pipeline/src/cd_eval/structured_extractor.py

"""Evidence-grounded structured extraction from the interim IMF corpus.

This is a transparent deterministic baseline, not an LLM-generated gold set.
Every observation, recommendation, and relationship includes page evidence and
an explicit extraction method/confidence. Contextual links are labeled as such
and are never represented as explicit causal claims.
"""

from __future__ import annotations

import argparse
import concurrent.futures
import datetime as dt
import hashlib
import json
import os
import re
import shutil
import sys
import tempfile
from collections import Counter, defaultdict
from pathlib import Path
from typing import Any, Iterable, Sequence

import jsonschema

SCHEMA_VERSION = "1.1.0"
EXTRACTOR_VERSION = "1.1.0"
EXTRACTION_METHOD = "deterministic_evidence_baseline"

HEADING_RE = re.compile(r"^\s*#{1,6}\s+(.+?)\s*$")
TABLE_SEPARATOR_RE = re.compile(r"^\s*\|?\s*:?-{3,}")
DOI_RE = re.compile(r"\b10\.5089/[A-Za-z0-9._;()/:-]+", re.I)
ISBN_RE = re.compile(r"\b(?:97[89][ -]?)?(?:\d[ -]?){9}[\dXx]\b")
PARAGRAPH_NUMBER_RE = re.compile(r"(?:¶|paragraph\s+)?(\d{1,3})", re.I)

RECOMMENDATION_HEADING_RE = re.compile(
    r"\b(recommendations?|recommended actions?|recomendaciones|recomenda(?:ç|c)[õo]es|recommandations?|"
    r"рекомендации|рекомендация)\b|التوصيات|توصيات",
    re.I,
)
OBSERVATION_HEADING_RE = re.compile(
    r"\b(executive summary|key findings?|main findings?|assessment|diagnostic|observations?|"
    r"resumen ejecutivo|principales hallazgos|constatations|sum[aá]rio executivo|"
    r"резюме|основные выводы|результаты)\b|ملخص|النتائج|الملاحظات",
    re.I,
)
EXCLUDED_SECTION_RE = re.compile(
    r"\b(contents|table of contents|glossary|preface|appendi(?:x|ces)|annex|bibliography|references)\b",
    re.I,
)
STRONG_EXPLICIT_RECOMMENDATION_RE = re.compile(
    r"\b(the mission (?:recommends?|recommended)|is recommended|are recommended|"
    r"we recommend|recommendation is to|it is recommended|se recomienda|recomenda-se|"
    r"il est recommand[ée]|рекомендуется)\b|توصي البعثة",
    re.I,
)
RECOMMENDATION_TABLE_HEADER_RE = re.compile(
    r"^(?:(?:main|key|priority)\s+)?recommendations?$|"
    r"^(?:short|medium|long)[- ]term projections?$|^recommended actions?$",
    re.I,
)

RECOMMENDATION_MODAL_RE = re.compile(
    r"\b(should|must|needs? to|is recommended|are recommended|the mission recommends?|"
    r"recommended that|recommendation is to|priority is to|deber[ií]a|debe(?:n)?|se recomienda|"
    r"devrait|doit|il est recommand[ée]|deveria|deve(?:m)?|recomenda-se|"
    r"следует|необходимо|долж(?:ен|на|ны)|рекомендуется)\b|ينبغي|يجب|يوصى",
    re.I,
)
IMPERATIVE_RE = re.compile(
    r"^(strengthen|establish|develop|adopt|implement|improve|ensure|create|prepare|finalize|"
    r"introduce|increase|reduce|review|revise|update|set up|initiate|start|continue|conduct|"
    r"align|clarify|define|enhance|formalize|operationalize|provide|require|maintain|"
    r"fortalecer|establecer|desarrollar|implementar|mejorar|garantizar|adoptar|"
    r"renforcer|[ée]tablir|am[ée]liorer|mettre en œuvre|adopter|"
    r"refor[çc]ar|estabelecer|desenvolver|implementar|melhorar|adotar|"
    r"укрепить|создать|разработать|внедрить|улучшить|обеспечить|принять)\b|"
    r"^(?:تعزيز|إنشاء|تطوير|تنفيذ|تحسين|ضمان|اعتماد)",
    re.I,
)
OBSERVATION_SIGNAL_RE = re.compile(
    r"\b(found|finds|finding|remains?|lacks?|weak(?:ness|nesses)?|limited|insufficient|"
    r"constraint|gap|shortcoming|challenge|risk|vulnerab|deficien|not yet|does not|do not|"
    r"has not|have not|however|progress|improved|effective|ineffective|fragmented|outdated|"
    r"ausencia|débil|limitad|insuficient|deficien|desaf[ií]o|riesgo|"
    r"faible|limit[ée]|insuffisant|lacune|risque|"
    r"fraco|limitado|insuficiente|defici[êe]ncia|desafio|risco|"
    r"недостат|слаб|огранич|риск|проблем|отсутств)\w*\b|"
    r"ضعف|يفتقر|محدود|تحديات|مخاطر|عدم",
    re.I,
)

