) {\n    if (!isMissing(error)) throw error\n  }\n}\n","/**\n * Canonical bytes and digests for durable journal rows.\n *\n * Serialization is RFC 8785 through `@tangle-network/agent-interface` — the\n * stack's single identity scheme — so a value carries the same digest in every\n * package that produces or verifies one.\n *\n * Values canonical JSON cannot represent faithfully are refused, never coerced.\n * `NaN`, `Infinity`, `undefined`, class instances, cycles, and unpaired\n * surrogates each have a lossy encoding that maps distinct records onto one\n * digest, and in a hash chain the digest IS the record's identity: a coercion\n * there is a collision, not a convenience.\n */\n\nimport {\n  canonicalCandidateBytes,\n  canonicalCandidateJson,\n  sha256Bytes,\n} from '@tangle-network/agent-interface'\nimport { ValidationError } from '../errors'\n\nexport type LedgerHash = `sha256:${string}`\n\n/** Shape of every ledger digest: `sha256:` followed by 64 lowercase hex chars. */\nexport const LEDGER_HASH_PATTERN = /^sha256:[a-f0-9]{64}$/\n\n/** A value has no faithful canonical-JSON form, so it cannot be stored or hashed. */\nexport class LedgerCanonicalizationError extends ValidationError {}\n\n/** Canonical JSON encoding (RFC 8785): object keys sorted by UTF-16 code unit,\n * recursively. This is the byte form that is hashed and stored, so equal values\n * always encode identically. */\nexport function canonicalString(value: unknown): string {\n  try {\n    return canonicalCandidateJson(value)\n  } catch (error) {\n    throw canonicalizationError(value, error)\n  }\n}\n\nexport function hashCanonical(value: unknown): LedgerHash {\n  try {\n    return sha256Bytes(canonicalCandidateBytes(value))\n  } catch (error) {\n    throw canonicalizationError(value, error)\n  }\n}\n\n/** Name the offending value and path so a refusal is actionable. The RFC 8785\n * encoder above is the sole authority on acceptance; this walk only explains a\n * rejection it already made. */\nfunction canonicalizationError(value: unknown, cause: unknown): LedgerCanonicalizationError {\n  const defect = firstNonCanonicalDefect(value, '$', new Set())\n  const reason = defect !== '' ? defect : cause instanceof Error ? cause.message : String(cause)\n  return new LedgerCanonicalizationError(`value has no canonical JSON form: ${reason}`, {\n    cause,\n  })\n}\n\nfunction firstNonCanonicalDefect(value: unknown, path: string, ancestors: Set<object>): string {\n  if (value === null) return ''\n  if (value === undefined) return `${path} is undefined`\n  const type = typeof value\n  if (type === 'number') {\n    return Number.isFinite(value) ? '' : `${path} is ${String(value)}`\n  }\n  if (type === 'string') {\n    return isWellFormedUtf16(value as string) ? '' : `${path} contains an unpaired surrogate`\n  }\n  if (type === 'boolean') return ''\n  if (type !== 'object') return `${path} is a ${type}`\n  const object = value as object\n  if (ancestors.has(object)) return `${path} closes a reference cycle`\n  const prototype = Object.getPrototypeOf(object)\n  if (!Array.isArray(object) && prototype !== Object.prototype && prototype !== null) {\n    return `${path} is a ${object.constructor?.name ?? 'non-plain'} instance, not plain JSON data`\n  }\n  const nested = new Set(ancestors).add(object)\n  if (Array.isArray(object)) {\n    for (let index = 0; index < object.length; index += 1) {\n      const defect = firstNonCanonicalDefect(object[index], `${path}[${index}]`, nested)\n      if (defect !== '') return defect\n    }\n    return ''\n  }\n  for (const [key, entry] of Object.entries(object)) {\n    if (!isWellFormedUtf16(key)) return `${path}.${key} has a key with an unpaired surrogate`\n    const defect = firstNonCanonicalDefect(entry, `${path}.${key}`, nested)\n    if (defect !== '') return defect\n  }\n  return ''\n}\n\nfunction isWellFormedUtf16(value: string): boolean {\n  for (let index = 0; index < value.length; index += 1) {\n    const code = value.charCodeAt(index)\n    if (code >= 0xd800 && code <= 0xdbff) {\n      const next = value.charCodeAt(index + 1)\n      if (!(next >= 0xdc00 && next <= 0xdfff)) return false\n      index += 1\n    } else if (code >= 0xdc00 && code <= 0xdfff) {\n      return false\n    }\n  }\n  return true\n}\n","/** Filesystem durability and cross-process exclusion for append-only journal files. */\n\nimport {\n  closeSync,\n  constants,\n  fsyncSync,\n  mkdirSync,\n  openSync,\n  renameSync,\n  writeSync,\n} from 'node:fs'\nimport { dirname } from 'node:path'\nimport { AtomicFileLockError, tryAcquireAtomicFileLock } from './atomic-file-lock'\n\n/** How a journal binding reports file-layer faults in its own error taxonomy. */\nexport interface LedgerFileContext {\n  /** Noun used in error messages, e.g. 'search ledger'. */\n  subject: string\n  integrityError(message: string, options?: { cause?: unknown }): Error\n}\n\n/** Append one already-serialized line; the write is fsynced (file and directory)\n * before returning so an acknowledged append survives a crash. */\nexport function appendLedgerLine(path: string, line: string, context: LedgerFileContext): void {\n  mkdirSync(dirname(path), { recursive: true })\n  const fd = openSync(path, constants.O_CREAT | constants.O_WRONLY | constants.O_APPEND, 0o600)\n  try {\n    writeAll(fd, Buffer.from(line, 'utf8'), context)\n    fsyncSync(fd)\n  } finally {\n    closeSync(fd)\n  }\n  fsyncDirectory(dirname(path))\n}\n\n/** Replace a small sidecar file whole: write a temporary sibling, fsync it,\n * then rename over the target. A reader therefore sees the previous contents\n * or the new ones, never a partial file. */\nexport function writeLedgerFileAtomically(\n  path: string,\n  contents: string,\n  context: LedgerFileContext,\n): void {\n  const directory = dirname(path)\n  mkdirSync(directory, { recursive: true })\n  const temporaryPath = `${path}.tmp`\n  const fd = openSync(\n    temporaryPath,\n    constants.O_CREAT | constants.O_WRONLY | constants.O_TRUNC,\n    0o600,\n  )\n  try {\n    writeAll(fd, Buffer.from(contents, 'utf8'), context)\n    fsyncSync(fd)\n  } finally {\n    closeSync(fd)\n  }\n  renameSync(temporaryPath, path)\n  fsyncDirectory(directory)\n}\n\nexport function withLedgerFileLock<T>(path: string, context: LedgerFileContext, run: () => T): T {\n  const result = tryWithLedgerFileLock(path, context, run)\n  if (!result.acquired) {\n    throw context.integrityError(`${context.subject} lock is held (${path})`)\n  }\n  return result.value\n}\n\nexport type FileLockResult<T> = { acquired: true; value: T } | { acquired: false }\n\nexport function tryWithLedgerFileLock<T>(\n  path: string,\n  context: LedgerFileContext,\n  run: () => T,\n): FileLockResult<T> {\n  mkdirSync(dirname(path), { recursive: true })\n  const lockPath = `${path}.lock`\n  try {\n    const acquisition = tryAcquireAtomicFileLock({ lockPath })\n    if (!acquisition.acquired) return { acquired: false }\n    try {\n      return { acquired: true, value: run() }\n    } finally {\n      acquisition.lock.release()\n    }\n  } catch (error) {\n    if (error instanceof AtomicFileLockError) {\n      throw context.integrityError(error.message, { cause: error })\n    }\n    throw error\n  }\n}\n\nfunction writeAll(fd: number, bytes: Buffer, context: LedgerFileContext): void {\n  let offset = 0\n  while (offset < bytes.byteLength) {\n    const written = writeSync(fd, bytes, offset, bytes.byteLength - offset)\n    if (written <= 0) throw context.integrityError('filesystem wrote zero bytes')\n    offset += written\n  }\n}\n\nfunction fsyncDirectory(path: string): void {\n  const fd = openSync(path, constants.O_RDONLY)\n  try {\n    fsyncSync(fd)\n  } finally {\n    closeSync(fd)\n  }\n}\n","/**\n * Trusted head — the pin that makes deletion detectable.\n *\n * ## What the chain binds\n *\n * Every entry hashes the journal header, its sequence, the previous entry's\n * hash, and its event, so editing, reordering, or removing an entry in the\n * middle of a journal invalidates every hash after it. What a hash chain\n * cannot bind is its own length. Dropping entries off the END leaves a shorter\n * journal that is internally perfect, and rebuilding a journal from sequence 0\n * with recomputed hashes produces a different, equally perfect chain. For an\n * audit log of candidate registrations and promotion decisions that is the\n * difference between a record of a search that happened and a record of one\n * that did not.\n *\n * ## What the pin binds\n *\n * `<journal>.head` holds `{sequence, entryHash}` for one entry a writer\n * actually appended, in a file separate from the journal it attests. Each\n * entry hash commits to the entire prefix before it, so checking that the\n * entry at exactly `sequence` still carries `entryHash` proves two things at\n * once: the journal is at least `sequence + 1` entries long, and those entries\n * are byte-for-byte the ones the writer saw. A journal that merely grew past\n * the pin still verifies; a truncated, rolled-back, or wholesale-rewritten one\n * cannot.\n *\n * ## What the pin does NOT defend against\n *\n * The pin is a second file, not a signature. A writer with access to BOTH the\n * journal and its pin can rewrite both consistently, and the result verifies.\n * Pinning raises the cost of undetected tampering from one file to two and\n * nothing more; defending against a writer who holds both requires an anchor\n * outside this store entirely — a signed head, or a digest published to an\n * append-only service the tamperer does not control.\n *\n * Deleting the pin file downgrades a pinned journal to an unpinned one, which\n * cannot be distinguished from a journal written before pinning existed. That\n * is what `requireTrustedHead` is for: a journal opened under it refuses to\n * read at all once a non-empty journal has no pin.\n */\n\nimport { existsSync, readFileSync } from 'node:fs'\nimport { canonicalString, LEDGER_HASH_PATTERN, type LedgerHash } from './canonical'\nimport { type LedgerFileContext, writeLedgerFileAtomically } from './journal-file'\n\n/** A `(sequence, entryHash)` pair a writer pinned outside the journal. */\nexport interface LedgerTrustedHead {\n  sequence: number\n  entryHash: LedgerHash\n}\n\n/** Minimum entry shape the anchor check needs. */\nexport interface LedgerAnchoredEntry {\n  sequence: number\n  entryHash: LedgerHash\n}\n\n/** Sibling pin file for a journal — deliberately NOT the journal, so rewriting\n * the log is not enough to rewrite the trust recorded about it. */\nexport function trustedHeadPathFor(journalPath: string): string {\n  return `${journalPath}.head`\n}\n\n/** The pin, or null when this journal has never been pinned. A pin that exists\n * but does not parse is corruption or tamper of the trust record itself and\n * fails loudly: reporting it as \"no pin\" would silently downgrade what the\n * journal can prove. */\nexport function readTrustedHeadFile(\n  path: string,\n  context: LedgerFileContext,\n): LedgerTrustedHead | null {\n  if (!existsSync(path)) return null\n  const text = readFileSync(path, 'utf8')\n  let raw: unknown\n  try {\n    raw = JSON.parse(text)\n  } catch (error) {\n    throw context.integrityError(`${context.subject} trusted head ${path} is not valid JSON`, {\n      cause: error,\n    })\n  }\n  if (typeof raw !== 'object' || raw === null || Array.isArray(raw)) {\n    throw context.integrityError(`${context.subject} trusted head ${path} is not an object`)\n  }\n  const record = raw as Record<string, unknown>\n  const keys = Object.keys(record).sort()\n  if (keys.length !== 2 || keys[0] !== 'entryHash' || keys[1] !== 'sequence') {\n    throw context.integrityError(\n      `${context.subject} trusted head ${path} has keys [${keys.join(', ')}], expected [entryHash, sequence]`,\n    )\n  }\n  const { sequence, entryHash } = record\n  if (typeof sequence !== 'number' || !Number.isSafeInteger(sequence) || sequence < 0) {\n    throw context.integrityError(\n      `${context.subject} trusted head ${path} sequence is not a non-negative integer`,\n    )\n  }\n  if (typeof entryHash !== 'string' || !LEDGER_HASH_PATTERN.test(entryHash)) {\n    throw context.integrityError(\n      `${context.subject} trusted head ${path} entryHash is not a sha256 digest`,\n    )\n  }\n  return { sequence, entryHash: entryHash as LedgerHash }\n}\n\n/** Publish a pin by write-then-rename, so a crash leaves either the previous\n * pin or the new one and never a torn file that would lock a journal out of\n * its own trust record. */\nexport function writeTrustedHeadFile(\n  path: string,\n  head: LedgerTrustedHead,\n  context: LedgerFileContext,\n): void {\n  writeLedgerFileAtomically(path, `${canonicalString(head)}\\n`, context)\n}\n\n/** The entry at exactly `head.sequence` must still carry `head.entryHash`.\n * Entries must already have passed chain verification, which is what makes\n * index and sequence interchangeable here. */\nexport function verifyEntriesAgainstTrustedHead(\n  entries: readonly LedgerAnchoredEntry[],\n  head: LedgerTrustedHead,\n  context: LedgerFileContext,\n  subject: string,\n): void {\n  const pinned = entries[head.sequence]\n  if (pinned === undefined) {\n    throw context.integrityError(\n      `${subject} trusted head pins sequence ${head.sequence} (${head.entryHash}) but the journal has ${entries.length} entries — pinned history is missing (truncation, rollback, or rewrite)`,\n    )\n  }\n  if (pinned.sequence !== head.sequence) {\n    throw context.integrityError(\n      `${subject} entry at index ${head.sequence} carries sequence ${pinned.sequence}; the chain was not verified before the trusted head`,\n    )\n  }\n  if (pinned.entryHash !== head.entryHash) {\n    throw context.integrityError(\n      `${subject} entry at pinned sequence ${head.sequence} carries ${pinned.entryHash} but the trusted head recorded ${head.entryHash} — the journal was rewritten`,\n    )\n  }\n}\n","/**\n * Generic durable append-only journal with a SHA-256 hash chain.\n *\n * The file format is canonical JSONL: every row is the canonical JSON encoding\n * (RFC 8785, object keys sorted recursively) of one entry, and every entry\n * hashes the previous entry's hash so the log is tamper-evident. Appends are\n * serialized in-process per path and across processes via a lock file, fsynced\n * before acknowledgement, and idempotent by `eventId`. A malformed,\n * non-canonical, truncated, reordered, or conflicting log fails loudly; a bad\n * row is never skipped.\n *\n * A hash chain binds each entry to its predecessor, which cannot prove that no\n * entry was removed from the end: a valid shorter prefix is still a valid\n * chain. `pinHead` records the appended entry in a sibling `<journal>.head`\n * file and every read re-verifies the journal against a pin that is present,\n * so truncation and wholesale rewrite are refused. `trusted-head.ts` carries\n * the full threat model, including what a pin does not defend against.\n *\n * Domain vocabulary lives entirely in the consumer's codec: entry schema\n * validation, the constant header fields stamped into every entry, the error\n * taxonomy, and the state machine replayed over verified entries.\n */\n\nimport { existsSync, readFileSync } from 'node:fs'\nimport { resolve } from 'node:path'\nimport { Mutex } from '../concurrency'\nimport { canonicalString, hashCanonical, type LedgerHash } from './canonical'\nimport { appendLedgerLine, type LedgerFileContext, withLedgerFileLock } from './journal-file'\nimport {\n  type LedgerTrustedHead,\n  readTrustedHeadFile,\n  trustedHeadPathFor,\n  verifyEntriesAgainstTrustedHead,\n  writeTrustedHeadFile,\n} from './trusted-head'\n\n/** Minimum shape of a journal event: the idempotency key. */\nexport interface LedgerEventBase {\n  eventId: string\n}\n\n/** Chain envelope every entry carries in addition to the journal's constant\n * header fields. `entryHash` covers the header, sequence, previous hash, and\n * event, so any rewrite of history invalidates every later entry. */\nexport interface LedgerChainFields<Event extends LedgerEventBase> {\n  sequence: number\n  previousHash: LedgerHash | null\n  event: Event\n  entryHash: LedgerHash\n}\n\nexport type LedgerEntryOf<Header extends object, Event extends LedgerEventBase> = Header &\n  LedgerChainFields<Event>\n\nexport interface LedgerLineContext {\n  path: string\n  line: number\n}\n\n/** Domain state machine replayed over chain-verified entries in order.\n * `apply` throws (via the codec's error taxonomy) on an invalid transition;\n * `finish` builds the consumer's projection once every entry was applied. */\nexport interface LedgerProjector<Entry, Projection> {\n  apply(entry: Entry, index: number): void\n  finish(entries: Entry[]): Projection\n}\n\nexport interface LedgerJournalCodec<\n  Header extends object,\n  Event extends LedgerEventBase,\n  Projection,\n> extends LedgerFileContext {\n  /** Constant fields stamped into every entry and covered by its hash. */\n  header: Header\n  conflictError(message: string): Error\n  /** Validate one parsed JSON row into a typed entry. Must reject malformed\n   * rows and non-canonical event encodings with the binding's own errors;\n   * the journal itself verifies the stored bytes and the hash chain. */\n  parseEntry(raw: unknown, context: LedgerLineContext): LedgerEntryOf<Header, Event>\n  /** Reject an entry whose constant header fields do not match this journal. */\n  checkEntryHeader(entry: LedgerEntryOf<Header, Event>, index: number): void\n  createProjector(): LedgerProjector<LedgerEntryOf<Header, Event>, Projection>\n}\n\nexport interface LedgerReplayResult<Entry, Projection> {\n  entries: Entry[]\n  projection: Projection\n}\n\nexport interface LedgerAppendResult<Entry, Projection> {\n  entry: Entry\n  /** False when the exact event was already durably present. */\n  appended: boolean\n  projection: Projection\n}\n\nexport interface LedgerAppendOptions {\n  /** Record the appended entry as this journal's trusted head, so a later read\n   * can prove nothing was deleted from the end. Ignored on the idempotent\n   * path: acknowledging an event that was already durable writes nothing, and\n   * moving the pin there would jump trust forward past entries this caller\n   * never saw. */\n  pinHead?: boolean\n}\n\nexport interface FileLedgerJournalOptions {\n  /** Refuse to read a non-empty journal that has no trusted head.\n   *\n   * Off by default because journals written before pinning have no pin and\n   * must keep opening. Turn it on for a journal this process pins: without it,\n   * deleting the sibling pin file silently downgrades the journal back to a\n   * chain that cannot detect deletion. */\n  requireTrustedHead?: boolean\n}\n\n// One async mutex per resolved journal path so concurrent appends from a single\n// process queue up instead of failing on the held cross-process lock file.\nconst journalMutexes = new Map<string, Mutex>()\n\nfunction mutexFor(path: string): Mutex {\n  const existing = journalMutexes.get(path)\n  if (existing) return existing\n  const mutex = new Mutex()\n  journalMutexes.set(path, mutex)\n  return mutex\n}\n\n/** Durable filesystem journal. Construction performs no I/O; `append` and\n * `replay` validate the complete existing file under the locks. */\nexport class FileLedgerJournal<Header extends object, Event extends LedgerEventBase, Projection> {\n  readonly path: string\n  /** Sibling file holding this journal's trusted head. */\n  readonly trustedHeadPath: string\n  private readonly codec: LedgerJournalCodec<Header, Event, Projection>\n  private readonly mutex: Mutex\n  private readonly requireTrustedHead: boolean\n\n  constructor(\n    path: string,\n    codec: LedgerJournalCodec<Header, Event, Projection>,\n    options: FileLedgerJournalOptions = {},\n  ) {\n    this.path = resolve(path)\n    this.trustedHeadPath = trustedHeadPathFor(this.path)\n    this.codec = codec\n    this.mutex = mutexFor(this.path)\n    this.requireTrustedHead = options.requireTrustedHead === true\n  }\n\n  async replay(): Promise<LedgerReplayResult<LedgerEntryOf<Header, Event>, Projection>> {\n    return this.mutex.runExclusive(() =>\n      withLedgerFileLock(this.path, this.codec, () => this.replayLocked()),\n    )\n  }\n\n  async append(\n    event: Event,\n    options: LedgerAppendOptions = {},\n  ): Promise<LedgerAppendResult<LedgerEntryOf<Header, Event>, Projection>> {\n    return this.mutex.runExclusive(() =>\n      withLedgerFileLock(this.path, this.codec, () => {\n        const before = this.replayLocked()\n        const existing = before.entries.find((entry) => entry.event.eventId === event.eventId)\n        if (existing) {\n          if (canonicalString(existing.event) !== canonicalString(event)) {\n            throw this.codec.conflictError(\n              `eventId ${event.eventId} already exists with different content`,\n            )\n          }\n          return { entry: existing, appended: false, projection: before.projection }\n        }\n\n        const material = {\n          ...this.codec.header,\n          sequence: before.entries.length,\n          previousHash: before.entries.at(-1)?.entryHash ?? null,\n          event,\n        }\n        const entry = { ...material, entryHash: hashCanonical(material) }\n\n        // Apply the state transition before spending an append: an entry the\n        // projector rejects must never reach the durable file.\n        const projection = this.project([...before.entries, entry])\n        appendLedgerLine(this.path, `${canonicalString(entry)}\\n`, this.codec)\n        if (options.pinHead === true) {\n          // Journal first, pin second: a crash between them leaves the pin one\n          // entry behind, which still verifies and is healed by the next\n          // pinning append. The reverse order would leave a pin naming an\n          // entry the journal does not carry, locking the journal out.\n          writeTrustedHeadFile(\n            this.trustedHeadPath,\n            { sequence: entry.sequence, entryHash: entry.entryHash },\n            this.codec,\n          )\n        }\n        return { entry, appended: true, projection }\n      }),\n    )\n  }\n\n  /** The pinned head, or null when this journal has never been pinned. */\n  async trustedHead(): Promise<LedgerTrustedHead | null> {\n    return this.mutex.runExclusive(() =>\n      withLedgerFileLock(this.path, this.codec, () =>\n        readTrustedHeadFile(this.trustedHeadPath, this.codec),\n      ),\n    )\n  }\n\n  /** Pin the current verified head. The chain and any existing pin are checked\n   * first, so the pin only ever moves forward along a journal that still\n   * carries the history it already recorded. */\n  async pinTrustedHead(): Promise<LedgerTrustedHead> {\n    return this.mutex.runExclusive(() =>\n      withLedgerFileLock(this.path, this.codec, () => {\n        const entries = this.readEntries()\n        this.project(entries)\n        // `requireTrustedHead` is deliberately not applied here: pinning is how\n        // a journal that lacks a pin acquires one.\n        this.verifyTrustedHead(entries, false)\n        const last = entries.at(-1)\n        if (last === undefined) {\n          throw this.codec.integrityError(\n            `${this.codec.subject} ${this.path} is empty; there is no head to pin`,\n          )\n        }\n        const head: LedgerTrustedHead = { sequence: last.sequence, entryHash: last.entryHash }\n        writeTrustedHeadFile(this.trustedHeadPath, head, this.codec)\n        return head\n      }),\n    )\n  }\n\n  private replayLocked(): LedgerReplayResult<LedgerEntryOf<Header, Event>, Projection> {\n    const entries = this.readEntries()\n    const projection = this.project(entries)\n    // The anchor is meaningful only over a chain that already verified, so it\n    // is checked last — and on every read path, so a violated pin is refused\n    // before an append can extend the journal over it.\n    this.verifyTrustedHead(entries, this.requireTrustedHead)\n    return { entries, projection }\n  }\n\n  private verifyTrustedHead(entries: LedgerEntryOf<Header, Event>[], requirePin: boolean): void {\n    const pinned = readTrustedHeadFile(this.trustedHeadPath, this.codec)\n    if (pinned === null) {\n      if (requirePin && entries.length > 0) {\n        throw this.codec.integrityError(\n          `${this.codec.subject} ${this.path} has ${entries.length} entries but no trusted head at ${this.trustedHeadPath} — without its pin the journal cannot prove nothing was deleted`,\n        )\n      }\n      return\n    }\n    verifyEntriesAgainstTrustedHead(\n      entries,\n      pinned,\n      this.codec,\n      `${this.codec.subject} ${this.path}`,\n    )\n  }\n\n  private readEntries(): LedgerEntryOf<Header, Event>[] {\n    if (!existsSync(this.path)) return []\n    const text = readFileSync(this.path, 'utf8')\n    if (text.length === 0) return []\n    if (!text.endsWith('\\n')) {\n      throw this.codec.integrityError(\n        `${this.codec.subject} ${this.path} has a truncated final record (missing newline)`,\n      )\n    }\n\n    const lines = text.slice(0, -1).split('\\n')\n    const entries: LedgerEntryOf<Header, Event>[] = []\n    for (let index = 0; index < lines.length; index += 1) {\n      const line = lines[index]!\n      if (line.length === 0) {\n        throw this.codec.integrityError(\n          `${this.codec.subject} ${this.path} has a blank row at line ${index + 1}`,\n        )\n      }\n      let raw: unknown\n      try {\n        raw = JSON.parse(line)\n      } catch (error) {\n        throw this.codec.integrityError(\n          `${this.codec.subject} ${this.path} has invalid JSON at line ${index + 1}`,\n          { cause: error },\n        )\n      }\n      const entry = this.codec.parseEntry(raw, { path: this.path, line: index + 1 })\n      let canonical: string\n      try {\n        canonical = canonicalString(entry)\n      } catch (error) {\n        throw this.codec.integrityError(\n          `${this.codec.subject} ${this.path} has a row with no canonical JSON form at line ${index + 1}`,\n          { cause: error },\n        )\n      }\n      if (line !== canonical) {\n        throw this.codec.integrityError(\n          `${this.codec.subject} ${this.path} has non-canonical bytes at line ${index + 1}`,\n        )\n      }\n      entries.push(entry)\n    }\n    return entries\n  }\n\n  private project(entries: LedgerEntryOf<Header, Event>[]): Projection {\n    const projector = this.codec.createProjector()\n    const eventIds = new Set<string>()\n    let expectedPrevious: LedgerHash | null = null\n    for (let index = 0; index < entries.length; index += 1) {\n      const entry = entries[index]!\n      this.codec.checkEntryHeader(entry, index)\n      if (entry.sequence !== index) {\n        throw this.codec.integrityError(\n          `entry ${entry.event.eventId} has sequence ${entry.sequence}, expected ${index}`,\n        )\n      }\n      if (entry.previousHash !== expectedPrevious) {\n        throw this.codec.integrityError(\n          `entry ${entry.event.eventId} does not extend the previous hash`,\n        )\n      }\n      const { entryHash: _entryHash, ...material } = entry\n      const expectedHash = hashCanonical(material)\n      if (entry.entryHash !== expectedHash) {\n        throw this.codec.integrityError(\n          `entry ${entry.event.eventId} hash mismatch: expected ${expectedHash}, got ${entry.entryHash}`,\n        )\n      }\n      expectedPrevious = entry.entryHash\n      if (eventIds.has(entry.event.eventId)) {\n        throw this.codec.integrityError(\n          `duplicate eventId ${entry.event.eventId} in durable ledger`,\n        )\n      }\n      eventIds.add(entry.event.eventId)\n      projector.apply(entry, index)\n    }\n    return projector.finish(entries)\n  }\n}\n"],"mappings":";;;;;;;;;AAyCA,IAAa,sBAAb,cAAyC,MAAM;CAC7C,OAAyB;AAC3B;;AAQA,SAAgB,oBACd,SACkC;CAClC,IAAI,WAAW,aAAa,QAAQ,QAAQ,CAAC,GAAG,OAAO;EAAE,UAAU;EAAO,QAAQ;CAAW;CAC7F,MAAM,QAAQ,WAAW,QAAQ,QAAQ;CACzC,IAAI,MAAM,UAAU,QAClB,OAAO;EAAE,UAAU;EAAO,QAAQ;EAAQ,QAAQ,MAAM;CAAM;CAEhE,OAAO;AACT;;;;;;;AAQA,SAAgB,yBACd,SAC2B;CAC3B,MAAM,MAAM,QAAQ,OAAO,QAAQ;CACnC,IAAI,CAAC,OAAO,cAAc,GAAG,KAAK,OAAO,GACvC,MAAM,IAAI,oBAAoB,iDAAiD;CAEjF,MAAM,QAA6B;EAAE;EAAK,MAAM,SAAS;EAAG,OAAO,WAAW;CAAE;CAChF,KAAK,IAAI,UAAU,GAAG,UAAU,GAAG,WAAW,GAAG;EAC/C,IAAI,WAAW,aAAa,QAAQ,QAAQ,CAAC,GAC3C,OAAO;GAAE,UAAU;GAAO,QAAQ;EAAW;EAE/C,IAAI,aAAa,QAAQ,UAAU,OAAO,WAAW,SAAS,GAC5D,OAAO,SAAS,QAAQ,UAAU,KAAK;EAGzC,MAAM,SAAS,WAAW,QAAQ,QAAQ;EAC1C,IAAI,OAAO,UAAU,WAAW;EAChC,IAAI,OAAO,UAAU,QACnB,OAAO;GAAE,UAAU;GAAO,QAAQ;GAAQ,QAAQ,OAAO;EAAM;EAGjE,MAAM,cAAc,aAAa,QAAQ,QAAQ;EACjD,IAAI,CAAC,aAAa,aAAa,OAAO,WAAW,SAAS,GACxD,OAAO;GAAE,UAAU;GAAO,QAAQ;EAAW;EAE/C,IAAI;GACF,MAAM,UAAU,WAAW,QAAQ,QAAQ;GAC3C,IAAI,QAAQ,UAAU,QACpB,OAAO;IAAE,UAAU;IAAO,QAAQ;IAAQ,QAAQ,QAAQ;GAAM;GAElE,IAAI,QAAQ,UAAU,SAAS;IAC7B,MAAM,YAAY,GAAG,QAAQ,SAAS,SAAS,MAAM,MAAM,GAAG;IAC9D,IAAI;KACF,WAAW,QAAQ,UAAU,SAAS;KACtC,WAAW,SAAS;IACtB,SAAS,OAAO;KACd,IAAI,CAAC,UAAU,KAAK,GAAG,MAAM;IAC/B;GACF;GACA,IAAI,aAAa,QAAQ,UAAU,OAAO,aAAa,SAAS,GAC9D,OAAO,SAAS,QAAQ,UAAU,KAAK;EAE3C,UAAU;GACR,iBAAiB,aAAa,KAAK;EACrC;CACF;CAEA,MAAM,IAAI,oBAAoB,sCAAsC,QAAQ,UAAU;AACxF;AAEA,SAAS,SAAS,UAAkB,OAAuD;CACzF,OAAO;EACL,UAAU;EACV,MAAM;GACJ;GACA,eAAe,iBAAiB,UAAU,KAAK;EACjD;CACF;AACF;AAEA,SAAS,aAAa,UAAkB,OAA4B,QAAyB;CAC3F,MAAM,YAAY,GAAG,SAAS,GAAG,MAAM,IAAI,GAAG,MAAM,MAAM,GAAG,OAAO;CACpE,MAAM,aAAa,SACjB,WACA,UAAU,UAAU,UAAU,SAAS,UAAU,UACjD,GACF;CACA,IAAI;EACF,cAAc,YAAY,GAAG,eAAe,KAAK,EAAE,KAAK,MAAM;EAC9D,UAAU,UAAU;CACtB,SAAS,OAAO;EACd,UAAU,UAAU;EACpB,eAAe,SAAS;EACxB,MAAM;CACR;CACA,UAAU,UAAU;CAEpB,IAAI;EACF,SAAS,WAAW,QAAQ;EAC5B,OAAO;CACT,SAAS,OAAO;EACd,IAAI,CAAC,gBAAgB,KAAK,GAAG,MAAM;EACnC,OAAO;CACT,UAAU;EACR,eAAe,SAAS;CAC1B;AACF;AAEA,SAAS,WAAW,UAA8B;CAChD,MAAM,QAAQ,UAAU,QAAQ;CAChC,IAAI,CAAC,OAAO,OAAO,EAAE,OAAO,UAAU;CACtC,IAAI,MAAM,SAAS,SAAS,GAAG,OAAO;EAAE,OAAO;EAAQ;CAAM;CAC7D,IAAI;EACF,QAAQ,KAAK,MAAM,KAAK,CAAC;EACzB,OAAO;GAAE,OAAO;GAAQ;EAAM;CAChC,SAAS,OAAO;EACd,IAAI,gBAAgB,KAAK,GAAG,OAAO;GAAE,OAAO;GAAS;EAAM;EAC3D,OAAO;GAAE,OAAO;GAAQ;EAAM;CAChC;AACF;AAEA,SAAS,UAAU,UAAmD;CACpE,IAAI;CACJ,IAAI;EACF,WAAW,aAAa,UAAU,MAAM,CAAC,CAAC,KAAK;CACjD,SAAS,OAAO;EACd,IAAI,UAAU,KAAK,GAAG,OAAO,KAAA;EAC7B,MAAM;CACR;CAEA,IAAI;CACJ,IAAI;EACF,QAAQ,KAAK,MAAM,QAAQ;CAC7B,SAAS,OAAO;EACd,MAAM,aAAa,UAAU,KAAK;CACpC;CACA,IAAI,CAAC,QAAQ,KAAK,GAAG,MAAM,aAAa,QAAQ;CAChD,OAAO;AACT;AAEA,SAAS,QAAQ,OAA8C;CAC7D,IAAI,OAAO,UAAU,YAAY,UAAU,QAAQ,MAAM,QAAQ,KAAK,GAAG,OAAO;CAChF,MAAM,SAAS;CACf,OACE,OAAO,KAAK,MAAM,CAAC,CAAC,WAAW,KAC/B,OAAO,cAAc,OAAO,GAAG,KAC/B,OAAO,OAAO,GAAG,IAAI,KACrB,OAAO,OAAO,SAAS,YACvB,OAAO,KAAK,KAAK,CAAC,CAAC,SAAS,KAC5B,OAAO,OAAO,UAAU,YACxB,OAAO,MAAM,KAAK,CAAC,CAAC,SAAS;AAEjC;AAEA,SAAS,aAAa,UAAkB,OAAsC;CAC5E,MAAM,QAAQ,IAAI,oBAChB,0CAA0C,SAAS,4BACrD;CACA,IAAI,UAAU,KAAA,GAAW,OAAO,eAAe,OAAO,SAAS,EAAE,OAAO,MAAM,CAAC;CAC/E,OAAO;AACT;AAEA,SAAS,iBAAiB,UAAkB,OAAkC;CAC5E,MAAM,UAAU,UAAU,QAAQ;CAClC,IAAI,CAAC,SAAS;CACd,IAAI,eAAe,OAAO,MAAM,eAAe,KAAK,GAClD,MAAM,IAAI,oBAAoB,kDAAkD,SAAS,EAAE;CAE7F,IAAI;EACF,WAAW,QAAQ;CACrB,SAAS,OAAO;EACd,IAAI,CAAC,UAAU,KAAK,GAAG,MAAM;CAC/B;AACF;AAEA,SAAS,eAAe,OAAoC;CAC1D,OAAO,KAAK,UAAU;EAAE,MAAM,MAAM;EAAM,OAAO,MAAM;EAAO,KAAK,MAAM;CAAI,CAAC;AAChF;AAEA,SAAS,aAAa,UAA0B;CAC9C,OAAO,GAAG,SAAS;AACrB;AAEA,SAAS,gBAAgB,OAAyB;CAChD,OAAO,iBAAiB,SAAS,UAAU,SAAS,MAAM,SAAS;AACrE;AAEA,SAAS,UAAU,OAAyB;CAC1C,OAAO,iBAAiB,SAAS,UAAU,SAAS,MAAM,SAAS;AACrE;AAEA,SAAS,gBAAgB,OAAyB;CAChD,OAAO,iBAAiB,SAAS,UAAU,SAAS,MAAM,SAAS;AACrE;AAEA,SAAS,eAAe,MAAoB;CAC1C,IAAI;EACF,WAAW,IAAI;CACjB,SAAS,OAAO;EACd,IAAI,CAAC,UAAU,KAAK,GAAG,MAAM;CAC/B;AACF;;;;;;;;;;;;;;;;;AClOA,MAAa,sBAAsB;;AAGnC,IAAa,8BAAb,cAAiD,gBAAgB,CAAC;;;;AAKlE,SAAgB,gBAAgB,OAAwB;CACtD,IAAI;EACF,OAAO,uBAAuB,KAAK;CACrC,SAAS,OAAO;EACd,MAAM,sBAAsB,OAAO,KAAK;CAC1C;AACF;AAEA,SAAgB,cAAc,OAA4B;CACxD,IAAI;EACF,OAAO,YAAY,wBAAwB,KAAK,CAAC;CACnD,SAAS,OAAO;EACd,MAAM,sBAAsB,OAAO,KAAK;CAC1C;AACF;;;;AAKA,SAAS,sBAAsB,OAAgB,OAA6C;CAC1F,MAAM,SAAS,wBAAwB,OAAO,qBAAK,IAAI,IAAI,CAAC;CAE5D,OAAO,IAAI,4BAA4B,qCADxB,WAAW,KAAK,SAAS,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,KACP,EACpF,MACF,CAAC;AACH;AAEA,SAAS,wBAAwB,OAAgB,MAAc,WAAgC;CAC7F,IAAI,UAAU,MAAM,OAAO;CAC3B,IAAI,UAAU,KAAA,GAAW,OAAO,GAAG,KAAK;CACxC,MAAM,OAAO,OAAO;CACpB,IAAI,SAAS,UACX,OAAO,OAAO,SAAS,KAAK,IAAI,KAAK,GAAG,KAAK,MAAM,OAAO,KAAK;CAEjE,IAAI,SAAS,UACX,OAAO,kBAAkB,KAAe,IAAI,KAAK,GAAG,KAAK;CAE3D,IAAI,SAAS,WAAW,OAAO;CAC/B,IAAI,SAAS,UAAU,OAAO,GAAG,KAAK,QAAQ;CAC9C,MAAM,SAAS;CACf,IAAI,UAAU,IAAI,MAAM,GAAG,OAAO,GAAG,KAAK;CAC1C,MAAM,YAAY,OAAO,eAAe,MAAM;CAC9C,IAAI,CAAC,MAAM,QAAQ,MAAM,KAAK,cAAc,OAAO,aAAa,cAAc,MAC5E,OAAO,GAAG,KAAK,QAAQ,OAAO,aAAa,QAAQ,YAAY;CAEjE,MAAM,SAAS,IAAI,IAAI,SAAS,CAAC,CAAC,IAAI,MAAM;CAC5C,IAAI,MAAM,QAAQ,MAAM,GAAG;EACzB,KAAK,IAAI,QAAQ,GAAG,QAAQ,OAAO,QAAQ,SAAS,GAAG;GACrD,MAAM,SAAS,wBAAwB,OAAO,QAAQ,GAAG,KAAK,GAAG,MAAM,IAAI,MAAM;GACjF,IAAI,WAAW,IAAI,OAAO;EAC5B;EACA,OAAO;CACT;CACA,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,MAAM,GAAG;EACjD,IAAI,CAAC,kBAAkB,GAAG,GAAG,OAAO,GAAG,KAAK,GAAG,IAAI;EACnD,MAAM,SAAS,wBAAwB,OAAO,GAAG,KAAK,GAAG,OAAO,MAAM;EACtE,IAAI,WAAW,IAAI,OAAO;CAC5B;CACA,OAAO;AACT;AAEA,SAAS,kBAAkB,OAAwB;CACjD,KAAK,IAAI,QAAQ,GAAG,QAAQ,MAAM,QAAQ,SAAS,GAAG;EACpD,MAAM,OAAO,MAAM,WAAW,KAAK;EACnC,IAAI,QAAQ,SAAU,QAAQ,OAAQ;GACpC,MAAM,OAAO,MAAM,WAAW,QAAQ,CAAC;GACvC,IAAI,EAAE,QAAQ,SAAU,QAAQ,QAAS,OAAO;GAChD,SAAS;EACX,OAAO,IAAI,QAAQ,SAAU,QAAQ,OACnC,OAAO;CAEX;CACA,OAAO;AACT;;;;;;AClFA,SAAgB,iBAAiB,MAAc,MAAc,SAAkC;CAC7F,UAAU,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;CAC5C,MAAM,KAAK,SAAS,MAAM,UAAU,UAAU,UAAU,WAAW,UAAU,UAAU,GAAK;CAC5F,IAAI;EACF,SAAS,IAAI,OAAO,KAAK,MAAM,MAAM,GAAG,OAAO;EAC/C,UAAU,EAAE;CACd,UAAU;EACR,UAAU,EAAE;CACd;CACA,eAAe,QAAQ,IAAI,CAAC;AAC9B;;;;AAKA,SAAgB,0BACd,MACA,UACA,SACM;CACN,MAAM,YAAY,QAAQ,IAAI;CAC9B,UAAU,WAAW,EAAE,WAAW,KAAK,CAAC;CACxC,MAAM,gBAAgB,GAAG,KAAK;CAC9B,MAAM,KAAK,SACT,eACA,UAAU,UAAU,UAAU,WAAW,UAAU,SACnD,GACF;CACA,IAAI;EACF,SAAS,IAAI,OAAO,KAAK,UAAU,MAAM,GAAG,OAAO;EACnD,UAAU,EAAE;CACd,UAAU;EACR,UAAU,EAAE;CACd;CACA,WAAW,eAAe,IAAI;CAC9B,eAAe,SAAS;AAC1B;AAEA,SAAgB,mBAAsB,MAAc,SAA4B,KAAiB;CAC/F,MAAM,SAAS,sBAAsB,MAAM,SAAS,GAAG;CACvD,IAAI,CAAC,OAAO,UACV,MAAM,QAAQ,eAAe,GAAG,QAAQ,QAAQ,iBAAiB,KAAK,EAAE;CAE1E,OAAO,OAAO;AAChB;AAIA,SAAgB,sBACd,MACA,SACA,KACmB;CACnB,UAAU,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;CAC5C,MAAM,WAAW,GAAG,KAAK;CACzB,IAAI;EACF,MAAM,cAAc,yBAAyB,EAAE,SAAS,CAAC;EACzD,IAAI,CAAC,YAAY,UAAU,OAAO,EAAE,UAAU,MAAM;EACpD,IAAI;GACF,OAAO;IAAE,UAAU;IAAM,OAAO,IAAI;GAAE;EACxC,UAAU;GACR,YAAY,KAAK,QAAQ;EAC3B;CACF,SAAS,OAAO;EACd,IAAI,iBAAiB,qBACnB,MAAM,QAAQ,eAAe,MAAM,SAAS,EAAE,OAAO,MAAM,CAAC;EAE9D,MAAM;CACR;AACF;AAEA,SAAS,SAAS,IAAY,OAAe,SAAkC;CAC7E,IAAI,SAAS;CACb,OAAO,SAAS,MAAM,YAAY;EAChC,MAAM,UAAU,UAAU,IAAI,OAAO,QAAQ,MAAM,aAAa,MAAM;EACtE,IAAI,WAAW,GAAG,MAAM,QAAQ,eAAe,6BAA6B;EAC5E,UAAU;CACZ;AACF;AAEA,SAAS,eAAe,MAAoB;CAC1C,MAAM,KAAK,SAAS,MAAM,UAAU,QAAQ;CAC5C,IAAI;EACF,UAAU,EAAE;CACd,UAAU;EACR,UAAU,EAAE;CACd;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACnDA,SAAgB,mBAAmB,aAA6B;CAC9D,OAAO,GAAG,YAAY;AACxB;;;;;AAMA,SAAgB,oBACd,MACA,SAC0B;CAC1B,IAAI,CAAC,WAAW,IAAI,GAAG,OAAO;CAC9B,MAAM,OAAO,aAAa,MAAM,MAAM;CACtC,IAAI;CACJ,IAAI;EACF,MAAM,KAAK,MAAM,IAAI;CACvB,SAAS,OAAO;EACd,MAAM,QAAQ,eAAe,GAAG,QAAQ,QAAQ,gBAAgB,KAAK,qBAAqB,EACxF,OAAO,MACT,CAAC;CACH;CACA,IAAI,OAAO,QAAQ,YAAY,QAAQ,QAAQ,MAAM,QAAQ,GAAG,GAC9D,MAAM,QAAQ,eAAe,GAAG,QAAQ,QAAQ,gBAAgB,KAAK,kBAAkB;CAEzF,MAAM,SAAS;CACf,MAAM,OAAO,OAAO,KAAK,MAAM,CAAC,CAAC,KAAK;CACtC,IAAI,KAAK,WAAW,KAAK,KAAK,OAAO,eAAe,KAAK,OAAO,YAC9D,MAAM,QAAQ,eACZ,GAAG,QAAQ,QAAQ,gBAAgB,KAAK,aAAa,KAAK,KAAK,IAAI,EAAE,kCACvE;CAEF,MAAM,EAAE,UAAU,cAAc;CAChC,IAAI,OAAO,aAAa,YAAY,CAAC,OAAO,cAAc,QAAQ,KAAK,WAAW,GAChF,MAAM,QAAQ,eACZ,GAAG,QAAQ,QAAQ,gBAAgB,KAAK,wCAC1C;CAEF,IAAI,OAAO,cAAc,YAAY,CAAC,oBAAoB,KAAK,SAAS,GACtE,MAAM,QAAQ,eACZ,GAAG,QAAQ,QAAQ,gBAAgB,KAAK,kCAC1C;CAEF,OAAO;EAAE;EAAqB;CAAwB;AACxD;;;;AAKA,SAAgB,qBACd,MACA,MACA,SACM;CACN,0BAA0B,MAAM,GAAG,gBAAgB,IAAI,EAAE,KAAK,OAAO;AACvE;;;;AAKA,SAAgB,gCACd,SACA,MACA,SACA,SACM;CACN,MAAM,SAAS,QAAQ,KAAK;CAC5B,IAAI,WAAW,KAAA,GACb,MAAM,QAAQ,eACZ,GAAG,QAAQ,8BAA8B,KAAK,SAAS,IAAI,KAAK,UAAU,wBAAwB,QAAQ,OAAO,wEACnH;CAEF,IAAI,OAAO,aAAa,KAAK,UAC3B,MAAM,QAAQ,eACZ,GAAG,QAAQ,kBAAkB,KAAK,SAAS,oBAAoB,OAAO,SAAS,qDACjF;CAEF,IAAI,OAAO,cAAc,KAAK,WAC5B,MAAM,QAAQ,eACZ,GAAG,QAAQ,4BAA4B,KAAK,SAAS,WAAW,OAAO,UAAU,iCAAiC,KAAK,UAAU,6BACnI;AAEJ;;;;;;;;;;;;;;;;;;;;;;;;;ACxBA,MAAM,iCAAiB,IAAI,IAAmB;AAE9C,SAAS,SAAS,MAAqB;CACrC,MAAM,WAAW,eAAe,IAAI,IAAI;CACxC,IAAI,UAAU,OAAO;CACrB,MAAM,QAAQ,IAAI,MAAM;CACxB,eAAe,IAAI,MAAM,KAAK;CAC9B,OAAO;AACT;;;AAIA,IAAa,oBAAb,MAAiG;CAC/F;;CAEA;CACA;CACA;CACA;CAEA,YACE,MACA,OACA,UAAoC,CAAC,GACrC;EACA,KAAK,OAAO,QAAQ,IAAI;EACxB,KAAK,kBAAkB,mBAAmB,KAAK,IAAI;EACnD,KAAK,QAAQ;EACb,KAAK,QAAQ,SAAS,KAAK,IAAI;EAC/B,KAAK,qBAAqB,QAAQ,uBAAuB;CAC3D;CAEA,MAAM,SAAgF;EACpF,OAAO,KAAK,MAAM,mBAChB,mBAAmB,KAAK,MAAM,KAAK,aAAa,KAAK,aAAa,CAAC,CACrE;CACF;CAEA,MAAM,OACJ,OACA,UAA+B,CAAC,GACuC;EACvE,OAAO,KAAK,MAAM,mBAChB,mBAAmB,KAAK,MAAM,KAAK,aAAa;GAC9C,MAAM,SAAS,KAAK,aAAa;GACjC,MAAM,WAAW,OAAO,QAAQ,MAAM,UAAU,MAAM,MAAM,YAAY,MAAM,OAAO;GACrF,IAAI,UAAU;IACZ,IAAI,gBAAgB,SAAS,KAAK,MAAM,gBAAgB,KAAK,GAC3D,MAAM,KAAK,MAAM,cACf,WAAW,MAAM,QAAQ,uCAC3B;IAEF,OAAO;KAAE,OAAO;KAAU,UAAU;KAAO,YAAY,OAAO;IAAW;GAC3E;GAEA,MAAM,WAAW;IACf,GAAG,KAAK,MAAM;IACd,UAAU,OAAO,QAAQ;IACzB,cAAc,OAAO,QAAQ,GAAG,EAAE,CAAC,EAAE,aAAa;IAClD;GACF;GACA,MAAM,QAAQ;IAAE,GAAG;IAAU,WAAW,cAAc,QAAQ;GAAE;GAIhE,MAAM,aAAa,KAAK,QAAQ,CAAC,GAAG,OAAO,SAAS,KAAK,CAAC;GAC1D,iBAAiB,KAAK,MAAM,GAAG,gBAAgB,KAAK,EAAE,KAAK,KAAK,KAAK;GACrE,IAAI,QAAQ,YAAY,MAKtB,qBACE,KAAK,iBACL;IAAE,UAAU,MAAM;IAAU,WAAW,MAAM;GAAU,GACvD,KAAK,KACP;GAEF,OAAO;IAAE;IAAO,UAAU;IAAM;GAAW;EAC7C,CAAC,CACH;CACF;;CAGA,MAAM,cAAiD;EACrD,OAAO,KAAK,MAAM,mBAChB,mBAAmB,KAAK,MAAM,KAAK,aACjC,oBAAoB,KAAK,iBAAiB,KAAK,KAAK,CACtD,CACF;CACF;;;;CAKA,MAAM,iBAA6C;EACjD,OAAO,KAAK,MAAM,mBAChB,mBAAmB,KAAK,MAAM,KAAK,aAAa;GAC9C,MAAM,UAAU,KAAK,YAAY;GACjC,KAAK,QAAQ,OAAO;GAGpB,KAAK,kBAAkB,SAAS,KAAK;GACrC,MAAM,OAAO,QAAQ,GAAG,EAAE;GAC1B,IAAI,SAAS,KAAA,GACX,MAAM,KAAK,MAAM,eACf,GAAG,KAAK,MAAM,QAAQ,GAAG,KAAK,KAAK,mCACrC;GAEF,MAAM,OAA0B;IAAE,UAAU,KAAK;IAAU,WAAW,KAAK;GAAU;GACrF,qBAAqB,KAAK,iBAAiB,MAAM,KAAK,KAAK;GAC3D,OAAO;EACT,CAAC,CACH;CACF;CAEA,eAAqF;EACnF,MAAM,UAAU,KAAK,YAAY;EACjC,MAAM,aAAa,KAAK,QAAQ,OAAO;EAIvC,KAAK,kBAAkB,SAAS,KAAK,kBAAkB;EACvD,OAAO;GAAE;GAAS;EAAW;CAC/B;CAEA,kBAA0B,SAAyC,YAA2B;EAC5F,MAAM,SAAS,oBAAoB,KAAK,iBAAiB,KAAK,KAAK;EACnE,IAAI,WAAW,MAAM;GACnB,IAAI,cAAc,QAAQ,SAAS,GACjC,MAAM,KAAK,MAAM,eACf,GAAG,KAAK,MAAM,QAAQ,GAAG,KAAK,KAAK,OAAO,QAAQ,OAAO,kCAAkC,KAAK,gBAAgB,gEAClH;GAEF;EACF;EACA,gCACE,SACA,QACA,KAAK,OACL,GAAG,KAAK,MAAM,QAAQ,GAAG,KAAK,MAChC;CACF;CAEA,cAAsD;EACpD,IAAI,CAAC,WAAW,KAAK,IAAI,GAAG,OAAO,CAAC;EACpC,MAAM,OAAO,aAAa,KAAK,MAAM,MAAM;EAC3C,IAAI,KAAK,WAAW,GAAG,OAAO,CAAC;EAC/B,IAAI,CAAC,KAAK,SAAS,IAAI,GACrB,MAAM,KAAK,MAAM,eACf,GAAG,KAAK,MAAM,QAAQ,GAAG,KAAK,KAAK,gDACrC;EAGF,MAAM,QAAQ,KAAK,MAAM,GAAG,EAAE,CAAC,CAAC,MAAM,IAAI;EAC1C,MAAM,UAA0C,CAAC;EACjD,KAAK,IAAI,QAAQ,GAAG,QAAQ,MAAM,QAAQ,SAAS,GAAG;GACpD,MAAM,OAAO,MAAM;GACnB,IAAI,KAAK,WAAW,GAClB,MAAM,KAAK,MAAM,eACf,GAAG,KAAK,MAAM,QAAQ,GAAG,KAAK,KAAK,2BAA2B,QAAQ,GACxE;GAEF,IAAI;GACJ,IAAI;IACF,MAAM,KAAK,MAAM,IAAI;GACvB,SAAS,OAAO;IACd,MAAM,KAAK,MAAM,eACf,GAAG,KAAK,MAAM,QAAQ,GAAG,KAAK,KAAK,4BAA4B,QAAQ,KACvE,EAAE,OAAO,MAAM,CACjB;GACF;GACA,MAAM,QAAQ,KAAK,MAAM,WAAW,KAAK;IAAE,MAAM,KAAK;IAAM,MAAM,QAAQ;GAAE,CAAC;GAC7E,IAAI;GACJ,IAAI;IACF,YAAY,gBAAgB,KAAK;GACnC,SAAS,OAAO;IACd,MAAM,KAAK,MAAM,eACf,GAAG,KAAK,MAAM,QAAQ,GAAG,KAAK,KAAK,iDAAiD,QAAQ,KAC5F,EAAE,OAAO,MAAM,CACjB;GACF;GACA,IAAI,SAAS,WACX,MAAM,KAAK,MAAM,eACf,GAAG,KAAK,MAAM,QAAQ,GAAG,KAAK,KAAK,mCAAmC,QAAQ,GAChF;GAEF,QAAQ,KAAK,KAAK;EACpB;EACA,OAAO;CACT;CAEA,QAAgB,SAAqD;EACnE,MAAM,YAAY,KAAK,MAAM,gBAAgB;EAC7C,MAAM,2BAAW,IAAI,IAAY;EACjC,IAAI,mBAAsC;EAC1C,KAAK,IAAI,QAAQ,GAAG,QAAQ,QAAQ,QAAQ,SAAS,GAAG;GACtD,MAAM,QAAQ,QAAQ;GACtB,KAAK,MAAM,iBAAiB,OAAO,KAAK;GACxC,IAAI,MAAM,aAAa,OACrB,MAAM,KAAK,MAAM,eACf,SAAS,MAAM,MAAM,QAAQ,gBAAgB,MAAM,SAAS,aAAa,OAC3E;GAEF,IAAI,MAAM,iBAAiB,kBACzB,MAAM,KAAK,MAAM,eACf,SAAS,MAAM,MAAM,QAAQ,mCAC/B;GAEF,MAAM,EAAE,WAAW,YAAY,GAAG,aAAa;GAC/C,MAAM,eAAe,cAAc,QAAQ;GAC3C,IAAI,MAAM,cAAc,cACtB,MAAM,KAAK,MAAM,eACf,SAAS,MAAM,MAAM,QAAQ,2BAA2B,aAAa,QAAQ,MAAM,WACrF;GAEF,mBAAmB,MAAM;GACzB,IAAI,SAAS,IAAI,MAAM,MAAM,OAAO,GAClC,MAAM,KAAK,MAAM,eACf,qBAAqB,MAAM,MAAM,QAAQ,mBAC3C;GAEF,SAAS,IAAI,MAAM,MAAM,OAAO;GAChC,UAAU,MAAM,OAAO,KAAK;EAC9B;EACA,OAAO,UAAU,OAAO,OAAO;CACjC;AACF"}