MONTHS = {
    "january": 1,
    "february": 2,
    "march": 3,
    "april": 4,
    "may": 5,
    "june": 6,
    "july": 7,
    "august": 8,
    "september": 9,
    "october": 10,
    "november": 11,
    "december": 12,
}
MONTH_PATTERN = "|".join(MONTHS)
SAME_MONTH_RANGE_RE = re.compile(
    rf"\b(?P<month>{MONTH_PATTERN})\s+(?P<start>\d{{1,2}})\s*[–—-]\s*"
    rf"(?P<end>\d{{1,2}}),?\s+(?P<year>20\d{{2}})\b",
    re.I,
)
CROSS_MONTH_RANGE_RE = re.compile(
    rf"\b(?P<month1>{MONTH_PATTERN})\s+(?P<start>\d{{1,2}})\s*[–—-]\s*"
    rf"(?P<month2>{MONTH_PATTERN})\s+(?P<end>\d{{1,2}}),?\s+(?P<year>20\d{{2}})\b",
    re.I,
)
SINGLE_DATE_RE = re.compile(
    rf"\b(?P<month>{MONTH_PATTERN})\s+(?P<day>\d{{1,2}}),?\s+(?P<year>20\d{{2}})\b",
    re.I,
)

_GLOBAL_FIGURES_CACHE: dict[str, dict[str, list[dict[str, Any]]]] = {}

STOPWORDS = {
    "the", "a", "an", "and", "or", "of", "to", "in", "for", "on", "with", "by",
    "that", "this", "these", "those", "is", "are", "be", "should", "must", "it", "its",
    "as", "from", "at", "has", "have", "will", "would", "could", "their", "which", "into",
    "imf", "mission", "recommend", "recommended", "recommendation",
}


def utc_now() -> str:
    return dt.datetime.now(dt.timezone.utc).replace(microsecond=0).isoformat()


def load_json(path: Path) -> Any:
    return json.loads(path.read_text(encoding="utf-8"))


def load_jsonl(path: Path) -> list[dict[str, Any]]:
    if not path.exists():
        return []
    with path.open(encoding="utf-8") as source:
        return [json.loads(line) for line in source if line.strip()]


def write_json(path: Path, value: Any) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    temporary = path.with_suffix(path.suffix + ".tmp")
    temporary.write_text(
        json.dumps(value, ensure_ascii=False, indent=2, sort_keys=False) + "\n",
        encoding="utf-8",
    )
    os.replace(temporary, path)


def write_jsonl(path: Path, rows: Iterable[dict[str, Any]]) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    temporary = path.with_suffix(path.suffix + ".tmp")
    with temporary.open("w", encoding="utf-8") as output:
        for row in rows:
            output.write(json.dumps(row, ensure_ascii=False, sort_keys=True) + "\n")
    os.replace(temporary, path)


def clean_markdown(value: str) -> str:
    value = re.sub(r"!\[[^\]]*\]\([^)]+\)", "", value)
    value = re.sub(r"\[([^\]]+)\]\([^)]+\)", r"\1", value)
    value = re.sub(r"<br\s*/?>", " ", value, flags=re.I)
    value = re.sub(r"</?[^>]+>", " ", value)
    value = re.sub(r"^\s*#{1,6}\s*", "", value)
    value = value.replace("**", "").replace("__", "").replace("`", "")
    value = value.replace("_", " ")
    value = re.sub(r"^\s*[-*•]\s+", "", value)
    value = re.sub(r"\s+", " ", value)
    return value.strip(" |\t\r\n")


def normalize_evidence(value: str) -> str:
    value = value.replace("_", " ")
    return re.sub(r"[^\w]+", " ", value.lower(), flags=re.UNICODE).strip()


def evidence_id(report_id: str, page: int, quote: str) -> str:
    digest = hashlib.sha256(
        f"{report_id}|{page}|{normalize_evidence(quote)}".encode("utf-8")
    ).hexdigest()[:16]
    return f"ev-{digest}"


def evidence(report_id: str, page: int, quote: str, source: str = "document") -> dict[str, Any]:
    return {
        "evidence_id": evidence_id(report_id, page, quote),
        "page": page,
        "quote": quote,
        "source": source,
    }


def token_set(value: str) -> set[str]:
    return {
        token
        for token in re.findall(r"\b[^\W\d_]\w{2,}\b", value.lower(), flags=re.UNICODE)
        if token not in STOPWORDS
    }