package/dist/meta-eval/index.js

@@ -1,8 +1,8 @@
 import { s as ValidationError } from "../errors-8YnH8WlF.js";
-import { A as spearmanR, E as pearsonR } from "../statistics-CnnxdpOg.js";
+import { A as spearmanR, E as pearsonR } from "../statistics-DWM_AyLe.js";
 import { o as llmSpans, t as aggregateLlm } from "../query-Di7eEQ79.js";
 import { t as analyzeSeries } from "../series-convergence-CjO2QdRW.js";
-import { t as rubricPredictiveValidity } from "../rubric-predictive-validity-B3xmbmS1.js";
+import { t as rubricPredictiveValidity } from "../rubric-predictive-validity-QG7ydk0s.js";
 import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "../outcome-store-ChBKlTd_.js";
 //#region src/meta-eval/calibration.ts
 async function calibrationCurve(traceStore, outcomeStore, evalMetric, outcomeMetric, options = {}) {

package/dist/{mint-D5_87M5L.js → mint-BvkwcYZU.js}

@@ -1,6 +1,6 @@
 import { s as ValidationError } from "./errors-8YnH8WlF.js";
 import { a as scoreOrigin, i as rolloutRewardFields } from "./reward-nw2xZGZG.js";
-import { o as runTaskScore } from "./run-record-CN8Zd21B.js";
+import { o as runTaskScore } from "./run-record-BIwU2wdV.js";
 import { i as ROLLOUT_SCHEMA, s as assertMinted } from "./schema-C6DW4ZHR.js";
 import { t as buildTrajectory } from "./trajectory-D_7rLrvE.js";
 //#region src/rollout/mint.ts
@@ -198,4 +198,4 @@ async function mintRolloutRows(records, store, options = {}) {
 //#endregion
 export { mintRolloutRows as t };
 