def similarity(left: str, right: str) -> float:
    left_tokens, right_tokens = token_set(left), token_set(right)
    if not left_tokens or not right_tokens:
        return 0.0
    return len(left_tokens & right_tokens) / len(left_tokens | right_tokens)


def split_sentences(paragraph: str) -> list[str]:
    paragraph = re.sub(r"\s*\n\s*", " ", paragraph).strip()
    if not paragraph:
        return []
    pieces = re.split(r"(?<=[.!?؟])\s+(?=\S)", paragraph)
    return [piece.strip() for piece in pieces if piece.strip()]


def parse_table_cells(line: str) -> list[str]:
    line = line.strip().strip("|")
    return [clean_markdown(cell) for cell in re.split(r"(?<!\\)\|", line)]


def table_blocks(markdown: str) -> list[list[str]]:
    blocks: list[list[str]] = []
    current: list[str] = []
    for line in markdown.splitlines():
        if line.strip().startswith("|") and line.count("|") >= 2:
            current.append(line)
        else:
            if len(current) >= 2:
                blocks.append(current)
            current = []
    if len(current) >= 2:
        blocks.append(current)
    return blocks


def column_index(headers: list[str], aliases: tuple[str, ...]) -> int | None:
    for index, header in enumerate(headers):
        lowered = header.lower()
        if any(alias in lowered for alias in aliases):
            return index
    return None


def fingerprint(value: str) -> str:
    return " ".join(sorted(token_set(value)))


def add_unique(
    collection: list[dict[str, Any]],
    item: dict[str, Any],
    *,
    similarity_threshold: float = 0.88,
) -> int:
    item_fp = fingerprint(item["verbatim"])
    for index, existing in enumerate(collection):
        existing_fp = existing.get("_fingerprint", "")
        if item_fp == existing_fp or similarity(item["verbatim"], existing["verbatim"]) >= similarity_threshold:
            known = {entry["evidence_id"] for entry in existing["evidence"]}
            existing["evidence"].extend(
                entry for entry in item["evidence"] if entry["evidence_id"] not in known
            )
            if item["confidence"] > existing["confidence"]:
                for key in (
                    "text", "verbatim", "actor", "priority", "timeframe",
                    "section", "extraction_method", "confidence", "_page", "_paragraph_key",
                ):
                    if key in item:
                        existing[key] = item[key]
            return index
    item["_fingerprint"] = item_fp
    collection.append(item)
    return len(collection) - 1


def parse_priority(value: str) -> str | None:
    match = re.search(r"\b(high|medium|low|critical|alta|media|baja|haute|moyenne|faible)\b", value, re.I)
    return match.group(1).lower() if match else None


def parse_timeframe(value: str) -> str | None:
    match = re.search(
        r"\b(near[- ]term|short[- ]term|medium[- ]term|long[- ]term|immediate|"
        r"NT|ST|MT|LT|\d+\s*(?:months?|years?))\b",
        value,
        re.I,
    )
    return match.group(1) if match else None


def recommendation_item(
    report_id: str,
    page: int,
    value: str,
    *,
    section: str,
    method: str,
    confidence: float,
    actor: str | None = None,
    priority: str | None = None,
    timeframe: str | None = None,
    paragraph_key: str | None = None,
) -> dict[str, Any]:
    quote = clean_markdown(value)
    paragraph_match = PARAGRAPH_NUMBER_RE.search(quote)
    return {
        "text": quote,
        "verbatim": quote,
        "actor": actor,
        "priority": priority,
        "timeframe": timeframe,
        "report_paragraph": int(paragraph_match.group(1)) if paragraph_match and "¶" in quote else None,
        "section": section or None,
        "evidence": [evidence(report_id, page, quote)],
        "confidence": confidence,
        "extraction_method": method,
        "review_status": "unreviewed",
        "_page": page,
        "_paragraph_key": paragraph_key,
    }


def observation_item(
    report_id: str,
    page: int,
    value: str,
    *,
    section: str,
    method: str,
    confidence: float,
    paragraph_key: str | None = None,
) -> dict[str, Any]:
    quote = clean_markdown(value)
    return {
        "text": quote,
        "verbatim": quote,
        "topic": section or None,
        "severity": None,
        "evidence": [evidence(report_id, page, quote)],
        "confidence": confidence,
        "extraction_method": method,
        "review_status": "unreviewed",
        "_page": page,
        "_paragraph_key": paragraph_key,
    }


def extract_tables(
    report_id: str,
    pages: list[dict[str, Any]],
    recommendations: list[dict[str, Any]],
    observations: list[dict[str, Any]],
) -> list[tuple[int, int, dict[str, Any]]]:
    explicit_links: list[tuple[int, int, dict[str, Any]]] = []
    for page in pages:
        page_number = page["page"]
        for table_number, lines in enumerate(table_blocks(page["markdown"]), start=1):
            cells = [parse_table_cells(line) for line in lines]
            if len(cells) < 2:
                continue
            headers = cells[0]
            row_start = 2 if len(cells) > 1 and TABLE_SEPARATOR_RE.match(lines[1]) else 1
            rec_col = column_index(
                headers,
                (
                    "recommend", "recommended action", "action required", "proposed action",
                    "recomenda", "рекоменд", "توص",
                ),
            )
            rtl_recommendation_table = False
            if (
                rec_col is None
                and len(headers) == 3
                and re.search(r"[\u0600-\u06ff\ufb50-\ufdff\ufe70-\ufeff]", " ".join(headers))
                and page_number <= 10
            ):
                # Arabic PDF extraction may emit presentation-form glyphs in
                # table headers. IMF action-plan tables consistently place the
                # recommendation/action in the center column after extraction.
                rec_col = 1
                rtl_recommendation_table = True
            if rec_col is None:
                continue
            obs_col = column_index(
                headers,
                (
                    "observation", "finding", "issue", "challenge", "weakness", "rationale",
                    "constat", "вывод", "проблем", "ملاحظ", "نتائج", "قضايا",
                ),
            )
            actor_col = column_index(
                headers, ("responsible", "authority", "institution", "agency", "actor")
            )
            if rtl_recommendation_table:
                actor_col = None
            priority_col = column_index(headers, ("priority", "prioridad", "priorité"))
            time_col = column_index(
                headers, ("timeframe", "timing", "timeline", "deadline", "term")
            )
            section = f"recommendation table {table_number}"
            for row_number, row in enumerate(cells[row_start:], start=1):
                if rec_col >= len(row):
                    continue
                rec_text = row[rec_col]
                if (
                    len(rec_text) < 15
                    or len(token_set(rec_text)) < 3
                    or RECOMMENDATION_HEADING_RE.fullmatch(rec_text)
                    or RECOMMENDATION_TABLE_HEADER_RE.fullmatch(rec_text)
                ):
                    continue
                actor = row[actor_col] if actor_col is not None and actor_col < len(row) else None
                priority_raw = row[priority_col] if priority_col is not None and priority_col < len(row) else ""
                time_raw = row[time_col] if time_col is not None and time_col < len(row) else ""
                paragraph_key = f"p{page_number}-table{table_number}-row{row_number}"
                rec_index = add_unique(
                    recommendations,
                    recommendation_item(
                        report_id,
                        page_number,
                        rec_text,
                        section=section,
                        method="recommendation_table",
                        confidence=0.98,
                        actor=actor or None,
                        priority=parse_priority(priority_raw),
                        timeframe=time_raw or parse_timeframe(rec_text),
                        paragraph_key=paragraph_key,
                    ),
                )
                if obs_col is not None and obs_col < len(row) and len(row[obs_col]) >= 15:
                    obs_text = row[obs_col]
                    obs_index = add_unique(
                        observations,
                        observation_item(
                            report_id,
                            page_number,
                            obs_text,
                            section=section,
                            method="observation_recommendation_table",
                            confidence=0.98,
                            paragraph_key=paragraph_key,
                        ),
                    )
                    explicit_links.append(
                        (
                            obs_index,
                            rec_index,
                            {
                                "relation": "addresses",
                                "link_basis": "explicit_table_row",
                                "confidence": 1.0,
                                "evidence": [evidence(report_id, page_number, " | ".join(row))],
                                "review_status": "unreviewed",
                            },
                        )
                    )
    return explicit_links


def paragraphs_with_sections(pages: list[dict[str, Any]]) -> list[dict[str, Any]]:
    output: list[dict[str, Any]] = []
    section = ""
    for page in pages:
        markdown = page["markdown"]
        table_line_numbers = {
            index
            for index, line in enumerate(markdown.splitlines())
            if line.strip().startswith("|") and line.count("|") >= 2
        }
        current: list[str] = []
        paragraph_counter = 0

        def flush() -> None:
            nonlocal paragraph_counter
            value = "\n".join(current).strip()
            current.clear()
            cleaned = clean_markdown(value)
            if cleaned:
                paragraph_counter += 1
                output.append(
                    {
                        "page": page["page"],
                        "section": section,
                        "raw": value,
                        "text": cleaned,
                        "key": f"p{page['page']}-para{paragraph_counter}",
                    }
                )

        for line_number, line in enumerate(markdown.splitlines()):
            heading_match = HEADING_RE.match(line)
            if heading_match:
                flush()
                section = clean_markdown(heading_match.group(1))
                continue
            if line_number in table_line_numbers:
                flush()
                continue
            if not line.strip():
                flush()
            else:
                current.append(line)
        flush()
    return output