-//# sourceMappingURL=mint-D5_87M5L.js.map
+//# sourceMappingURL=mint-BvkwcYZU.js.map

package/dist/{mint-D5_87M5L.js.map → mint-BvkwcYZU.js.map}

@@ -1 +1 @@
-{"version":3,"file":"mint-D5_87M5L.js","names":[],"sources":["../src/rollout/mint.ts"],"sourcesContent":["/**\n * Rollout minting — `tangle.rollout.v1` lines joined from the records the\n * substrate ALREADY keeps. There is no separate rollout store: a rollout\n * is the JOIN of a RunRecord (identity, provenance, cost, outcome) with\n * its trace (spans share `runId`), projected into the canonical line.\n *\n * Composition, not duplication:\n *   - identity/provenance     → `RunRecord` (candidateId, splitTag, agentProfile, hashes)\n *   - step structure          → `buildTrajectory` over the shared TraceStore\n *   - preference-pair export  → `feedbackTrajectoryToOptimizerRow` (feedback-trajectory.ts)\n *   - PRM / reward-model      → `reward-model-export.ts`\n *\n * Anti-Goodhart invariant: a run whose `outcome.realness.gated` is true is\n * never exported with a positive reward OR with any of the numbers that reward\n * was computed from. The gate travels into the training data (`reward` forced\n * to 0, `realness_gated: true`) and the whole outcome is transformed by\n * `gateGamedOutcome` inside `assertMinted` below, which relocates `metrics` and\n * `verdict` to `provenance.gated_evidence`. Mint returns\n * `MintedRolloutLine[]`: the brand the training exporters require, which only\n * this function, `readRolloutLedger`, and an explicit `assertMinted` can mint.\n *\n * A record carrying NEITHER split score is REJECTED (`ValidationError`), never\n * minted at 0 — \"nobody graded this\" is not the same claim as \"graded a total\n * failure\", and a trainer reading 0 learns the second. Lines that already\n * carry `reward: null` (interchange imports, existing ledgers) remain valid on\n * the wire; only the RunRecord→line door refuses.\n *\n * Records without spans become labeled GAP LINES (messages: [],\n * provenance.gap) — present in the output AND surfaced in\n * `missingTraces`; a capture gap is a finding, never a silent omission.\n */\n\nimport { ValidationError } from '../errors'\nimport { type RunRecord, runTaskScore } from '../run-record'\nimport type { LlmSpan, Message, Span, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory } from '../trajectory'\nimport { rolloutRewardFields, scoreOrigin } from './reward'\nimport {\n  assertMinted,\n  type ChatMessage,\n  type MintedRolloutLine,\n  ROLLOUT_SCHEMA,\n  type RolloutRole,\n  type RolloutSplit,\n  type RolloutStep,\n} from './schema'\n\n/** Redactor applied to every exported string (secrets, PII). Identity by default. */\nexport type RolloutScrubber = (text: string) => string\n\nexport interface MintRolloutOptions {\n  scrub?: RolloutScrubber\n  /** Cap steps per line (longest runs first drop middle steps). Default: no cap. */\n  maxSteps?: number\n  /** Role recorded on every minted line. Default 'agent' (a solo eval run). */\n  role?: RolloutRole\n  /** Task suite label. Default: the record's `experimentId`. */\n  suite?: string\n  /** Injected clock for deterministic output. */\n  now?: () => Date\n}\n\nexport interface MintRolloutResult {\n  rows: MintedRolloutLine[]\n  /** runIds that had a RunRecord but no spans — emitted as gap lines AND listed here. */\n  missingTraces: string[]\n}\n\nconst asText = (v: unknown, scrub: RolloutScrubber): string => {\n  const s = typeof v === 'string' ? v : JSON.stringify(v)\n  return scrub(s ?? '')\n}\n\nfunction projectStep(span: Span, scrub: RolloutScrubber): RolloutStep {\n  const base: RolloutStep = {\n    kind: span.kind,\n    name: scrub(span.name),\n    status: span.status,\n    durationMs: span.endedAt !== undefined ? span.endedAt - span.startedAt : undefined,\n  }\n  if (span.kind === 'llm') {\n    const llm = span as LlmSpan\n    const last = llm.messages[llm.messages.length - 1]\n    if (last) base.input = scrub(last.content)\n    if (llm.output !== undefined) base.output = scrub(llm.output)\n  } else if (span.kind === 'tool') {\n    const tool = span as ToolSpan\n    base.input = asText(tool.args, scrub)\n    if (tool.result !== undefined) base.output = asText(tool.result, scrub)\n  }\n  return base\n}\n\n/** The final llm span's history + output is the completed conversation. */\nfunction finalConversation(spans: Span[], scrub: RolloutScrubber): ChatMessage[] {\n  const llms = spans.filter((s): s is LlmSpan => s.kind === 'llm')\n  const last = llms[llms.length - 1]\n  if (!last) return []\n  const messages: ChatMessage[] = last.messages.map((m: Message) => ({\n    role: m.role,\n    content: scrub(m.content),\n  }))\n  if (last.output !== undefined && last.output !== '') {\n    messages.push({ role: 'assistant', content: scrub(last.output) })\n  }\n  return messages\n}\n\n// The reward derivations live in the leaf module `./reward` so gate and\n// reporting code can import them without dragging in the trace store; they are\n// re-exported here because the derivations shipped from this path.\nexport {\n  isRealnessGated,\n  observedScore,\n  observedSplitScore,\n  type ScoreOrigin,\n  type ScorePreference,\n  scoreOrigin,\n  trainingReward,\n  trainingScore,\n} from './reward'\n\nconst REWARD_SOURCE: Record<ReturnType<typeof scoreOrigin>, string> = {\n  holdout: 'run-record/holdout-score',\n  search: 'run-record/search-score',\n  unscored: 'run-record/unscored',\n}\n\n/**\n * The mint door refuses an execution-only record: a missing training label is\n * not a zero reward, and not a mintable line either. Lines that already carry\n * `reward: null` — interchange imports, existing ledgers — stay valid on the\n * wire and keep their labeled gap; this guard is only about the\n * RunRecord→line door, where the producer can still be told to go score the\n * run instead of shipping an unlabeled row.\n */\nfunction requireTaskScore(record: RunRecord): void {\n  if (runTaskScore(record) === undefined) {\n    throw new ValidationError(`Cannot mint rollout for run ${record.runId}: task score is missing`)\n  }\n}\n\nconst SPLIT_FROM_TAG: Record<RunRecord['splitTag'], RolloutSplit> = {\n  search: 'search',\n  dev: 'dev',\n  holdout: 'holdout',\n}\n\nfunction mintLine(\n  record: RunRecord,\n  steps: RolloutStep[],\n  messages: ChatMessage[],\n  options: MintRolloutOptions,\n  capturedAt: string,\n  gap?: string,\n): MintedRolloutLine {\n  // A missing task score is refused before anything is built: an\n  // execution-only record has no training label, and a missing label is\n  // neither a zero reward nor a mintable row.\n  requireTaskScore(record)\n  // `reward` and `realness_gated` come out of one call, so neither door into\n  // the waist can write one and forget the other.\n  const rewardFields = rolloutRewardFields(record)\n  const uncaptured = record.costProvenance.kind === 'uncaptured'\n  const terminalOutcome = record.terminalOutcome\n  const isCompleted = terminalOutcome === 'succeeded' || terminalOutcome === 'failed'\n  const isTruncated = terminalOutcome === 'cancelled' || terminalOutcome === 'incomplete'\n  const terminalError =\n    terminalOutcome === 'failed' ||\n    terminalOutcome === 'cancelled' ||\n    terminalOutcome === 'incomplete'\n      ? (record.terminalFailureReason ?? `run ended ${terminalOutcome}`)\n      : null\n  // `assertMinted` rather than a cast: mint is the producer the whole gate\n  // rests on, so it proves the line it just built is valid instead of asserting\n  // it by fiat. The brand is unforgeable precisely because nobody casts to it.\n  return assertMinted(\n    {\n      schema: ROLLOUT_SCHEMA,\n      rollout_id: record.runId,\n      parent_rollout_id: null,\n      run_id: record.runId,\n      experiment_id: record.experimentId,\n      candidate_id: record.candidateId,\n      generation: null,\n      candidate_index: null,\n      role: options.role ?? 'agent',\n      task: {\n        suite: options.suite ?? record.experimentId,\n        instance_id: record.scenarioId,\n        split: SPLIT_FROM_TAG[record.splitTag],\n        seed: record.seed,\n        rep: 0,\n      },\n      policy: {\n        harness: null,\n        harness_version: null,\n        model: record.model,\n        provider: null,\n        profile_commit: record.commitSha,\n        prompt_hash: record.promptHash,\n        config_hash: record.configHash,\n        agent_profile_cell_id: record.agentProfile?.cellId ?? null,\n        sampling: null,\n      },\n      messages,\n      tool_defs: [],\n      ...(steps.length > 0 ? { steps } : {}),\n      outcome: {\n        ...rewardFields,\n        reward_source: REWARD_SOURCE[scoreOrigin(record)],\n        verdict: null,\n        // A verbatim bulk copy, deliberately UNFILTERED here. `outcome.raw`\n        // holds the per-layer verifier scores (`layer.*`) that the reward was\n        // derived from, so on a gated run this dict is the reward signal in\n        // component form — but filtering it at this call site is the pattern\n        // that has now leaked twice, because the next producer to write a\n        // reward-bearing field forgets. The gate is applied to the whole\n        // outcome once, in `assertMinted` below (`gateGamedOutcome`), which\n        // moves the block to `provenance.gated_evidence` when the run is gated\n        // and leaves it here untouched when it is not.\n        metrics: { ...record.outcome.raw },\n        is_completed: isCompleted,\n        is_truncated: isTruncated,\n        error: terminalError,\n      },\n      cost: {\n        usd: uncaptured ? null : record.costUsd,\n        tokens_in: record.tokenUsage.input,\n        tokens_out: record.tokenUsage.output,\n        tokens_reasoning: record.tokenUsage.reasoning ?? null,\n        cache_read: record.tokenUsage.cached ?? null,\n        cache_write: record.tokenUsage.cacheWrite ?? null,\n        wall_s: Math.round(record.wallMs / 1000),\n      },\n      artifacts: { patch_path: null, run_dir: null, transcript_ref: null },\n      provenance: {\n        captured_at: capturedAt,\n        capture: 'mint',\n        ...(gap !== undefined ? { gap } : {}),\n      },\n    },\n    `minted rollout line for run ${record.runId}`,\n  )\n}\n\n/**\n * Join RunRecords with their traces into canonical rollout lines. Records\n * without spans are emitted as labeled gap lines and reported in\n * `missingTraces`. Execution-only records without a task score are rejected\n * because a missing training label is not a zero reward.\n */\nexport async function mintRolloutRows(\n  records: RunRecord[],\n  store: TraceStore,\n  options: MintRolloutOptions = {},\n): Promise<MintRolloutResult> {\n  const scrub = options.scrub ?? ((t) => t)\n  const capturedAt = (options.now?.() ?? new Date()).toISOString()\n  const rows: MintedRolloutLine[] = []\n  const missingTraces: string[] = []\n  for (const record of records) {\n    const trajectory = await buildTrajectory(store, record.runId)\n    if (trajectory.steps.length === 0) {\n      missingTraces.push(record.runId)\n      rows.push(\n        mintLine(record, [], [], options, capturedAt, 'no trace spans recorded for this runId'),\n      )\n      continue\n    }\n    let steps = trajectory.steps.map((s) => projectStep(s.span, scrub))\n    if (options.maxSteps !== undefined && steps.length > options.maxSteps) {\n      // Keep the head and tail — the middle of a long run is the least\n      // informative for outcome attribution.\n      const head = Math.ceil(options.maxSteps / 2)\n      const tail = options.maxSteps - head\n      steps = [...steps.slice(0, head), ...steps.slice(steps.length - tail)]\n    }\n    const conversation = finalConversation(\n      trajectory.steps.map((s) => s.span),\n      scrub,\n    )\n    const gap =\n      conversation.length === 0 ? 'trace has no llm spans — no conversation to inline' : undefined\n    rows.push(mintLine(record, steps, conversation, options, capturedAt, gap))\n  }\n  return { rows, missingTraces }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqEA,MAAM,UAAU,GAAY,UAAmC;CAE7D,OAAO,OADG,OAAO,MAAM,WAAW,IAAI,KAAK,UAAU,CAAC,MACpC,EAAE;AACtB;AAEA,SAAS,YAAY,MAAY,OAAqC;CACpE,MAAM,OAAoB;EACxB,MAAM,KAAK;EACX,MAAM,MAAM,KAAK,IAAI;EACrB,QAAQ,KAAK;EACb,YAAY,KAAK,YAAY,KAAA,IAAY,KAAK,UAAU,KAAK,YAAY,KAAA;CAC3E;CACA,IAAI,KAAK,SAAS,OAAO;EACvB,MAAM,MAAM;EACZ,MAAM,OAAO,IAAI,SAAS,IAAI,SAAS,SAAS;EAChD,IAAI,MAAM,KAAK,QAAQ,MAAM,KAAK,OAAO;EACzC,IAAI,IAAI,WAAW,KAAA,GAAW,KAAK,SAAS,MAAM,IAAI,MAAM;CAC9D,OAAO,IAAI,KAAK,SAAS,QAAQ;EAC/B,MAAM,OAAO;EACb,KAAK,QAAQ,OAAO,KAAK,MAAM,KAAK;EACpC,IAAI,KAAK,WAAW,KAAA,GAAW,KAAK,SAAS,OAAO,KAAK,QAAQ,KAAK;CACxE;CACA,OAAO;AACT;;AAGA,SAAS,kBAAkB,OAAe,OAAuC;CAC/E,MAAM,OAAO,MAAM,QAAQ,MAAoB,EAAE,SAAS,KAAK;CAC/D,MAAM,OAAO,KAAK,KAAK,SAAS;CAChC,IAAI,CAAC,MAAM,OAAO,CAAC;CACnB,MAAM,WAA0B,KAAK,SAAS,KAAK,OAAgB;EACjE,MAAM,EAAE;EACR,SAAS,MAAM,EAAE,OAAO;CAC1B,EAAE;CACF,IAAI,KAAK,WAAW,KAAA,KAAa,KAAK,WAAW,IAC/C,SAAS,KAAK;EAAE,MAAM;EAAa,SAAS,MAAM,KAAK,MAAM;CAAE,CAAC;CAElE,OAAO;AACT;AAgBA,MAAM,gBAAgE;CACpE,SAAS;CACT,QAAQ;CACR,UAAU;AACZ;;;;;;;;;AAUA,SAAS,iBAAiB,QAAyB;CACjD,IAAI,aAAa,MAAM,MAAM,KAAA,GAC3B,MAAM,IAAI,gBAAgB,+BAA+B,OAAO,MAAM,wBAAwB;AAElG;AAEA,MAAM,iBAA8D;CAClE,QAAQ;CACR,KAAK;CACL,SAAS;AACX;AAEA,SAAS,SACP,QACA,OACA,UACA,SACA,YACA,KACmB;CAInB,iBAAiB,MAAM;CAGvB,MAAM,eAAe,oBAAoB,MAAM;CAC/C,MAAM,aAAa,OAAO,eAAe,SAAS;CAClD,MAAM,kBAAkB,OAAO;CAC/B,MAAM,cAAc,oBAAoB,eAAe,oBAAoB;CAC3E,MAAM,cAAc,oBAAoB,eAAe,oBAAoB;CAC3E,MAAM,gBACJ,oBAAoB,YACpB,oBAAoB,eACpB,oBAAoB,eACf,OAAO,yBAAyB,aAAa,oBAC9C;CAIN,OAAO,aACL;EACE,QAAQ;EACR,YAAY,OAAO;EACnB,mBAAmB;EACnB,QAAQ,OAAO;EACf,eAAe,OAAO;EACtB,cAAc,OAAO;EACrB,YAAY;EACZ,iBAAiB;EACjB,MAAM,QAAQ,QAAQ;EACtB,MAAM;GACJ,OAAO,QAAQ,SAAS,OAAO;GAC/B,aAAa,OAAO;GACpB,OAAO,eAAe,OAAO;GAC7B,MAAM,OAAO;GACb,KAAK;EACP;EACA,QAAQ;GACN,SAAS;GACT,iBAAiB;GACjB,OAAO,OAAO;GACd,UAAU;GACV,gBAAgB,OAAO;GACvB,aAAa,OAAO;GACpB,aAAa,OAAO;GACpB,uBAAuB,OAAO,cAAc,UAAU;GACtD,UAAU;EACZ;EACA;EACA,WAAW,CAAC;EACZ,GAAI,MAAM,SAAS,IAAI,EAAE,MAAM,IAAI,CAAC;EACpC,SAAS;GACP,GAAG;GACH,eAAe,cAAc,YAAY,MAAM;GAC/C,SAAS;GAUT,SAAS,EAAE,GAAG,OAAO,QAAQ,IAAI;GACjC,cAAc;GACd,cAAc;GACd,OAAO;EACT;EACA,MAAM;GACJ,KAAK,aAAa,OAAO,OAAO;GAChC,WAAW,OAAO,WAAW;GAC7B,YAAY,OAAO,WAAW;GAC9B,kBAAkB,OAAO,WAAW,aAAa;GACjD,YAAY,OAAO,WAAW,UAAU;GACxC,aAAa,OAAO,WAAW,cAAc;GAC7C,QAAQ,KAAK,MAAM,OAAO,SAAS,GAAI;EACzC;EACA,WAAW;GAAE,YAAY;GAAM,SAAS;GAAM,gBAAgB;EAAK;EACnE,YAAY;GACV,aAAa;GACb,SAAS;GACT,GAAI,QAAQ,KAAA,IAAY,EAAE,IAAI,IAAI,CAAC;EACrC;CACF,GACA,+BAA+B,OAAO,OACxC;AACF;;;;;;;AAQA,eAAsB,gBACpB,SACA,OACA,UAA8B,CAAC,GACH;CAC5B,MAAM,QAAQ,QAAQ,WAAW,MAAM;CACvC,MAAM,cAAc,QAAQ,MAAM,qBAAK,IAAI,KAAK,EAAA,CAAG,YAAY;CAC/D,MAAM,OAA4B,CAAC;CACnC,MAAM,gBAA0B,CAAC;CACjC,KAAK,MAAM,UAAU,SAAS;EAC5B,MAAM,aAAa,MAAM,gBAAgB,OAAO,OAAO,KAAK;EAC5D,IAAI,WAAW,MAAM,WAAW,GAAG;GACjC,cAAc,KAAK,OAAO,KAAK;GAC/B,KAAK,KACH,SAAS,QAAQ,CAAC,GAAG,CAAC,GAAG,SAAS,YAAY,wCAAwC,CACxF;GACA;EACF;EACA,IAAI,QAAQ,WAAW,MAAM,KAAK,MAAM,YAAY,EAAE,MAAM,KAAK,CAAC;EAClE,IAAI,QAAQ,aAAa,KAAA,KAAa,MAAM,SAAS,QAAQ,UAAU;GAGrE,MAAM,OAAO,KAAK,KAAK,QAAQ,WAAW,CAAC;GAC3C,MAAM,OAAO,QAAQ,WAAW;GAChC,QAAQ,CAAC,GAAG,MAAM,MAAM,GAAG,IAAI,GAAG,GAAG,MAAM,MAAM,MAAM,SAAS,IAAI,CAAC;EACvE;EACA,MAAM,eAAe,kBACnB,WAAW,MAAM,KAAK,MAAM,EAAE,IAAI,GAClC,KACF;EACA,MAAM,MACJ,aAAa,WAAW,IAAI,uDAAuD,KAAA;EACrF,KAAK,KAAK,SAAS,QAAQ,OAAO,cAAc,SAAS,YAAY,GAAG,CAAC;CAC3E;CACA,OAAO;EAAE;EAAM;CAAc;AAC/B"}
+{"version":3,"file":"mint-BvkwcYZU.js","names":[],"sources":["../src/rollout/mint.ts"],"sourcesContent":["/**\n * Rollout minting — `tangle.rollout.v1` lines joined from the records the\n * substrate ALREADY keeps. There is no separate rollout store: a rollout\n * is the JOIN of a RunRecord (identity, provenance, cost, outcome) with\n * its trace (spans share `runId`), projected into the canonical line.\n *\n * Composition, not duplication:\n *   - identity/provenance     → `RunRecord` (candidateId, splitTag, agentProfile, hashes)\n *   - step structure          → `buildTrajectory` over the shared TraceStore\n *   - preference-pair export  → `feedbackTrajectoryToOptimizerRow` (feedback-trajectory.ts)\n *   - PRM / reward-model      → `reward-model-export.ts`\n *\n * Anti-Goodhart invariant: a run whose `outcome.realness.gated` is true is\n * never exported with a positive reward OR with any of the numbers that reward\n * was computed from. The gate travels into the training data (`reward` forced\n * to 0, `realness_gated: true`) and the whole outcome is transformed by\n * `gateGamedOutcome` inside `assertMinted` below, which relocates `metrics` and\n * `verdict` to `provenance.gated_evidence`. Mint returns\n * `MintedRolloutLine[]`: the brand the training exporters require, which only\n * this function, `readRolloutLedger`, and an explicit `assertMinted` can mint.\n *\n * A record carrying NEITHER split score is REJECTED (`ValidationError`), never\n * minted at 0 — \"nobody graded this\" is not the same claim as \"graded a total\n * failure\", and a trainer reading 0 learns the second. Lines that already\n * carry `reward: null` (interchange imports, existing ledgers) remain valid on\n * the wire; only the RunRecord→line door refuses.\n *\n * Records without spans become labeled GAP LINES (messages: [],\n * provenance.gap) — present in the output AND surfaced in\n * `missingTraces`; a capture gap is a finding, never a silent omission.\n */\n\nimport { ValidationError } from '../errors'\nimport { type RunRecord, runTaskScore } from '../run-record'\nimport type { LlmSpan, Message, Span, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory } from '../trajectory'\nimport { rolloutRewardFields, scoreOrigin } from './reward'\nimport {\n  assertMinted,\n  type ChatMessage,\n  type MintedRolloutLine,\n  ROLLOUT_SCHEMA,\n  type RolloutRole,\n  type RolloutSplit,\n  type RolloutStep,\n} from './schema'\n\n/** Redactor applied to every exported string (secrets, PII). Identity by default. */\nexport type RolloutScrubber = (text: string) => string\n\nexport interface MintRolloutOptions {\n  scrub?: RolloutScrubber\n  /** Cap steps per line (longest runs first drop middle steps). Default: no cap. */\n  maxSteps?: number\n  /** Role recorded on every minted line. Default 'agent' (a solo eval run). */\n  role?: RolloutRole\n  /** Task suite label. Default: the record's `experimentId`. */\n  suite?: string\n  /** Injected clock for deterministic output. */\n  now?: () => Date\n}\n\nexport interface MintRolloutResult {\n  rows: MintedRolloutLine[]\n  /** runIds that had a RunRecord but no spans — emitted as gap lines AND listed here. */\n  missingTraces: string[]\n}\n\nconst asText = (v: unknown, scrub: RolloutScrubber): string => {\n  const s = typeof v === 'string' ? v : JSON.stringify(v)\n  return scrub(s ?? '')\n}\n\nfunction projectStep(span: Span, scrub: RolloutScrubber): RolloutStep {\n  const base: RolloutStep = {\n    kind: span.kind,\n    name: scrub(span.name),\n    status: span.status,\n    durationMs: span.endedAt !== undefined ? span.endedAt - span.startedAt : undefined,\n  }\n  if (span.kind === 'llm') {\n    const llm = span as LlmSpan\n    const last = llm.messages[llm.messages.length - 1]\n    if (last) base.input = scrub(last.content)\n    if (llm.output !== undefined) base.output = scrub(llm.output)\n  } else if (span.kind === 'tool') {\n    const tool = span as ToolSpan\n    base.input = asText(tool.args, scrub)\n    if (tool.result !== undefined) base.output = asText(tool.result, scrub)\n  }\n  return base\n}\n\n/** The final llm span's history + output is the completed conversation. */\nfunction finalConversation(spans: Span[], scrub: RolloutScrubber): ChatMessage[] {\n  const llms = spans.filter((s): s is LlmSpan => s.kind === 'llm')\n  const last = llms[llms.length - 1]\n  if (!last) return []\n  const messages: ChatMessage[] = last.messages.map((m: Message) => ({\n    role: m.role,\n    content: scrub(m.content),\n  }))\n  if (last.output !== undefined && last.output !== '') {\n    messages.push({ role: 'assistant', content: scrub(last.output) })\n  }\n  return messages\n}\n\n// The reward derivations live in the leaf module `./reward` so gate and\n// reporting code can import them without dragging in the trace store; they are\n// re-exported here because the derivations shipped from this path.\nexport {\n  isRealnessGated,\n  observedScore,\n  observedSplitScore,\n  type ScoreOrigin,\n  type ScorePreference,\n  scoreOrigin,\n  trainingReward,\n  trainingScore,\n} from './reward'\n\nconst REWARD_SOURCE: Record<ReturnType<typeof scoreOrigin>, string> = {\n  holdout: 'run-record/holdout-score',\n  search: 'run-record/search-score',\n  unscored: 'run-record/unscored',\n}\n\n/**\n * The mint door refuses an execution-only record: a missing training label is\n * not a zero reward, and not a mintable line either. Lines that already carry\n * `reward: null` — interchange imports, existing ledgers — stay valid on the\n * wire and keep their labeled gap; this guard is only about the\n * RunRecord→line door, where the producer can still be told to go score the\n * run instead of shipping an unlabeled row.\n */\nfunction requireTaskScore(record: RunRecord): void {\n  if (runTaskScore(record) === undefined) {\n    throw new ValidationError(`Cannot mint rollout for run ${record.runId}: task score is missing`)\n  }\n}\n\nconst SPLIT_FROM_TAG: Record<RunRecord['splitTag'], RolloutSplit> = {\n  search: 'search',\n  dev: 'dev',\n  holdout: 'holdout',\n}\n\nfunction mintLine(\n  record: RunRecord,\n  steps: RolloutStep[],\n  messages: ChatMessage[],\n  options: MintRolloutOptions,\n  capturedAt: string,\n  gap?: string,\n): MintedRolloutLine {\n  // A missing task score is refused before anything is built: an\n  // execution-only record has no training label, and a missing label is\n  // neither a zero reward nor a mintable row.\n  requireTaskScore(record)\n  // `reward` and `realness_gated` come out of one call, so neither door into\n  // the waist can write one and forget the other.\n  const rewardFields = rolloutRewardFields(record)\n  const uncaptured = record.costProvenance.kind === 'uncaptured'\n  const terminalOutcome = record.terminalOutcome\n  const isCompleted = terminalOutcome === 'succeeded' || terminalOutcome === 'failed'\n  const isTruncated = terminalOutcome === 'cancelled' || terminalOutcome === 'incomplete'\n  const terminalError =\n    terminalOutcome === 'failed' ||\n    terminalOutcome === 'cancelled' ||\n    terminalOutcome === 'incomplete'\n      ? (record.terminalFailureReason ?? `run ended ${terminalOutcome}`)\n      : null\n  // `assertMinted` rather than a cast: mint is the producer the whole gate\n  // rests on, so it proves the line it just built is valid instead of asserting\n  // it by fiat. The brand is unforgeable precisely because nobody casts to it.\n  return assertMinted(\n    {\n      schema: ROLLOUT_SCHEMA,\n      rollout_id: record.runId,\n      parent_rollout_id: null,\n      run_id: record.runId,\n      experiment_id: record.experimentId,\n      candidate_id: record.candidateId,\n      generation: null,\n      candidate_index: null,\n      role: options.role ?? 'agent',\n      task: {\n        suite: options.suite ?? record.experimentId,\n        instance_id: record.scenarioId,\n        split: SPLIT_FROM_TAG[record.splitTag],\n        seed: record.seed,\n        rep: 0,\n      },\n      policy: {\n        harness: null,\n        harness_version: null,\n        model: record.model,\n        provider: null,\n        profile_commit: record.commitSha,\n        prompt_hash: record.promptHash,\n        config_hash: record.configHash,\n        agent_profile_cell_id: record.agentProfile?.cellId ?? null,\n        sampling: null,\n      },\n      messages,\n      tool_defs: [],\n      ...(steps.length > 0 ? { steps } : {}),\n      outcome: {\n        ...rewardFields,\n        reward_source: REWARD_SOURCE[scoreOrigin(record)],\n        verdict: null,\n        // A verbatim bulk copy, deliberately UNFILTERED here. `outcome.raw`\n        // holds the per-layer verifier scores (`layer.*`) that the reward was\n        // derived from, so on a gated run this dict is the reward signal in\n        // component form — but filtering it at this call site is the pattern\n        // that has now leaked twice, because the next producer to write a\n        // reward-bearing field forgets. The gate is applied to the whole\n        // outcome once, in `assertMinted` below (`gateGamedOutcome`), which\n        // moves the block to `provenance.gated_evidence` when the run is gated\n        // and leaves it here untouched when it is not.\n        metrics: { ...record.outcome.raw },\n        is_completed: isCompleted,\n        is_truncated: isTruncated,\n        error: terminalError,\n      },\n      cost: {\n        usd: uncaptured ? null : record.costUsd,\n        tokens_in: record.tokenUsage.input,\n        tokens_out: record.tokenUsage.output,\n        tokens_reasoning: record.tokenUsage.reasoning ?? null,\n        cache_read: record.tokenUsage.cached ?? null,\n        cache_write: record.tokenUsage.cacheWrite ?? null,\n        wall_s: Math.round(record.wallMs / 1000),\n      },\n      artifacts: { patch_path: null, run_dir: null, transcript_ref: null },\n      provenance: {\n        captured_at: capturedAt,\n        capture: 'mint',\n        ...(gap !== undefined ? { gap } : {}),\n      },\n    },\n    `minted rollout line for run ${record.runId}`,\n  )\n}\n\n/**\n * Join RunRecords with their traces into canonical rollout lines. Records\n * without spans are emitted as labeled gap lines and reported in\n * `missingTraces`. Execution-only records without a task score are rejected\n * because a missing training label is not a zero reward.\n */\nexport async function mintRolloutRows(\n  records: RunRecord[],\n  store: TraceStore,\n  options: MintRolloutOptions = {},\n): Promise<MintRolloutResult> {\n  const scrub = options.scrub ?? ((t) => t)\n  const capturedAt = (options.now?.() ?? new Date()).toISOString()\n  const rows: MintedRolloutLine[] = []\n  const missingTraces: string[] = []\n  for (const record of records) {\n    const trajectory = await buildTrajectory(store, record.runId)\n    if (trajectory.steps.length === 0) {\n      missingTraces.push(record.runId)\n      rows.push(\n        mintLine(record, [], [], options, capturedAt, 'no trace spans recorded for this runId'),\n      )\n      continue\n    }\n    let steps = trajectory.steps.map((s) => projectStep(s.span, scrub))\n    if (options.maxSteps !== undefined && steps.length > options.maxSteps) {\n      // Keep the head and tail — the middle of a long run is the least\n      // informative for outcome attribution.\n      const head = Math.ceil(options.maxSteps / 2)\n      const tail = options.maxSteps - head\n      steps = [...steps.slice(0, head), ...steps.slice(steps.length - tail)]\n    }\n    const conversation = finalConversation(\n      trajectory.steps.map((s) => s.span),\n      scrub,\n    )\n    const gap =\n      conversation.length === 0 ? 'trace has no llm spans — no conversation to inline' : undefined\n    rows.push(mintLine(record, steps, conversation, options, capturedAt, gap))\n  }\n  return { rows, missingTraces }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqEA,MAAM,UAAU,GAAY,UAAmC;CAE7D,OAAO,OADG,OAAO,MAAM,WAAW,IAAI,KAAK,UAAU,CAAC,MACpC,EAAE;AACtB;AAEA,SAAS,YAAY,MAAY,OAAqC;CACpE,MAAM,OAAoB;EACxB,MAAM,KAAK;EACX,MAAM,MAAM,KAAK,IAAI;EACrB,QAAQ,KAAK;EACb,YAAY,KAAK,YAAY,KAAA,IAAY,KAAK,UAAU,KAAK,YAAY,KAAA;CAC3E;CACA,IAAI,KAAK,SAAS,OAAO;EACvB,MAAM,MAAM;EACZ,MAAM,OAAO,IAAI,SAAS,IAAI,SAAS,SAAS;EAChD,IAAI,MAAM,KAAK,QAAQ,MAAM,KAAK,OAAO;EACzC,IAAI,IAAI,WAAW,KAAA,GAAW,KAAK,SAAS,MAAM,IAAI,MAAM;CAC9D,OAAO,IAAI,KAAK,SAAS,QAAQ;EAC/B,MAAM,OAAO;EACb,KAAK,QAAQ,OAAO,KAAK,MAAM,KAAK;EACpC,IAAI,KAAK,WAAW,KAAA,GAAW,KAAK,SAAS,OAAO,KAAK,QAAQ,KAAK;CACxE;CACA,OAAO;AACT;;AAGA,SAAS,kBAAkB,OAAe,OAAuC;CAC/E,MAAM,OAAO,MAAM,QAAQ,MAAoB,EAAE,SAAS,KAAK;CAC/D,MAAM,OAAO,KAAK,KAAK,SAAS;CAChC,IAAI,CAAC,MAAM,OAAO,CAAC;CACnB,MAAM,WAA0B,KAAK,SAAS,KAAK,OAAgB;EACjE,MAAM,EAAE;EACR,SAAS,MAAM,EAAE,OAAO;CAC1B,EAAE;CACF,IAAI,KAAK,WAAW,KAAA,KAAa,KAAK,WAAW,IAC/C,SAAS,KAAK;EAAE,MAAM;EAAa,SAAS,MAAM,KAAK,MAAM;CAAE,CAAC;CAElE,OAAO;AACT;AAgBA,MAAM,gBAAgE;CACpE,SAAS;CACT,QAAQ;CACR,UAAU;AACZ;;;;;;;;;AAUA,SAAS,iBAAiB,QAAyB;CACjD,IAAI,aAAa,MAAM,MAAM,KAAA,GAC3B,MAAM,IAAI,gBAAgB,+BAA+B,OAAO,MAAM,wBAAwB;AAElG;AAEA,MAAM,iBAA8D;CAClE,QAAQ;CACR,KAAK;CACL,SAAS;AACX;AAEA,SAAS,SACP,QACA,OACA,UACA,SACA,YACA,KACmB;CAInB,iBAAiB,MAAM;CAGvB,MAAM,eAAe,oBAAoB,MAAM;CAC/C,MAAM,aAAa,OAAO,eAAe,SAAS;CAClD,MAAM,kBAAkB,OAAO;CAC/B,MAAM,cAAc,oBAAoB,eAAe,oBAAoB;CAC3E,MAAM,cAAc,oBAAoB,eAAe,oBAAoB;CAC3E,MAAM,gBACJ,oBAAoB,YACpB,oBAAoB,eACpB,oBAAoB,eACf,OAAO,yBAAyB,aAAa,oBAC9C;CAIN,OAAO,aACL;EACE,QAAQ;EACR,YAAY,OAAO;EACnB,mBAAmB;EACnB,QAAQ,OAAO;EACf,eAAe,OAAO;EACtB,cAAc,OAAO;EACrB,YAAY;EACZ,iBAAiB;EACjB,MAAM,QAAQ,QAAQ;EACtB,MAAM;GACJ,OAAO,QAAQ,SAAS,OAAO;GAC/B,aAAa,OAAO;GACpB,OAAO,eAAe,OAAO;GAC7B,MAAM,OAAO;GACb,KAAK;EACP;EACA,QAAQ;GACN,SAAS;GACT,iBAAiB;GACjB,OAAO,OAAO;GACd,UAAU;GACV,gBAAgB,OAAO;GACvB,aAAa,OAAO;GACpB,aAAa,OAAO;GACpB,uBAAuB,OAAO,cAAc,UAAU;GACtD,UAAU;EACZ;EACA;EACA,WAAW,CAAC;EACZ,GAAI,MAAM,SAAS,IAAI,EAAE,MAAM,IAAI,CAAC;EACpC,SAAS;GACP,GAAG;GACH,eAAe,cAAc,YAAY,MAAM;GAC/C,SAAS;GAUT,SAAS,EAAE,GAAG,OAAO,QAAQ,IAAI;GACjC,cAAc;GACd,cAAc;GACd,OAAO;EACT;EACA,MAAM;GACJ,KAAK,aAAa,OAAO,OAAO;GAChC,WAAW,OAAO,WAAW;GAC7B,YAAY,OAAO,WAAW;GAC9B,kBAAkB,OAAO,WAAW,aAAa;GACjD,YAAY,OAAO,WAAW,UAAU;GACxC,aAAa,OAAO,WAAW,cAAc;GAC7C,QAAQ,KAAK,MAAM,OAAO,SAAS,GAAI;EACzC;EACA,WAAW;GAAE,YAAY;GAAM,SAAS;GAAM,gBAAgB;EAAK;EACnE,YAAY;GACV,aAAa;GACb,SAAS;GACT,GAAI,QAAQ,KAAA,IAAY,EAAE,IAAI,IAAI,CAAC;EACrC;CACF,GACA,+BAA+B,OAAO,OACxC;AACF;;;;;;;AAQA,eAAsB,gBACpB,SACA,OACA,UAA8B,CAAC,GACH;CAC5B,MAAM,QAAQ,QAAQ,WAAW,MAAM;CACvC,MAAM,cAAc,QAAQ,MAAM,qBAAK,IAAI,KAAK,EAAA,CAAG,YAAY;CAC/D,MAAM,OAA4B,CAAC;CACnC,MAAM,gBAA0B,CAAC;CACjC,KAAK,MAAM,UAAU,SAAS;EAC5B,MAAM,aAAa,MAAM,gBAAgB,OAAO,OAAO,KAAK;EAC5D,IAAI,WAAW,MAAM,WAAW,GAAG;GACjC,cAAc,KAAK,OAAO,KAAK;GAC/B,KAAK,KACH,SAAS,QAAQ,CAAC,GAAG,CAAC,GAAG,SAAS,YAAY,wCAAwC,CACxF;GACA;EACF;EACA,IAAI,QAAQ,WAAW,MAAM,KAAK,MAAM,YAAY,EAAE,MAAM,KAAK,CAAC;EAClE,IAAI,QAAQ,aAAa,KAAA,KAAa,MAAM,SAAS,QAAQ,UAAU;GAGrE,MAAM,OAAO,KAAK,KAAK,QAAQ,WAAW,CAAC;GAC3C,MAAM,OAAO,QAAQ,WAAW;GAChC,QAAQ,CAAC,GAAG,MAAM,MAAM,GAAG,IAAI,GAAG,GAAG,MAAM,MAAM,MAAM,SAAS,IAAI,CAAC;EACvE;EACA,MAAM,eAAe,kBACnB,WAAW,MAAM,KAAK,MAAM,EAAE,IAAI,GAClC,KACF;EACA,MAAM,MACJ,aAAa,WAAW,IAAI,uDAAuD,KAAA;EACrF,KAAK,KAAK,SAAS,QAAQ,OAAO,cAAc,SAAS,YAAY,GAAG,CAAC;CAC3E;CACA,OAAO;EAAE;EAAM;CAAc;AAC/B"}