def valid_candidate(value: str) -> bool:
    if not 25 <= len(value) <= 900:
        return False
    if value.count("_") > 5 or re.search(r"_{5,}|\.{5,}", value):
        return False
    if re.fullmatch(r"[\W\d_]+", value):
        return False
    if re.search(r"IMF (?:Technical Assistance|Country) Report\s*\|?\s*\d+", value, re.I):
        return False
    return True


def extract_body_candidates(
    report_id: str,
    pages: list[dict[str, Any]],
    recommendations: list[dict[str, Any]],
    observations: list[dict[str, Any]],
) -> None:
    for paragraph in paragraphs_with_sections(pages):
        section = paragraph["section"]
        if EXCLUDED_SECTION_RE.search(section):
            continue
        recommendation_section = bool(RECOMMENDATION_HEADING_RE.search(section))
        observation_section = bool(OBSERVATION_HEADING_RE.search(section))
        raw_starts_bullet = bool(re.match(r"\s*(?:[-*•]|\d+[.)])\s+", paragraph["raw"]))
        for sentence in split_sentences(paragraph["text"]):
            sentence = clean_markdown(sentence)
            if not valid_candidate(sentence):
                continue
            has_modal = bool(RECOMMENDATION_MODAL_RE.search(sentence))
            imperative = bool(IMPERATIVE_RE.search(sentence))
            if has_modal or (recommendation_section and (imperative or raw_starts_bullet)):
                confidence = 0.88 if recommendation_section else 0.72
                method = (
                    "recommendation_section_sentence"
                    if recommendation_section
                    else "explicit_recommendation_modal"
                )
                add_unique(
                    recommendations,
                    recommendation_item(
                        report_id,
                        paragraph["page"],
                        sentence,
                        section=section,
                        method=method,
                        confidence=confidence,
                        priority=parse_priority(sentence),
                        timeframe=parse_timeframe(sentence),
                        paragraph_key=paragraph["key"],
                    ),
                )
                continue
            has_signal = bool(OBSERVATION_SIGNAL_RE.search(sentence))
            numbered_finding = bool(re.match(r"^\d+\.\s+", paragraph["text"]))
            if has_signal and (observation_section or numbered_finding or len(sentence) >= 50):
                confidence = 0.82 if observation_section else 0.62
                add_unique(
                    observations,
                    observation_item(
                        report_id,
                        paragraph["page"],
                        sentence,
                        section=section,
                        method=(
                            "finding_section_sentence"
                            if observation_section
                            else "diagnostic_signal_sentence"
                        ),
                        confidence=confidence,
                        paragraph_key=paragraph["key"],
                    ),
                )


def parse_iso_date(year: int, month: int, day: int) -> str | None:
    try:
        return dt.date(year, month, day).isoformat()
    except ValueError:
        return None


def extract_date_mentions(
    report_id: str,
    title: str,
    pages: list[dict[str, Any]],
    publication_date: str | None,
    source_page_url: str | None,
) -> list[dict[str, Any]]:
    dates: list[dict[str, Any]] = []
    url_date = None
    if source_page_url:
        match = re.search(r"/issues/(\d{4})/(\d{2})/(\d{2})/", source_page_url, re.I)
        if match:
            url_date = parse_iso_date(int(match.group(1)), int(match.group(2)), int(match.group(3)))
    primary_publication_date = url_date or (publication_date[:10] if publication_date else None)
    if primary_publication_date:
        dates.append(
            {
                "type": "publication",
                "start": primary_publication_date,
                "end": primary_publication_date,
                "precision": "day",
                "verbatim": primary_publication_date,
                "evidence": [
                    {"source": "imf_publication_url" if url_date else "imf_index_metadata"}
                ],
                "confidence": 1.0 if not url_date else 0.98,
            }
        )
    if publication_date and publication_date[:10] != primary_publication_date:
        dates.append(
            {
                "type": "imf_index_date",
                "start": publication_date[:10],
                "end": publication_date[:10],
                "precision": "day",
                "verbatim": publication_date,
                "evidence": [{"source": "imf_index_metadata"}],
                "confidence": 1.0,
            }
        )
    search_sources = [(0, title)] + [
        (page["page"], page["text"]) for page in pages[:6]
    ]
    seen: set[tuple[str, str | None, str | None]] = set()
    for page_number, text in search_sources:
        for pattern, cross_month in ((CROSS_MONTH_RANGE_RE, True), (SAME_MONTH_RANGE_RE, False)):
            for match in pattern.finditer(text):
                year = int(match.group("year"))
                if cross_month:
                    start_month = MONTHS[match.group("month1").lower()]
                    end_month = MONTHS[match.group("month2").lower()]
                else:
                    start_month = end_month = MONTHS[match.group("month").lower()]
                start = parse_iso_date(year, start_month, int(match.group("start")))
                end = parse_iso_date(year, end_month, int(match.group("end")))
                key = ("mission_or_report_range", start, end)
                if start and end and key not in seen:
                    seen.add(key)
                    quote = match.group(0)
                    dates.append(
                        {
                            "type": "mission_or_report_range",
                            "start": start,
                            "end": end,
                            "precision": "day",
                            "verbatim": quote,
                            "evidence": (
                                [evidence(report_id, page_number, quote)]
                                if page_number
                                else [{"source": "title", "quote": quote}]
                            ),
                            "confidence": 0.8,
                        }
                    )
        for match in SINGLE_DATE_RE.finditer(text):
            year = int(match.group("year"))
            value = parse_iso_date(year, MONTHS[match.group("month").lower()], int(match.group("day")))
            key = ("date_mention", value, value)
            if value and key not in seen:
                seen.add(key)
                quote = match.group(0)
                dates.append(
                    {
                        "type": "date_mention",
                        "start": value,
                        "end": value,
                        "precision": "day",
                        "verbatim": quote,
                        "evidence": (
                            [evidence(report_id, page_number, quote)]
                            if page_number
                            else [{"source": "title", "quote": quote}]
                        ),
                        "confidence": 0.65,
                    }
                )
    return dates


def prepared_by_statement(pages: list[dict[str, Any]]) -> tuple[int, str, str] | None:
    stop_re = re.compile(
        r"^(?:authoring\s+)?departments?\b|^approved\b|^authorized\b|"
        r"^international monetary fund\b|^the mission\b|^prepared for\b",
        re.I,
    )
    department_phrase_re = re.compile(
        r"\b(fiscal affairs|monetary and capital markets|statistics|legal|"
        r"institute for capacity development|finance|research|department)\b",
        re.I,
    )
    for page in pages[:8]:
        lines = [line.strip(" \t:;") for line in page["text"].splitlines()]
        for index, line in enumerate(lines):
            match = re.search(r"\bPrepared\s+by\b\s*[:\-]?\s*(.*)$", line, re.I)
            if not match:
                continue
            collected = [match.group(1).strip()] if match.group(1).strip() else []
            stop_line = ""
            for candidate in lines[index + 1 : index + 10]:
                if not candidate:
                    continue
                if stop_re.search(candidate):
                    stop_line = candidate
                    break
                if re.fullmatch(r"(?:[A-Z][A-Z .&/-]+|\d{4})", candidate) and collected:
                    break
                collected.append(candidate)
                if len(collected) >= 4:
                    break
            if stop_line.lower() == "department" and len(collected) > 1:
                if department_phrase_re.search(collected[-1]):
                    collected.pop()
            statement = re.sub(
                r"\s+", " ", " ".join(collected).replace("_", " ")
            ).strip(" .,;")
            if 2 <= len(statement) <= 300:
                raw_quote = "Prepared By\n" + "\n".join(collected)
                return page["page"], statement, raw_quote
    return None


def split_prepared_by_names(statement: str) -> list[str]:
    normalized = re.sub(r"\s+(?:and|&)\s+", ",", statement, flags=re.I)
    parts = [part.strip(" .;,") for part in re.split(r"[,;]", normalized)]
    plausible = []
    for part in parts:
        words = re.findall(r"[^\W\d_]+", part, flags=re.UNICODE)
        if 2 <= len(words) <= 10 and not re.search(r"\b(department|division|team|staff)\b", part, re.I):
            plausible.append(part)
    return plausible or [statement]


def extract_authors(
    report_id: str, report: dict[str, Any], document: dict[str, Any], pages: list[dict[str, Any]]
) -> list[dict[str, Any]]:
    authors: list[dict[str, Any]] = []
    indexed = report.get("author_indexed")
    if indexed:
        authors.append(
            {
                "name": indexed,
                "role": "indexed_institutional_author",
                "evidence": [{"source": "imf_index_metadata"}],
                "confidence": 1.0,
            }
        )
    pdf_author = (document.get("pdf_metadata") or {}).get("author")
    if pdf_author and pdf_author.lower() not in {str(indexed).lower(), "imf"}:
        authors.append(
            {
                "name": pdf_author,
                "role": "pdf_metadata_author",
                "evidence": [{"source": "pdf_metadata"}],
                "confidence": 0.9,
            }
        )
    prepared = prepared_by_statement(pages)
    if prepared:
        page_number, statement, quote = prepared
        for name in split_prepared_by_names(statement):
            if all(name.lower() != author["name"].lower() for author in authors):
                authors.append(
                    {
                        "name": name,
                        "role": "prepared_by",
                        "verbatim_statement": statement,
                        "evidence": [evidence(report_id, page_number, quote)],
                        "confidence": 0.82,
                    }
                )
    return authors