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.133.0",
+    "version": "0.133.2",
     "description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.0.0. Bumps on breaking changes to request/response schemas.",
     "contact": {
       "name": "Tangle Network",

package/dist/{opencode-sqlite-BGrHeDu3.js → opencode-sqlite-8r6WUfHc.js}

@@ -187,11 +187,10 @@ function transcriptFromEntries(entries, options = {}) {
 * throwing — callers record a gap line, never crash the backfill.
 */
 const DEFAULT_OPENCODE_DB = join(homedir(), ".local", "share", "opencode", "opencode.db");
-const nodeRequire = createRequire(import.meta.url);
 /** Open the store read-only; null = unavailable/corrupt (caller records a gap). */
 async function openOpencodeDb(path = DEFAULT_OPENCODE_DB) {
 	try {
-		const { DatabaseSync } = nodeRequire("node:sqlite");
+		const { DatabaseSync } = createRequire(import.meta.url)("node:sqlite");
 		const db = new DatabaseSync(path, { readOnly: true });
 		db.prepare("SELECT id FROM session LIMIT 1").get();
 		return db;
@@ -315,4 +314,4 @@ function readOpencodeSessionMessages(db, sessionId) {
 //#endregion
 export { readOpencodeSessionMessages as a, findClaudeTranscripts as c, transcriptFromEntries as d, openOpencodeDb as i, parseClaudeEntries as l, findOpencodeSessionById as n, DEFAULT_CLAUDE_PROJECTS_DIR as o, findOpencodeSessionsByDirectory as r, claudeProjectSlug as s, DEFAULT_OPENCODE_DB as t, readClaudeTranscript as u };
 
-//# sourceMappingURL=opencode-sqlite-BGrHeDu3.js.map
+//# sourceMappingURL=opencode-sqlite-8r6WUfHc.js.map

package/dist/opencode-sqlite-8r6WUfHc.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"opencode-sqlite-8r6WUfHc.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\nexport function findOpencodeSessionById(\n  db: DatabaseSync,\n  sessionId: string,\n): OpencodeSessionRow | null {\n  const row = db.prepare(`SELECT ${SESSION_COLUMNS} FROM session WHERE id = ?`).get(sessionId) as\n    | Record<string, unknown>\n    | undefined\n  return row === undefined ? null : parseSessionRow(row)\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;AAEA,SAAgB,wBACd,IACA,WAC2B;CAC3B,MAAM,MAAM,GAAG,QAAQ,UAAU,gBAAgB,2BAA2B,CAAC,CAAC,IAAI,SAAS;CAG3F,OAAO,QAAQ,KAAA,IAAY,OAAO,gBAAgB,GAAG;AACvD;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/{paired-arms-D9D0wXj2.js → paired-arms-6XItKzd1.js}

@@ -1,5 +1,5 @@
 import { s as ValidationError } from "./errors-8YnH8WlF.js";
-import { N as wilcoxonSignedRank, p as mcnemar, v as pairedBootstrap, x as pairedRiskDifference } from "./statistics-CnnxdpOg.js";
+import { N as wilcoxonSignedRank, p as mcnemar, v as pairedBootstrap, x as pairedRiskDifference } from "./statistics-DWM_AyLe.js";
 //#region src/paired-arms.ts
 /**
 * Matched-pair arm comparison — "did the treatment arm beat the baseline arm
@@ -257,4 +257,4 @@ function metricValue(row, name) {
 //#endregion
 export { pairArms as n, pairRunRecords as r, comparePairedArms as t };
 
-//# sourceMappingURL=paired-arms-D9D0wXj2.js.map
+//# sourceMappingURL=paired-arms-6XItKzd1.js.map

package/dist/{paired-arms-D9D0wXj2.js.map → paired-arms-6XItKzd1.js.map}

@@ -1 +1 @@
-{"version":3,"file":"paired-arms-D9D0wXj2.js","names":[],"sources":["../src/paired-arms.ts"],"sourcesContent":["/**\n * Matched-pair arm comparison — \"did the treatment arm beat the baseline arm\n * on the SAME work items?\"\n *\n * An arm A/B over run records is only trustworthy when it is PAIRED: the same\n * task/scenario/seed evaluated under both arms, compared item-by-item, so\n * inter-item difficulty variance cancels instead of masquerading as an arm\n * effect. This module owns the two error-prone steps every consumer otherwise\n * hand-rolls:\n *\n *   1. Pairing — matching rows across arms by `pairKey` (and by `repKey`\n *      within multi-rep items), with leftovers REPORTED rather than silently\n *      dropped (a silently unbalanced pairing biases every paired statistic\n *      downstream). Pairing never keys on outcome content: matching reps by\n *      their outcomes deflates discordant-pair counts and makes McNemar\n *      anti-conservative, so reps pair only by (`pairKey`, `repKey`) identity.\n *   2. Composition — feeding the matched pairs to the correct paired\n *      estimators that already live in `statistics`: `mcnemar` +\n *      `pairedRiskDifference` for pass/fail, `pairedBootstrap` +\n *      `wilcoxonSignedRank` for continuous metrics. No statistic is\n *      re-implemented here.\n *\n * The row shape is deliberately structural — callers project a `RunRecord`\n * (or any record) into `{ pairKey, arm, pass?, metrics? }`. Arm names are\n * caller-supplied parameters; the module ships no domain literal.\n */\n\nimport { ValidationError } from './errors'\nimport type { RunRecord } from './run-record'\nimport type { McNemarResult, PairedBootstrapOptions, PairedBootstrapResult } from './statistics'\nimport {\n  mcnemar,\n  pairedBootstrap,\n  pairedRiskDifference,\n  type RiskDifferenceResult,\n  wilcoxonSignedRank,\n} from './statistics'\n\n/** One arm observation of one work item. Structural on purpose: callers\n *  project their own record type (e.g. a `RunRecord`) into this shape. */\nexport interface PairedArmRow {\n  /** Matching key — rows sharing a `pairKey` across both arms form pairs\n   *  (typically the task/scenario/seed identity). */\n  pairKey: string\n  /** Rep identity within a `pairKey` (e.g. a seed or rep number). Required on\n   *  every row of a `pairKey` that has more than one rep in either arm; reps\n   *  then pair only on exact (`pairKey`, `repKey`) match, never on outcome\n   *  content. Optional when each arm has at most one rep of the item. */\n  repKey?: string\n  /** Arm label this row was produced under. */\n  arm: string\n  /** Binary outcome; omit when the comparison has no pass/fail notion. */\n  pass?: boolean\n  /** Named numeric measurements (score, cost, latency, …). */\n  metrics?: Record<string, number>\n}\n\nexport interface PairArmsOptions {\n  /** Arm treated as the control side of every pair. */\n  baselineArm: string\n  /** Arm treated as the treatment side of every pair. */\n  treatmentArm: string\n}\n\n/** One matched (baseline, treatment) observation of the same work item. */\nexport interface MatchedPair {\n  pairKey: string\n  /** 0-based position of this pair within its `pairKey`, ordered by sorted\n   *  `repKey` (always 0 for a single-rep item). The rep identity itself is on\n   *  the rows (`baseline.repKey` / `treatment.repKey`). */\n  repIndex: number\n  baseline: PairedArmRow\n  treatment: PairedArmRow\n}\n\nexport interface PairArmsResult {\n  /** Matched pairs, ordered by (`pairKey`, `repIndex`). */\n  pairs: MatchedPair[]\n  /** Baseline rows left without a treatment counterpart — reported, never\n   *  silently dropped. */\n  unpairedBaseline: PairedArmRow[]\n  /** Treatment rows left without a baseline counterpart. */\n  unpairedTreatment: PairedArmRow[]\n}\n\n/**\n * Match rows across two arms into (baseline, treatment) pairs by `pairKey`.\n *\n * A `pairKey` with at most one row per arm pairs directly, no `repKey`\n * needed. A `pairKey` with multiple reps in either arm requires `repKey` on\n * every one of its rows, and reps pair only on exact (`pairKey`, `repKey`)\n * match — pairing is keyed purely on row identity, never on outcome content\n * (outcome-keyed matching deflates discordant counts and biases McNemar), and\n * is therefore independent of input order. Reps whose `repKey` has no\n * counterpart in the other arm, and items present in only one arm, land in\n * the unpaired lists — reported, never truncated.\n *\n * Fail-loud: throws when either named arm has zero rows (an unknown arm\n * name would otherwise read as \"everything unpaired\"), when the two arm\n * names are equal, when a multi-rep `pairKey` has a row without `repKey`, or\n * when a (`pairKey`, arm) group repeats a `repKey` (the match would be\n * ambiguous).\n */\nexport function pairArms(rows: readonly PairedArmRow[], opts: PairArmsOptions): PairArmsResult {\n  const { baselineArm, treatmentArm } = opts\n  if (baselineArm === treatmentArm) {\n    throw new ValidationError(\n      `pairArms: baselineArm and treatmentArm are both '${baselineArm}' — an arm cannot be compared to itself`,\n    )\n  }\n\n  // arm → pairKey → rows\n  const byArm = new Map<string, Map<string, PairedArmRow[]>>()\n  const armsSeen = new Set<string>()\n  for (const row of rows) {\n    armsSeen.add(row.arm)\n    if (row.arm !== baselineArm && row.arm !== treatmentArm) continue\n    const byKey = byArm.get(row.arm) ?? new Map<string, PairedArmRow[]>()\n    const group = byKey.get(row.pairKey) ?? []\n    group.push(row)\n    byKey.set(row.pairKey, group)\n    byArm.set(row.arm, byKey)\n  }\n\n  for (const arm of [baselineArm, treatmentArm]) {\n    if (!byArm.has(arm)) {\n      const seen = [...armsSeen].sort().join(', ') || '<none>'\n      throw new ValidationError(`pairArms: no rows for arm '${arm}' (arms present: ${seen})`)\n    }\n  }\n\n  const baselineByKey = byArm.get(baselineArm)!\n  const treatmentByKey = byArm.get(treatmentArm)!\n\n  const allKeys = [...new Set([...baselineByKey.keys(), ...treatmentByKey.keys()])].sort()\n  const pairs: MatchedPair[] = []\n  const unpairedBaseline: PairedArmRow[] = []\n  const unpairedTreatment: PairedArmRow[] = []\n  for (const pairKey of allKeys) {\n    const b = baselineByKey.get(pairKey) ?? []\n    const t = treatmentByKey.get(pairKey) ?? []\n\n    if (b.length <= 1 && t.length <= 1) {\n      if (b.length === 1 && t.length === 1) {\n        const baseline = b[0]!\n        const treatment = t[0]!\n        if (baseline.repKey !== undefined || treatment.repKey !== undefined) {\n          if (\n            baseline.repKey === undefined ||\n            treatment.repKey === undefined ||\n            baseline.repKey !== treatment.repKey\n          ) {\n            unpairedBaseline.push(baseline)\n            unpairedTreatment.push(treatment)\n            continue\n          }\n        }\n        pairs.push({ pairKey, repIndex: 0, baseline, treatment })\n      } else {\n        unpairedBaseline.push(...b)\n        unpairedTreatment.push(...t)\n      }\n      continue\n    }\n\n    const bByRep = indexByRepKey(b, pairKey, baselineArm)\n    const tByRep = indexByRepKey(t, pairKey, treatmentArm)\n    const repKeys = [...new Set([...bByRep.keys(), ...tByRep.keys()])].sort()\n    let repIndex = 0\n    for (const repKey of repKeys) {\n      const baseline = bByRep.get(repKey)\n      const treatment = tByRep.get(repKey)\n      if (baseline !== undefined && treatment !== undefined) {\n        pairs.push({ pairKey, repIndex: repIndex++, baseline, treatment })\n      } else if (baseline !== undefined) {\n        unpairedBaseline.push(baseline)\n      } else if (treatment !== undefined) {\n        unpairedTreatment.push(treatment)\n      }\n    }\n  }\n\n  return { pairs, unpairedBaseline, unpairedTreatment }\n}\n\n/** Index a multi-rep (pairKey, arm) group by `repKey`, enforcing that every\n *  row carries one and that no repKey repeats within the group. */\nfunction indexByRepKey(\n  group: readonly PairedArmRow[],\n  pairKey: string,\n  arm: string,\n): Map<string, PairedArmRow> {\n  const byRep = new Map<string, PairedArmRow>()\n  for (const row of group) {\n    if (row.repKey === undefined) {\n      throw new ValidationError(\n        `pairArms: pairKey '${pairKey}' has multiple reps in an arm, but a row in arm '${arm}' ` +\n          `is missing repKey — multi-rep items require an explicit repKey on every row so reps ` +\n          `pair by identity (pairing reps by outcome or by index would bias the paired statistics)`,\n      )\n    }\n    if (byRep.has(row.repKey)) {\n      throw new ValidationError(\n        `pairArms: duplicate repKey '${row.repKey}' for pairKey '${pairKey}' in arm '${arm}' — ` +\n          `(pairKey, repKey) must uniquely identify a rep within an arm`,\n      )\n    }\n    byRep.set(row.repKey, row)\n  }\n  return byRep\n}\n\n/** Paired pass/fail comparison over the pairs where BOTH sides carry `pass`. */\nexport interface PairedCorrectness {\n  /** Discordant pairs where the treatment passed and the baseline failed. */\n  b10: number\n  /** Discordant pairs where the baseline passed and the treatment failed. */\n  b01: number\n  /** Exact McNemar significance over the paired outcomes (`b === b10`, `c === b01`). */\n  mcnemar: McNemarResult\n  /** Paired effect size: p(treatment) − p(baseline) with a paired-variance CI. */\n  riskDifference: RiskDifferenceResult\n}\n\n/** Paired delta summary for one named metric (delta = treatment − baseline). */\nexport interface PairedMetricDelta {\n  name: string\n  /** Pairs where BOTH sides carry a finite value for this metric. */\n  n: number\n  /** Pairs where at least one side does not carry the metric. */\n  nMissing: number\n  /** Median paired delta, or null when `n === 0`. */\n  medianDelta: number | null\n  /** Mean paired delta, or null when `n === 0`. */\n  meanDelta: number | null\n  /** Bootstrap CI on the paired delta (`pairedBootstrap`); null when\n   *  `n === 0` — a zero-width [0, 0] interval on no data would read as a\n   *  measured tight null. */\n  bootstrapCi: PairedBootstrapResult | null\n  /** Wilcoxon signed-rank test on the paired deltas; null when `n === 0`. */\n  wilcoxon: { w: number; p: number } | null\n}\n\nexport interface ComparePairedArmsOptions extends PairArmsOptions {\n  /** Metrics to compare. Default: every metric name observed on any matched\n   *  pair, sorted. A name that appears on no pair is still reported (with\n   *  `n = 0`) so a misspelled metric is visible instead of vanishing. */\n  metricNames?: string[]\n  /** Passed through to `pairedBootstrap` — set `seed` for reproducible CIs. */\n  bootstrap?: PairedBootstrapOptions\n}\n\nexport interface PairedArmsComparison {\n  nPairs: number\n  nUnpairedBaseline: number\n  nUnpairedTreatment: number\n  /** null when no matched pair carries `pass` on both sides — a pass/fail\n   *  verdict over rows that never measured pass/fail would be fabricated. */\n  correctness: PairedCorrectness | null\n  metricDeltas: PairedMetricDelta[]\n}\n\n/**\n * Full matched-pair arm comparison: pair via {@link pairArms}, then compose\n * the paired estimators from `statistics` over the matched pairs.\n *\n * Correctness uses only the pairs where both sides carry `pass` (`mcnemar.n`\n * is that subset's size); each metric uses only the pairs where both sides\n * carry a finite value for it, with the remainder counted in `nMissing`.\n * Deltas are treatment − baseline throughout.\n *\n * Fail-loud: inherits {@link pairArms}'s unknown-arm throw, and throws on a\n * non-finite metric value — silently treating corrupt telemetry as \"metric\n * absent\" would misreport it as missing coverage.\n */\nexport function comparePairedArms(\n  rows: readonly PairedArmRow[],\n  opts: ComparePairedArmsOptions,\n): PairedArmsComparison {\n  const { pairs, unpairedBaseline, unpairedTreatment } = pairArms(rows, opts)\n\n  let correctness: PairedCorrectness | null = null\n  const baselinePass: number[] = []\n  const treatmentPass: number[] = []\n  for (const pair of pairs) {\n    if (pair.baseline.pass === undefined || pair.treatment.pass === undefined) continue\n    baselinePass.push(pair.baseline.pass ? 1 : 0)\n    treatmentPass.push(pair.treatment.pass ? 1 : 0)\n  }\n  if (baselinePass.length > 0) {\n    const mc = mcnemar(baselinePass, treatmentPass)\n    correctness = {\n      b10: mc.b,\n      b01: mc.c,\n      mcnemar: mc,\n      riskDifference: pairedRiskDifference(baselinePass, treatmentPass),\n    }\n  }\n\n  const metricNames =\n    opts.metricNames ??\n    [\n      ...new Set(\n        pairs.flatMap((p) => [\n          ...Object.keys(p.baseline.metrics ?? {}),\n          ...Object.keys(p.treatment.metrics ?? {}),\n        ]),\n      ),\n    ].sort()\n\n  const metricDeltas: PairedMetricDelta[] = metricNames.map((name) => {\n    const before: number[] = []\n    const after: number[] = []\n    let nMissing = 0\n    for (const pair of pairs) {\n      const b = metricValue(pair.baseline, name)\n      const t = metricValue(pair.treatment, name)\n      if (b === undefined || t === undefined) {\n        nMissing++\n        continue\n      }\n      before.push(b)\n      after.push(t)\n    }\n    const bootstrapCi = before.length === 0 ? null : pairedBootstrap(before, after, opts.bootstrap)\n    return {\n      name,\n      n: before.length,\n      nMissing,\n      medianDelta: bootstrapCi?.median ?? null,\n      meanDelta: bootstrapCi?.mean ?? null,\n      bootstrapCi,\n      wilcoxon: before.length === 0 ? null : wilcoxonSignedRank(before, after),\n    }\n  })\n\n  return {\n    nPairs: pairs.length,\n    nUnpairedBaseline: unpairedBaseline.length,\n    nUnpairedTreatment: unpairedTreatment.length,\n    correctness,\n    metricDeltas,\n  }\n}\n\nexport interface MatchedRunRecordPair {\n  pairKey: string\n  repKey: string\n  baseline: RunRecord\n  treatment: RunRecord\n}\n\nexport interface PairRunRecordsResult {\n  pairs: MatchedRunRecordPair[]\n  unpairedBaseline: RunRecord[]\n  unpairedTreatment: RunRecord[]\n}\n\ninterface RunRecordArmRow extends PairedArmRow {\n  run: RunRecord\n  repKey: string\n}\n\n/**\n * Pair two RunRecord arms by the identity of the evaluated work:\n * `(experimentId, scenarioId, seed)`.\n *\n * Falling back to array order, candidate id, or experiment id can compare\n * different tasks and fabricate lift. Duplicate identities throw.