def extract_authoring_departments(report_id: str, pages: list[dict[str, Any]]) -> list[dict[str, Any]]:
    departments: list[dict[str, Any]] = []
    known_re = re.compile(
        r"\b(Fiscal Affairs Department|Monetary and Capital Markets Department|"
        r"Statistics Department|Legal Department|Institute for Capacity Development|"
        r"Research Department|Finance Department)\b",
        re.I,
    )
    for page in pages[:8]:
        for match in known_re.finditer(re.sub(r"\s+", " ", page["text"])):
            name = match.group(1)
            if all(name.lower() != item["name"].lower() for item in departments):
                departments.append(
                    {
                        "name": name,
                        "evidence": [evidence(report_id, page["page"], match.group(0))],
                        "confidence": 0.9,
                    }
                )
    return departments


def extract_identifiers(pages: list[dict[str, Any]], report: dict[str, Any]) -> dict[str, Any]:
    text = "\n".join(page["text"] for page in pages[:8])
    dois = sorted({match.group(0).rstrip(".,;)") for match in DOI_RE.finditer(text)})
    isbns = []
    for match in ISBN_RE.finditer(text):
        compact = re.sub(r"[ -]", "", match.group(0)).upper()
        if len(compact) in {10, 13} and compact not in isbns:
            isbns.append(compact)
    return {
        "series": report.get("series", []),
        "series_volume_no": report.get("series_volume_no"),
        "doi": dois,
        "isbn": isbns,
        "subjects": report.get("subjects", []),
        "topics": report.get("topics", []),
        "keywords": report.get("keywords", []),
        "description_indexed": report.get("description"),
    }


def classify_recommendation(item: dict[str, Any]) -> None:
    """Assign an explicitness taxonomy without discarding recall-oriented candidates."""
    text = item["verbatim"]
    method = item["extraction_method"]
    strong_explicit = bool(STRONG_EXPLICIT_RECOMMENDATION_RE.search(text))
    direct_action = bool(IMPERATIVE_RE.search(text) or RECOMMENDATION_MODAL_RE.search(text))

    if method == "recommendation_table":
        recommendation_type = "explicit_table"
        explicitness = "explicit"
        tier = "high"
        conservative = True
        confidence = 0.98
    elif strong_explicit:
        recommendation_type = "explicit_attributed_statement"
        explicitness = "explicit"
        tier = "high"
        conservative = True
        confidence = 0.92
    elif method == "recommendation_section_sentence" and direct_action:
        recommendation_type = "direct_action_in_recommendation_section"
        explicitness = "direct_normative"
        tier = "high"
        conservative = True
        confidence = 0.85
    elif method == "explicit_recommendation_modal":
        recommendation_type = "normative_modal_candidate"
        explicitness = "implicit_candidate"
        tier = "medium"
        conservative = False
        confidence = 0.65
    else:
        recommendation_type = "recommendation_section_context_candidate"
        explicitness = "context_candidate"
        tier = "low"
        conservative = False
        confidence = 0.35

    item["recommendation_type"] = recommendation_type
    item["explicitness"] = explicitness
    item["confidence_tier"] = tier
    item["in_conservative_set"] = conservative
    item["confidence"] = confidence


def finalize_entities(
    report_id: str,
    recommendations: list[dict[str, Any]],
    observations: list[dict[str, Any]],
    explicit_links: list[tuple[int, int, dict[str, Any]]],
) -> tuple[list[dict[str, Any]], list[dict[str, Any]], list[dict[str, Any]]]:
    for item in recommendations:
        classify_recommendation(item)
    recommendations.sort(key=lambda item: (item["_page"], -item["confidence"], item["verbatim"]))
    observations.sort(key=lambda item: (item["_page"], -item["confidence"], item["verbatim"]))
    # Sorting invalidates explicit temporary indices, so remap by paragraph keys and
    # text. Explicit rows have unique shared paragraph keys.
    rec_by_paragraph = defaultdict(list)
    obs_by_paragraph = defaultdict(list)
    for index, item in enumerate(recommendations):
        rec_by_paragraph[item.get("_paragraph_key")].append(index)
    for index, item in enumerate(observations):
        obs_by_paragraph[item.get("_paragraph_key")].append(index)

    for index, item in enumerate(recommendations, start=1):
        item["recommendation_id"] = f"{report_id}-rec-{index:04d}"
    for index, item in enumerate(observations, start=1):
        item["observation_id"] = f"{report_id}-obs-{index:04d}"

    links: list[dict[str, Any]] = []
    seen_pairs: set[tuple[str, str, str]] = set()