\n */\nexport function pairRunRecords(\n  baselineRuns: readonly RunRecord[],\n  treatmentRuns: readonly RunRecord[],\n): PairRunRecordsResult {\n  const baselineRows = runRecordArmRows(baselineRuns, 'baseline')\n  const treatmentRows = runRecordArmRows(treatmentRuns, 'treatment')\n  validateRunRecordArmRows(baselineRows, 'baseline')\n  validateRunRecordArmRows(treatmentRows, 'treatment')\n  if (baselineRows.length === 0 || treatmentRows.length === 0) {\n    return {\n      pairs: [],\n      unpairedBaseline: baselineRows.map((row) => row.run),\n      unpairedTreatment: treatmentRows.map((row) => row.run),\n    }\n  }\n\n  const result = pairArms([...baselineRows, ...treatmentRows], {\n    baselineArm: 'baseline',\n    treatmentArm: 'treatment',\n  })\n  return {\n    pairs: result.pairs.map((pair) => {\n      const baseline = pair.baseline as RunRecordArmRow\n      const treatment = pair.treatment as RunRecordArmRow\n      return {\n        pairKey: pair.pairKey,\n        repKey: baseline.repKey,\n        baseline: baseline.run,\n        treatment: treatment.run,\n      }\n    }),\n    unpairedBaseline: result.unpairedBaseline.map((row) => (row as RunRecordArmRow).run),\n    unpairedTreatment: result.unpairedTreatment.map((row) => (row as RunRecordArmRow).run),\n  }\n}\n\nfunction runRecordArmRows(runs: readonly RunRecord[], arm: string): RunRecordArmRow[] {\n  return runs.map((run) => {\n    const scenarioId = run.scenarioId.trim()\n    if (!scenarioId) {\n      throw new ValidationError(\n        `pairRunRecords: run '${run.runId}' is missing scenarioId; paired comparisons require explicit scenario identity`,\n      )\n    }\n    return {\n      pairKey: JSON.stringify([run.experimentId, scenarioId]),\n      repKey: String(run.seed),\n      arm,\n      run,\n    }\n  })\n}\n\nfunction validateRunRecordArmRows(rows: readonly RunRecordArmRow[], arm: string): void {\n  const byPairKey = new Map<string, RunRecordArmRow[]>()\n  for (const row of rows) {\n    const group = byPairKey.get(row.pairKey) ?? []\n    group.push(row)\n    byPairKey.set(row.pairKey, group)\n  }\n  for (const [pairKey, group] of byPairKey) {\n    if (group.length > 1) indexByRepKey(group, pairKey, arm)\n  }\n}\n\nfunction metricValue(row: PairedArmRow, name: string): number | undefined {\n  const v = row.metrics?.[name]\n  if (v === undefined) return undefined\n  if (!Number.isFinite(v)) {\n    throw new ValidationError(\n      `comparePairedArms: non-finite value for metric '${name}' on pairKey '${row.pairKey}' (arm '${row.arm}'): ${v}`,\n    )\n  }\n  return v\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuGA,SAAgB,SAAS,MAA+B,MAAuC;CAC7F,MAAM,EAAE,aAAa,iBAAiB;CACtC,IAAI,gBAAgB,cAClB,MAAM,IAAI,gBACR,oDAAoD,YAAY,wCAClE;CAIF,MAAM,wBAAQ,IAAI,IAAyC;CAC3D,MAAM,2BAAW,IAAI,IAAY;CACjC,KAAK,MAAM,OAAO,MAAM;EACtB,SAAS,IAAI,IAAI,GAAG;EACpB,IAAI,IAAI,QAAQ,eAAe,IAAI,QAAQ,cAAc;EACzD,MAAM,QAAQ,MAAM,IAAI,IAAI,GAAG,qBAAK,IAAI,IAA4B;EACpE,MAAM,QAAQ,MAAM,IAAI,IAAI,OAAO,KAAK,CAAC;EACzC,MAAM,KAAK,GAAG;EACd,MAAM,IAAI,IAAI,SAAS,KAAK;EAC5B,MAAM,IAAI,IAAI,KAAK,KAAK;CAC1B;CAEA,KAAK,MAAM,OAAO,CAAC,aAAa,YAAY,GAC1C,IAAI,CAAC,MAAM,IAAI,GAAG,GAEhB,MAAM,IAAI,gBAAgB,8BAA8B,IAAI,mBAD/C,CAAC,GAAG,QAAQ,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,IAAI,KAAK,SACoC,EAAE;CAI1F,MAAM,gBAAgB,MAAM,IAAI,WAAW;CAC3C,MAAM,iBAAiB,MAAM,IAAI,YAAY;CAE7C,MAAM,UAAU,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,cAAc,KAAK,GAAG,GAAG,eAAe,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK;CACvF,MAAM,QAAuB,CAAC;CAC9B,MAAM,mBAAmC,CAAC;CAC1C,MAAM,oBAAoC,CAAC;CAC3C,KAAK,MAAM,WAAW,SAAS;EAC7B,MAAM,IAAI,cAAc,IAAI,OAAO,KAAK,CAAC;EACzC,MAAM,IAAI,eAAe,IAAI,OAAO,KAAK,CAAC;EAE1C,IAAI,EAAE,UAAU,KAAK,EAAE,UAAU,GAAG;GAClC,IAAI,EAAE,WAAW,KAAK,EAAE,WAAW,GAAG;IACpC,MAAM,WAAW,EAAE;IACnB,MAAM,YAAY,EAAE;IACpB,IAAI,SAAS,WAAW,KAAA,KAAa,UAAU,WAAW,KAAA,GAEtD;SAAA,SAAS,WAAW,KAAA,KACpB,UAAU,WAAW,KAAA,KACrB,SAAS,WAAW,UAAU,QAC9B;MACA,iBAAiB,KAAK,QAAQ;MAC9B,kBAAkB,KAAK,SAAS;MAChC;KACF;;IAEF,MAAM,KAAK;KAAE;KAAS,UAAU;KAAG;KAAU;IAAU,CAAC;GAC1D,OAAO;IACL,iBAAiB,KAAK,GAAG,CAAC;IAC1B,kBAAkB,KAAK,GAAG,CAAC;GAC7B;GACA;EACF;EAEA,MAAM,SAAS,cAAc,GAAG,SAAS,WAAW;EACpD,MAAM,SAAS,cAAc,GAAG,SAAS,YAAY;EACrD,MAAM,UAAU,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,OAAO,KAAK,GAAG,GAAG,OAAO,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK;EACxE,IAAI,WAAW;EACf,KAAK,MAAM,UAAU,SAAS;GAC5B,MAAM,WAAW,OAAO,IAAI,MAAM;GAClC,MAAM,YAAY,OAAO,IAAI,MAAM;GACnC,IAAI,aAAa,KAAA,KAAa,cAAc,KAAA,GAC1C,MAAM,KAAK;IAAE;IAAS,UAAU;IAAY;IAAU;GAAU,CAAC;QAC5D,IAAI,aAAa,KAAA,GACtB,iBAAiB,KAAK,QAAQ;QACzB,IAAI,cAAc,KAAA,GACvB,kBAAkB,KAAK,SAAS;EAEpC;CACF;CAEA,OAAO;EAAE;EAAO;EAAkB;CAAkB;AACtD;;;AAIA,SAAS,cACP,OACA,SACA,KAC2B;CAC3B,MAAM,wBAAQ,IAAI,IAA0B;CAC5C,KAAK,MAAM,OAAO,OAAO;EACvB,IAAI,IAAI,WAAW,KAAA,GACjB,MAAM,IAAI,gBACR,sBAAsB,QAAQ,mDAAmD,IAAI,8KAGvF;EAEF,IAAI,MAAM,IAAI,IAAI,MAAM,GACtB,MAAM,IAAI,gBACR,+BAA+B,IAAI,OAAO,iBAAiB,QAAQ,YAAY,IAAI,iEAErF;EAEF,MAAM,IAAI,IAAI,QAAQ,GAAG;CAC3B;CACA,OAAO;AACT;;;;;;;;;;;;;;AAiEA,SAAgB,kBACd,MACA,MACsB;CACtB,MAAM,EAAE,OAAO,kBAAkB,sBAAsB,SAAS,MAAM,IAAI;CAE1E,IAAI,cAAwC;CAC5C,MAAM,eAAyB,CAAC;CAChC,MAAM,gBAA0B,CAAC;CACjC,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,KAAK,SAAS,SAAS,KAAA,KAAa,KAAK,UAAU,SAAS,KAAA,GAAW;EAC3E,aAAa,KAAK,KAAK,SAAS,OAAO,IAAI,CAAC;EAC5C,cAAc,KAAK,KAAK,UAAU,OAAO,IAAI,CAAC;CAChD;CACA,IAAI,aAAa,SAAS,GAAG;EAC3B,MAAM,KAAK,QAAQ,cAAc,aAAa;EAC9C,cAAc;GACZ,KAAK,GAAG;GACR,KAAK,GAAG;GACR,SAAS;GACT,gBAAgB,qBAAqB,cAAc,aAAa;EAClE;CACF;CAaA,MAAM,gBAVJ,KAAK,eACL,CACE,GAAG,IAAI,IACL,MAAM,SAAS,MAAM,CACnB,GAAG,OAAO,KAAK,EAAE,SAAS,WAAW,CAAC,CAAC,GACvC,GAAG,OAAO,KAAK,EAAE,UAAU,WAAW,CAAC,CAAC,CAC1C,CAAC,CACH,CACF,CAAC,CAAC,KAAK,EAAA,CAE6C,KAAK,SAAS;EAClE,MAAM,SAAmB,CAAC;EAC1B,MAAM,QAAkB,CAAC;EACzB,IAAI,WAAW;EACf,KAAK,MAAM,QAAQ,OAAO;GACxB,MAAM,IAAI,YAAY,KAAK,UAAU,IAAI;GACzC,MAAM,IAAI,YAAY,KAAK,WAAW,IAAI;GAC1C,IAAI,MAAM,KAAA,KAAa,MAAM,KAAA,GAAW;IACtC;IACA;GACF;GACA,OAAO,KAAK,CAAC;GACb,MAAM,KAAK,CAAC;EACd;EACA,MAAM,cAAc,OAAO,WAAW,IAAI,OAAO,gBAAgB,QAAQ,OAAO,KAAK,SAAS;EAC9F,OAAO;GACL;GACA,GAAG,OAAO;GACV;GACA,aAAa,aAAa,UAAU;GACpC,WAAW,aAAa,QAAQ;GAChC;GACA,UAAU,OAAO,WAAW,IAAI,OAAO,mBAAmB,QAAQ,KAAK;EACzE;CACF,CAAC;CAED,OAAO;EACL,QAAQ,MAAM;EACd,mBAAmB,iBAAiB;EACpC,oBAAoB,kBAAkB;EACtC;EACA;CACF;AACF;;;;;;;;AA2BA,SAAgB,eACd,cACA,eACsB;CACtB,MAAM,eAAe,iBAAiB,cAAc,UAAU;CAC9D,MAAM,gBAAgB,iBAAiB,eAAe,WAAW;CACjE,yBAAyB,cAAc,UAAU;CACjD,yBAAyB,eAAe,WAAW;CACnD,IAAI,aAAa,WAAW,KAAK,cAAc,WAAW,GACxD,OAAO;EACL,OAAO,CAAC;EACR,kBAAkB,aAAa,KAAK,QAAQ,IAAI,GAAG;EACnD,mBAAmB,cAAc,KAAK,QAAQ,IAAI,GAAG;CACvD;CAGF,MAAM,SAAS,SAAS,CAAC,GAAG,cAAc,GAAG,aAAa,GAAG;EAC3D,aAAa;EACb,cAAc;CAChB,CAAC;CACD,OAAO;EACL,OAAO,OAAO,MAAM,KAAK,SAAS;GAChC,MAAM,WAAW,KAAK;GACtB,MAAM,YAAY,KAAK;GACvB,OAAO;IACL,SAAS,KAAK;IACd,QAAQ,SAAS;IACjB,UAAU,SAAS;IACnB,WAAW,UAAU;GACvB;EACF,CAAC;EACD,kBAAkB,OAAO,iBAAiB,KAAK,QAAS,IAAwB,GAAG;EACnF,mBAAmB,OAAO,kBAAkB,KAAK,QAAS,IAAwB,GAAG;CACvF;AACF;AAEA,SAAS,iBAAiB,MAA4B,KAAgC;CACpF,OAAO,KAAK,KAAK,QAAQ;EACvB,MAAM,aAAa,IAAI,WAAW,KAAK;EACvC,IAAI,CAAC,YACH,MAAM,IAAI,gBACR,wBAAwB,IAAI,MAAM,+EACpC;EAEF,OAAO;GACL,SAAS,KAAK,UAAU,CAAC,IAAI,cAAc,UAAU,CAAC;GACtD,QAAQ,OAAO,IAAI,IAAI;GACvB;GACA;EACF;CACF,CAAC;AACH;AAEA,SAAS,yBAAyB,MAAkC,KAAmB;CACrF,MAAM,4BAAY,IAAI,IAA+B;CACrD,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,QAAQ,UAAU,IAAI,IAAI,OAAO,KAAK,CAAC;EAC7C,MAAM,KAAK,GAAG;EACd,UAAU,IAAI,IAAI,SAAS,KAAK;CAClC;CACA,KAAK,MAAM,CAAC,SAAS,UAAU,WAC7B,IAAI,MAAM,SAAS,GAAG,cAAc,OAAO,SAAS,GAAG;AAE3D;AAEA,SAAS,YAAY,KAAmB,MAAkC;CACxE,MAAM,IAAI,IAAI,UAAU;CACxB,IAAI,MAAM,KAAA,GAAW,OAAO,KAAA;CAC5B,IAAI,CAAC,OAAO,SAAS,CAAC,GACpB,MAAM,IAAI,gBACR,mDAAmD,KAAK,gBAAgB,IAAI,QAAQ,UAAU,IAAI,IAAI,MAAM,GAC9G;CAEF,OAAO;AACT"}
+{"version":3,"file":"paired-arms-6XItKzd1.js","names":[],"sources":["../src/paired-arms.ts"],"sourcesContent":["/**\n * Matched-pair arm comparison — \"did the treatment arm beat the baseline arm\n * on the SAME work items?\"\n *\n * An arm A/B over run records is only trustworthy when it is PAIRED: the same\n * task/scenario/seed evaluated under both arms, compared item-by-item, so\n * inter-item difficulty variance cancels instead of masquerading as an arm\n * effect. This module owns the two error-prone steps every consumer otherwise\n * hand-rolls:\n *\n *   1. Pairing — matching rows across arms by `pairKey` (and by `repKey`\n *      within multi-rep items), with leftovers REPORTED rather than silently\n *      dropped (a silently unbalanced pairing biases every paired statistic\n *      downstream). Pairing never keys on outcome content: matching reps by\n *      their outcomes deflates discordant-pair counts and makes McNemar\n *      anti-conservative, so reps pair only by (`pairKey`, `repKey`) identity.\n *   2. Composition — feeding the matched pairs to the correct paired\n *      estimators that already live in `statistics`: `mcnemar` +\n *      `pairedRiskDifference` for pass/fail, `pairedBootstrap` +\n *      `wilcoxonSignedRank` for continuous metrics. No statistic is\n *      re-implemented here.\n *\n * The row shape is deliberately structural — callers project a `RunRecord`\n * (or any record) into `{ pairKey, arm, pass?, metrics? }`. Arm names are\n * caller-supplied parameters; the module ships no domain literal.\n */\n\nimport { ValidationError } from './errors'\nimport type { RunRecord } from './run-record'\nimport type { McNemarResult, PairedBootstrapOptions, PairedBootstrapResult } from './statistics'\nimport {\n  mcnemar,\n  pairedBootstrap,\n  pairedRiskDifference,\n  type RiskDifferenceResult,\n  wilcoxonSignedRank,\n} from './statistics'\n\n/** One arm observation of one work item. Structural on purpose: callers\n *  project their own record type (e.g. a `RunRecord`) into this shape. */\nexport interface PairedArmRow {\n  /** Matching key — rows sharing a `pairKey` across both arms form pairs\n   *  (typically the task/scenario/seed identity). */\n  pairKey: string\n  /** Rep identity within a `pairKey` (e.g. a seed or rep number). Required on\n   *  every row of a `pairKey` that has more than one rep in either arm; reps\n   *  then pair only on exact (`pairKey`, `repKey`) match, never on outcome\n   *  content. Optional when each arm has at most one rep of the item. */\n  repKey?: string\n  /** Arm label this row was produced under. */\n  arm: string\n  /** Binary outcome; omit when the comparison has no pass/fail notion. */\n  pass?: boolean\n  /** Named numeric measurements (score, cost, latency, …). */\n  metrics?: Record<string, number>\n}\n\nexport interface PairArmsOptions {\n  /** Arm treated as the control side of every pair. */\n  baselineArm: string\n  /** Arm treated as the treatment side of every pair. */\n  treatmentArm: string\n}\n\n/** One matched (baseline, treatment) observation of the same work item. */\nexport interface MatchedPair {\n  pairKey: string\n  /** 0-based position of this pair within its `pairKey`, ordered by sorted\n   *  `repKey` (always 0 for a single-rep item). The rep identity itself is on\n   *  the rows (`baseline.repKey` / `treatment.repKey`). */\n  repIndex: number\n  baseline: PairedArmRow\n  treatment: PairedArmRow\n}\n\nexport interface PairArmsResult {\n  /** Matched pairs, ordered by (`pairKey`, `repIndex`). */\n  pairs: MatchedPair[]\n  /** Baseline rows left without a treatment counterpart — reported, never\n   *  silently dropped. */\n  unpairedBaseline: PairedArmRow[]\n  /** Treatment rows left without a baseline counterpart. */\n  unpairedTreatment: PairedArmRow[]\n}\n\n/**\n * Match rows across two arms into (baseline, treatment) pairs by `pairKey`.\n *\n * A `pairKey` with at most one row per arm pairs directly, no `repKey`\n * needed. A `pairKey` with multiple reps in either arm requires `repKey` on\n * every one of its rows, and reps pair only on exact (`pairKey`, `repKey`)\n * match — pairing is keyed purely on row identity, never on outcome content\n * (outcome-keyed matching deflates discordant counts and biases McNemar), and\n * is therefore independent of input order. Reps whose `repKey` has no\n * counterpart in the other arm, and items present in only one arm, land in\n * the unpaired lists — reported, never truncated.\n *\n * Fail-loud: throws when either named arm has zero rows (an unknown arm\n * name would otherwise read as \"everything unpaired\"), when the two arm\n * names are equal, when a multi-rep `pairKey` has a row without `repKey`, or\n * when a (`pairKey`, arm) group repeats a `repKey` (the match would be\n * ambiguous).\n */\nexport function pairArms(rows: readonly PairedArmRow[], opts: PairArmsOptions): PairArmsResult {\n  const { baselineArm, treatmentArm } = opts\n  if (baselineArm === treatmentArm) {\n    throw new ValidationError(\n      `pairArms: baselineArm and treatmentArm are both '${baselineArm}' — an arm cannot be compared to itself`,\n    )\n  }\n\n  // arm → pairKey → rows\n  const byArm = new Map<string, Map<string, PairedArmRow[]>>()\n  const armsSeen = new Set<string>()\n  for (const row of rows) {\n    armsSeen.add(row.arm)\n    if (row.arm !== baselineArm && row.arm !== treatmentArm) continue\n    const byKey = byArm.get(row.arm) ?? new Map<string, PairedArmRow[]>()\n    const group = byKey.get(row.pairKey) ?? []\n    group.push(row)\n    byKey.set(row.pairKey, group)\n    byArm.set(row.arm, byKey)\n  }\n\n  for (const arm of [baselineArm, treatmentArm]) {\n    if (!byArm.has(arm)) {\n      const seen = [...armsSeen].sort().join(', ') || '<none>'\n      throw new ValidationError(`pairArms: no rows for arm '${arm}' (arms present: ${seen})`)\n    }\n  }\n\n  const baselineByKey = byArm.get(baselineArm)!\n  const treatmentByKey = byArm.get(treatmentArm)!\n\n  const allKeys = [...new Set([...baselineByKey.keys(), ...treatmentByKey.keys()])].sort()\n  const pairs: MatchedPair[] = []\n  const unpairedBaseline: PairedArmRow[] = []\n  const unpairedTreatment: PairedArmRow[] = []\n  for (const pairKey of allKeys) {\n    const b = baselineByKey.get(pairKey) ?? []\n    const t = treatmentByKey.get(pairKey) ?? []\n\n    if (b.length <= 1 && t.length <= 1) {\n      if (b.length === 1 && t.length === 1) {\n        const baseline = b[0]!\n        const treatment = t[0]!\n        if (baseline.repKey !== undefined || treatment.repKey !== undefined) {\n          if (\n            baseline.repKey === undefined ||\n            treatment.repKey === undefined ||\n            baseline.repKey !== treatment.repKey\n          ) {\n            unpairedBaseline.push(baseline)\n            unpairedTreatment.push(treatment)\n            continue\n          }\n        }\n        pairs.push({ pairKey, repIndex: 0, baseline, treatment })\n      } else {\n        unpairedBaseline.push(...b)\n        unpairedTreatment.push(...t)\n      }\n      continue\n    }\n\n    const bByRep = indexByRepKey(b, pairKey, baselineArm)\n    const tByRep = indexByRepKey(t, pairKey, treatmentArm)\n    const repKeys = [...new Set([...bByRep.keys(), ...tByRep.keys()])].sort()\n    let repIndex = 0\n    for (const repKey of repKeys) {\n      const baseline = bByRep.get(repKey)\n      const treatment = tByRep.get(repKey)\n      if (baseline !== undefined && treatment !== undefined) {\n        pairs.push({ pairKey, repIndex: repIndex++, baseline, treatment })\n      } else if (baseline !== undefined) {\n        unpairedBaseline.push(baseline)\n      } else if (treatment !== undefined) {\n        unpairedTreatment.push(treatment)\n      }\n    }\n  }\n\n  return { pairs, unpairedBaseline, unpairedTreatment }\n}\n\n/** Index a multi-rep (pairKey, arm) group by `repKey`, enforcing that every\n *  row carries one and that no repKey repeats within the group. */\nfunction indexByRepKey(\n  group: readonly PairedArmRow[],\n  pairKey: string,\n  arm: string,\n): Map<string, PairedArmRow> {\n  const byRep = new Map<string, PairedArmRow>()\n  for (const row of group) {\n    if (row.repKey === undefined) {\n      throw new ValidationError(\n        `pairArms: pairKey '${pairKey}' has multiple reps in an arm, but a row in arm '${arm}' ` +\n          `is missing repKey — multi-rep items require an explicit repKey on every row so reps ` +\n          `pair by identity (pairing reps by outcome or by index would bias the paired statistics)`,\n      )\n    }\n    if (byRep.has(row.repKey)) {\n      throw new ValidationError(\n        `pairArms: duplicate repKey '${row.repKey}' for pairKey '${pairKey}' in arm '${arm}' — ` +\n          `(pairKey, repKey) must uniquely identify a rep within an arm`,\n      )\n    }\n    byRep.set(row.repKey, row)\n  }\n  return byRep\n}\n\n/** Paired pass/fail comparison over the pairs where BOTH sides carry `pass`. */\nexport interface PairedCorrectness {\n  /** Discordant pairs where the treatment passed and the baseline failed. */\n  b10: number\n  /** Discordant pairs where the baseline passed and the treatment failed. */\n  b01: number\n  /** Exact McNemar significance over the paired outcomes (`b === b10`, `c === b01`). */\n  mcnemar: McNemarResult\n  /** Paired effect size: p(treatment) − p(baseline) with a paired-variance CI. */\n  riskDifference: RiskDifferenceResult\n}\n\n/** Paired delta summary for one named metric (delta = treatment − baseline). */\nexport interface PairedMetricDelta {\n  name: string\n  /** Pairs where BOTH sides carry a finite value for this metric. */\n  n: number\n  /** Pairs where at least one side does not carry the metric. */\n  nMissing: number\n  /** Median paired delta, or null when `n === 0`. */\n  medianDelta: number | null\n  /** Mean paired delta, or null when `n === 0`. */\n  meanDelta: number | null\n  /** Bootstrap CI on the paired delta (`pairedBootstrap`); null when\n   *  `n === 0` — a zero-width [0, 0] interval on no data would read as a\n   *  measured tight null. */\n  bootstrapCi: PairedBootstrapResult | null\n  /** Wilcoxon signed-rank test on the paired deltas; null when `n === 0`. */\n  wilcoxon: { w: number; p: number } | null\n}\n\nexport interface ComparePairedArmsOptions extends PairArmsOptions {\n  /** Metrics to compare. Default: every metric name observed on any matched\n   *  pair, sorted. A name that appears on no pair is still reported (with\n   *  `n = 0`) so a misspelled metric is visible instead of vanishing. */\n  metricNames?: string[]\n  /** Passed through to `pairedBootstrap` — set `seed` for reproducible CIs. */\n  bootstrap?: PairedBootstrapOptions\n}\n\nexport interface PairedArmsComparison {\n  nPairs: number\n  nUnpairedBaseline: number\n  nUnpairedTreatment: number\n  /** null when no matched pair carries `pass` on both sides — a pass/fail\n   *  verdict over rows that never measured pass/fail would be fabricated. */\n  correctness: PairedCorrectness | null\n  metricDeltas: PairedMetricDelta[]\n}\n\n/**\n * Full matched-pair arm comparison: pair via {@link pairArms}, then compose\n * the paired estimators from `statistics` over the matched pairs.\n *\n * Correctness uses only the pairs where both sides carry `pass` (`mcnemar.n`\n * is that subset's size); each metric uses only the pairs where both sides\n * carry a finite value for it, with the remainder counted in `nMissing`.\n * Deltas are treatment − baseline throughout.\n *\n * Fail-loud: inherits {@link pairArms}'s unknown-arm throw, and throws on a\n * non-finite metric value — silently treating corrupt telemetry as \"metric\n * absent\" would misreport it as missing coverage.\n */\nexport function comparePairedArms(\n  rows: readonly PairedArmRow[],\n  opts: ComparePairedArmsOptions,\n): PairedArmsComparison {\n  const { pairs, unpairedBaseline, unpairedTreatment } = pairArms(rows, opts)\n\n  let correctness: PairedCorrectness | null = null\n  const baselinePass: number[] = []\n  const treatmentPass: number[] = []\n  for (const pair of pairs) {\n    if (pair.baseline.pass === undefined || pair.treatment.pass === undefined) continue\n    baselinePass.push(pair.baseline.pass ? 1 : 0)\n    treatmentPass.push(pair.treatment.pass ? 1 : 0)\n  }\n  if (baselinePass.length > 0) {\n    const mc = mcnemar(baselinePass, treatmentPass)\n    correctness = {\n      b10: mc.b,\n      b01: mc.c,\n      mcnemar: mc,\n      riskDifference: pairedRiskDifference(baselinePass, treatmentPass),\n    }\n  }\n\n  const metricNames =\n    opts.metricNames ??\n    [\n      ...new Set(\n        pairs.flatMap((p) => [\n          ...Object.keys(p.baseline.metrics ?? {}),\n          ...Object.keys(p.treatment.metrics ?? {}),\n        ]),\n      ),\n    ].sort()\n\n  const metricDeltas: PairedMetricDelta[] = metricNames.map((name) => {\n    const before: number[] = []\n    const after: number[] = []\n    let nMissing = 0\n    for (const pair of pairs) {\n      const b = metricValue(pair.baseline, name)\n      const t = metricValue(pair.treatment, name)\n      if (b === undefined || t === undefined) {\n        nMissing++\n        continue\n      }\n      before.push(b)\n      after.push(t)\n    }\n    const bootstrapCi = before.length === 0 ? null : pairedBootstrap(before, after, opts.bootstrap)\n    return {\n      name,\n      n: before.length,\n      nMissing,\n      medianDelta: bootstrapCi?.median ?? null,\n      meanDelta: bootstrapCi?.mean ?? null,\n      bootstrapCi,\n      wilcoxon: before.length === 0 ? null : wilcoxonSignedRank(before, after),\n    }\n  })\n\n  return {\n    nPairs: pairs.length,\n    nUnpairedBaseline: unpairedBaseline.length,\n    nUnpairedTreatment: unpairedTreatment.length,\n    correctness,\n    metricDeltas,\n  }\n}\n\nexport interface MatchedRunRecordPair {\n  pairKey: string\n  repKey: string\n  baseline: RunRecord\n  treatment: RunRecord\n}\n\nexport interface PairRunRecordsResult {\n  pairs: MatchedRunRecordPair[]\n  unpairedBaseline: RunRecord[]\n  unpairedTreatment: RunRecord[]\n}\n\ninterface RunRecordArmRow extends PairedArmRow {\n  run: RunRecord\n  repKey: string\n}\n\n/**\n * Pair two RunRecord arms by the identity of the evaluated work:\n * `(experimentId, scenarioId, seed)`.\n *\n * Falling back to array order, candidate id, or experiment id can compare\n * different tasks and fabricate lift. Duplicate identities throw.\n */\nexport function pairRunRecords(\n  baselineRuns: readonly RunRecord[],\n  treatmentRuns: readonly RunRecord[],\n): PairRunRecordsResult {\n  const baselineRows = runRecordArmRows(baselineRuns, 'baseline')\n  const treatmentRows = runRecordArmRows(treatmentRuns, 'treatment')\n  validateRunRecordArmRows(baselineRows, 'baseline')\n  validateRunRecordArmRows(treatmentRows, 'treatment')\n  if (baselineRows.length === 0 || treatmentRows.length === 0) {\n    return {\n      pairs: [],\n      unpairedBaseline: baselineRows.map((row) => row.run),\n      unpairedTreatment: treatmentRows.map((row) => row.run),\n    }\n  }\n\n  const result = pairArms([...baselineRows, ...treatmentRows], {\n    baselineArm: 'baseline',\n    treatmentArm: 'treatment',\n  })\n  return {\n    pairs: result.pairs.map((pair) => {\n      const baseline = pair.baseline as RunRecordArmRow\n      const treatment = pair.treatment as RunRecordArmRow\n      return {\n        pairKey: pair.pairKey,\n        repKey: baseline.repKey,\n        baseline: baseline.run,\n        treatment: treatment.run,\n      }\n    }),\n    unpairedBaseline: result.unpairedBaseline.map((row) => (row as RunRecordArmRow).run),\n    unpairedTreatment: result.unpairedTreatment.map((row) => (row as RunRecordArmRow).run),\n  }\n}\n\nfunction runRecordArmRows(runs: readonly RunRecord[], arm: string): RunRecordArmRow[] {\n  return runs.map((run) => {\n    const scenarioId = run.scenarioId.trim()\n    if (!scenarioId) {\n      throw new ValidationError(\n        `pairRunRecords: run '${run.runId}' is missing scenarioId; paired comparisons require explicit scenario identity`,\n      )\n    }\n    return {\n      pairKey: JSON.stringify([run.experimentId, scenarioId]),\n      repKey: String(run.seed),\n      arm,\n      run,\n    }\n  })\n}\n\nfunction validateRunRecordArmRows(rows: readonly RunRecordArmRow[], arm: string): void {\n  const byPairKey = new Map<string, RunRecordArmRow[]>()\n  for (const row of rows) {\n    const group = byPairKey.get(row.pairKey) ?? []\n    group.push(row)\n    byPairKey.set(row.pairKey, group)\n  }\n  for (const [pairKey, group] of byPairKey) {\n    if (group.length > 1) indexByRepKey(group, pairKey, arm)\n  }\n}\n\nfunction metricValue(row: PairedArmRow, name: string): number | undefined {\n  const v = row.metrics?.[name]\n  if (v === undefined) return undefined\n  if (!Number.isFinite(v)) {\n    throw new ValidationError(\n      `comparePairedArms: non-finite value for metric '${name}' on pairKey '${row.pairKey}' (arm '${row.arm}'): ${v}`,\n    )\n  }\n  return v\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuGA,SAAgB,SAAS,MAA+B,MAAuC;CAC7F,MAAM,EAAE,aAAa,iBAAiB;CACtC,IAAI,gBAAgB,cAClB,MAAM,IAAI,gBACR,oDAAoD,YAAY,wCAClE;CAIF,MAAM,wBAAQ,IAAI,IAAyC;CAC3D,MAAM,2BAAW,IAAI,IAAY;CACjC,KAAK,MAAM,OAAO,MAAM;EACtB,SAAS,IAAI,IAAI,GAAG;EACpB,IAAI,IAAI,QAAQ,eAAe,IAAI,QAAQ,cAAc;EACzD,MAAM,QAAQ,MAAM,IAAI,IAAI,GAAG,qBAAK,IAAI,IAA4B;EACpE,MAAM,QAAQ,MAAM,IAAI,IAAI,OAAO,KAAK,CAAC;EACzC,MAAM,KAAK,GAAG;EACd,MAAM,IAAI,IAAI,SAAS,KAAK;EAC5B,MAAM,IAAI,IAAI,KAAK,KAAK;CAC1B;CAEA,KAAK,MAAM,OAAO,CAAC,aAAa,YAAY,GAC1C,IAAI,CAAC,MAAM,IAAI,GAAG,GAEhB,MAAM,IAAI,gBAAgB,8BAA8B,IAAI,mBAD/C,CAAC,GAAG,QAAQ,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,IAAI,KAAK,SACoC,EAAE;CAI1F,MAAM,gBAAgB,MAAM,IAAI,WAAW;CAC3C,MAAM,iBAAiB,MAAM,IAAI,YAAY;CAE7C,MAAM,UAAU,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,cAAc,KAAK,GAAG,GAAG,eAAe,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK;CACvF,MAAM,QAAuB,CAAC;CAC9B,MAAM,mBAAmC,CAAC;CAC1C,MAAM,oBAAoC,CAAC;CAC3C,KAAK,MAAM,WAAW,SAAS;EAC7B,MAAM,IAAI,cAAc,IAAI,OAAO,KAAK,CAAC;EACzC,MAAM,IAAI,eAAe,IAAI,OAAO,KAAK,CAAC;EAE1C,IAAI,EAAE,UAAU,KAAK,EAAE,UAAU,GAAG;GAClC,IAAI,EAAE,WAAW,KAAK,EAAE,WAAW,GAAG;IACpC,MAAM,WAAW,EAAE;IACnB,MAAM,YAAY,EAAE;IACpB,IAAI,SAAS,WAAW,KAAA,KAAa,UAAU,WAAW,KAAA,GAEtD;SAAA,SAAS,WAAW,KAAA,KACpB,UAAU,WAAW,KAAA,KACrB,SAAS,WAAW,UAAU,QAC9B;MACA,iBAAiB,KAAK,QAAQ;MAC9B,kBAAkB,KAAK,SAAS;MAChC;KACF;;IAEF,MAAM,KAAK;KAAE;KAAS,UAAU;KAAG;KAAU;IAAU,CAAC;GAC1D,OAAO;IACL,iBAAiB,KAAK,GAAG,CAAC;IAC1B,kBAAkB,KAAK,GAAG,CAAC;GAC7B;GACA;EACF;EAEA,MAAM,SAAS,cAAc,GAAG,SAAS,WAAW;EACpD,MAAM,SAAS,cAAc,GAAG,SAAS,YAAY;EACrD,MAAM,UAAU,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,OAAO,KAAK,GAAG,GAAG,OAAO,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK;EACxE,IAAI,WAAW;EACf,KAAK,MAAM,UAAU,SAAS;GAC5B,MAAM,WAAW,OAAO,IAAI,MAAM;GAClC,MAAM,YAAY,OAAO,IAAI,MAAM;GACnC,IAAI,aAAa,KAAA,KAAa,cAAc,KAAA,GAC1C,MAAM,KAAK;IAAE;IAAS,UAAU;IAAY;IAAU;GAAU,CAAC;QAC5D,IAAI,aAAa,KAAA,GACtB,iBAAiB,KAAK,QAAQ;QACzB,IAAI,cAAc,KAAA,GACvB,kBAAkB,KAAK,SAAS;EAEpC;CACF;CAEA,OAAO;EAAE;EAAO;EAAkB;CAAkB;AACtD;;;AAIA,SAAS,cACP,OACA,SACA,KAC2B;CAC3B,MAAM,wBAAQ,IAAI,IAA0B;CAC5C,KAAK,MAAM,OAAO,OAAO;EACvB,IAAI,IAAI,WAAW,KAAA,GACjB,MAAM,IAAI,gBACR,sBAAsB,QAAQ,mDAAmD,IAAI,8KAGvF;EAEF,IAAI,MAAM,IAAI,IAAI,MAAM,GACtB,MAAM,IAAI,gBACR,+BAA+B,IAAI,OAAO,iBAAiB,QAAQ,YAAY,IAAI,iEAErF;EAEF,MAAM,IAAI,IAAI,QAAQ,GAAG;CAC3B;CACA,OAAO;AACT;;;;;;;;;;;;;;AAiEA,SAAgB,kBACd,MACA,MACsB;CACtB,MAAM,EAAE,OAAO,kBAAkB,sBAAsB,SAAS,MAAM,IAAI;CAE1E,IAAI,cAAwC;CAC5C,MAAM,eAAyB,CAAC;CAChC,MAAM,gBAA0B,CAAC;CACjC,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,KAAK,SAAS,SAAS,KAAA,KAAa,KAAK,UAAU,SAAS,KAAA,GAAW;EAC3E,aAAa,KAAK,KAAK,SAAS,OAAO,IAAI,CAAC;EAC5C,cAAc,KAAK,KAAK,UAAU,OAAO,IAAI,CAAC;CAChD;CACA,IAAI,aAAa,SAAS,GAAG;EAC3B,MAAM,KAAK,QAAQ,cAAc,aAAa;EAC9C,cAAc;GACZ,KAAK,GAAG;GACR,KAAK,GAAG;GACR,SAAS;GACT,gBAAgB,qBAAqB,cAAc,aAAa;EAClE;CACF;CAaA,MAAM,gBAVJ,KAAK,eACL,CACE,GAAG,IAAI,IACL,MAAM,SAAS,MAAM,CACnB,GAAG,OAAO,KAAK,EAAE,SAAS,WAAW,CAAC,CAAC,GACvC,GAAG,OAAO,KAAK,EAAE,UAAU,WAAW,CAAC,CAAC,CAC1C,CAAC,CACH,CACF,CAAC,CAAC,KAAK,EAAA,CAE6C,KAAK,SAAS;EAClE,MAAM,SAAmB,CAAC;EAC1B,MAAM,QAAkB,CAAC;EACzB,IAAI,WAAW;EACf,KAAK,MAAM,QAAQ,OAAO;GACxB,MAAM,IAAI,YAAY,KAAK,UAAU,IAAI;GACzC,MAAM,IAAI,YAAY,KAAK,WAAW,IAAI;GAC1C,IAAI,MAAM,KAAA,KAAa,MAAM,KAAA,GAAW;IACtC;IACA;GACF;GACA,OAAO,KAAK,CAAC;GACb,MAAM,KAAK,CAAC;EACd;EACA,MAAM,cAAc,OAAO,WAAW,IAAI,OAAO,gBAAgB,QAAQ,OAAO,KAAK,SAAS;EAC9F,OAAO;GACL;GACA,GAAG,OAAO;GACV;GACA,aAAa,aAAa,UAAU;GACpC,WAAW,aAAa,QAAQ;GAChC;GACA,UAAU,OAAO,WAAW,IAAI,OAAO,mBAAmB,QAAQ,KAAK;EACzE;CACF,CAAC;CAED,OAAO;EACL,QAAQ,MAAM;EACd,mBAAmB,iBAAiB;EACpC,oBAAoB,kBAAkB;EACtC;EACA;CACF;AACF;;;;;;;;AA2BA,SAAgB,eACd,cACA,eACsB;CACtB,MAAM,eAAe,iBAAiB,cAAc,UAAU;CAC9D,MAAM,gBAAgB,iBAAiB,eAAe,WAAW;CACjE,yBAAyB,cAAc,UAAU;CACjD,yBAAyB,eAAe,WAAW;CACnD,IAAI,aAAa,WAAW,KAAK,cAAc,WAAW,GACxD,OAAO;EACL,OAAO,CAAC;EACR,kBAAkB,aAAa,KAAK,QAAQ,IAAI,GAAG;EACnD,mBAAmB,cAAc,KAAK,QAAQ,IAAI,GAAG;CACvD;CAGF,MAAM,SAAS,SAAS,CAAC,GAAG,cAAc,GAAG,aAAa,GAAG;EAC3D,aAAa;EACb,cAAc;CAChB,CAAC;CACD,OAAO;EACL,OAAO,OAAO,MAAM,KAAK,SAAS;GAChC,MAAM,WAAW,KAAK;GACtB,MAAM,YAAY,KAAK;GACvB,OAAO;IACL,SAAS,KAAK;IACd,QAAQ,SAAS;IACjB,UAAU,SAAS;IACnB,WAAW,UAAU;GACvB;EACF,CAAC;EACD,kBAAkB,OAAO,iBAAiB,KAAK,QAAS,IAAwB,GAAG;EACnF,mBAAmB,OAAO,kBAAkB,KAAK,QAAS,IAAwB,GAAG;CACvF;AACF;AAEA,SAAS,iBAAiB,MAA4B,KAAgC;CACpF,OAAO,KAAK,KAAK,QAAQ;EACvB,MAAM,aAAa,IAAI,WAAW,KAAK;EACvC,IAAI,CAAC,YACH,MAAM,IAAI,gBACR,wBAAwB,IAAI,MAAM,+EACpC;EAEF,OAAO;GACL,SAAS,KAAK,UAAU,CAAC,IAAI,cAAc,UAAU,CAAC;GACtD,QAAQ,OAAO,IAAI,IAAI;GACvB;GACA;EACF;CACF,CAAC;AACH;AAEA,SAAS,yBAAyB,MAAkC,KAAmB;CACrF,MAAM,4BAAY,IAAI,IAA+B;CACrD,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,QAAQ,UAAU,IAAI,IAAI,OAAO,KAAK,CAAC;EAC7C,MAAM,KAAK,GAAG;EACd,UAAU,IAAI,IAAI,SAAS,KAAK;CAClC;CACA,KAAK,MAAM,CAAC,SAAS,UAAU,WAC7B,IAAI,MAAM,SAAS,GAAG,cAAc,OAAO,SAAS,GAAG;AAE3D;AAEA,SAAS,YAAY,KAAmB,MAAkC;CACxE,MAAM,IAAI,IAAI,UAAU;CACxB,IAAI,MAAM,KAAA,GAAW,OAAO,KAAA;CAC5B,IAAI,CAAC,OAAO,SAAS,CAAC,GACpB,MAAM,IAAI,gBACR,mDAAmD,KAAK,gBAAgB,IAAI,QAAQ,UAAU,IAAI,IAAI,MAAM,GAC9G;CAEF,OAAO;AACT"}

package/dist/pipelines/index.js

@@ -1,9 +1,9 @@
 import { t as executionTrackByLane } from "../execution-tracks-CpgFPpS5.js";
-import { E as pearsonR, u as interRaterReliability } from "../statistics-CnnxdpOg.js";
+import { E as pearsonR, u as interRaterReliability } from "../statistics-DWM_AyLe.js";
 import { s as isToolSpan } from "../schema-CRhEY1SO.js";
 import { t as buildTrajectory } from "../trajectory-D_7rLrvE.js";
 import { i as hasCapturedToolArgs, l as toolSpans, n as argHash, o as llmSpans, s as runFailureClass, t as aggregateLlm } from "../query-Di7eEQ79.js";
-import { a as DEFAULT_RULES, i as computeToolUseMetrics, o as classifyFailure, t as compareToBaseline } from "../baseline-HsBvw_dk.js";
+import { a as DEFAULT_RULES, i as computeToolUseMetrics, o as classifyFailure, t as compareToBaseline } from "../baseline-DcX5hQDv.js";
 //#region src/pipelines/budget-breach.ts
 async function budgetBreachView(store, options = {}) {
 	const runs = await store.listRuns({