    # Reconstruct explicit table-row links from matching paragraph keys.
    explicit_keys = {
        recommendations[rec_index].get("_paragraph_key")
        for _, rec_index, _ in explicit_links
        if 0 <= rec_index < len(recommendations)
    }
    # The list above may refer to pre-sort indices; all explicit rows are also
    # identifiable by the table-row paragraph key.
    explicit_keys.update(
        item.get("_paragraph_key")
        for item in recommendations
        if item.get("extraction_method") == "recommendation_table"
    )
    for key in explicit_keys:
        if not key:
            continue
        for obs_index in obs_by_paragraph.get(key, []):
            for rec_index in rec_by_paragraph.get(key, []):
                obs = observations[obs_index]
                rec = recommendations[rec_index]
                pair = (obs["observation_id"], rec["recommendation_id"], "explicit_table_row")
                if pair in seen_pairs:
                    continue
                seen_pairs.add(pair)
                links.append(
                    {
                        "observation_id": obs["observation_id"],
                        "recommendation_id": rec["recommendation_id"],
                        "relation": "addresses",
                        "link_basis": "explicit_table_row",
                        "evidence": rec["evidence"],
                        "confidence": 1.0,
                        "review_status": "unreviewed",
                        "recommendation_type": rec["recommendation_type"],
                        "recommendation_confidence_tier": rec["confidence_tier"],
                        "conservative_recommendation": rec["in_conservative_set"],
                    }
                )

    # Same-paragraph modal links are strong contextual evidence, but not declared causal.
    for rec in recommendations:
        key = rec.get("_paragraph_key")
        if not key:
            continue
        for obs_index in obs_by_paragraph.get(key, []):
            obs = observations[obs_index]
            pair = (obs["observation_id"], rec["recommendation_id"], "same_paragraph")
            if any(existing[:2] == pair[:2] for existing in seen_pairs):
                continue
            seen_pairs.add(pair)
            links.append(
                {
                    "observation_id": obs["observation_id"],
                    "recommendation_id": rec["recommendation_id"],
                    "relation": "responds_to_context",
                    "link_basis": "same_paragraph",
                    "evidence": rec["evidence"],
                    "confidence": 0.78,
                    "review_status": "unreviewed",
                    "recommendation_type": rec["recommendation_type"],
                    "recommendation_confidence_tier": rec["confidence_tier"],
                    "conservative_recommendation": rec["in_conservative_set"],
                }
            )

    # If no stronger link exists, retain one clearly labeled lexical/contextual link.
    linked_recommendations = {link["recommendation_id"] for link in links}
    for rec in recommendations:
        if rec["recommendation_id"] in linked_recommendations:
            continue
        candidates = []
        for obs in observations:
            if abs(obs["_page"] - rec["_page"]) > 1:
                continue
            score = similarity(obs["verbatim"], rec["verbatim"])
            if score >= 0.12:
                candidates.append((score, obs))
        if candidates:
            score, obs = max(candidates, key=lambda item: item[0])
            links.append(
                {
                    "observation_id": obs["observation_id"],
                    "recommendation_id": rec["recommendation_id"],
                    "relation": "contextually_associated_with",
                    "link_basis": "same_or_adjacent_page_lexical_similarity",
                    "evidence": rec["evidence"],
                    "confidence": round(min(0.65, 0.4 + score), 3),
                    "review_status": "unreviewed",
                    "recommendation_type": rec["recommendation_type"],
                    "recommendation_confidence_tier": rec["confidence_tier"],
                    "conservative_recommendation": rec["in_conservative_set"],
                }
            )

    for collection in (recommendations, observations):
        for item in collection:
            for key in list(item):
                if key.startswith("_"):
                    del item[key]
    links.sort(key=lambda item: (item["recommendation_id"], item["observation_id"]))
    return recommendations, observations, links


def report_schema() -> dict[str, Any]:
    evidence_schema = {
        "type": "object",
        "properties": {
            "evidence_id": {"type": "string"},
            "page": {"type": "integer", "minimum": 1},
            "quote": {"type": "string"},
            "source": {"type": "string"},
        },
        "required": ["source"],
        "additionalProperties": True,
    }
    return {
        "$schema": "https://json-schema.org/draft/2020-12/schema",
        "$id": "https://cd-eval.local/schemas/report.schema.json",
        "title": "IMF Technical Assistance Structured Report",
        "type": "object",
        "required": [
            "schema_version", "report_id", "source_sha256", "title", "authors",
            "countries", "dates", "metadata", "observations", "recommendations",
            "observation_recommendation_links", "figures", "extraction",
        ],
        "properties": {
            "schema_version": {"const": SCHEMA_VERSION},
            "report_id": {"type": "string"},
            "source_sha256": {"type": "string", "pattern": "^[0-9a-f]{64}$"},
            "title": {"type": "string"},
            "language": {"type": "array", "items": {"type": "string"}},
            "authors": {"type": "array", "items": {"type": "object"}},
            "countries": {"type": "array", "items": {"type": "object"}},
            "dates": {"type": "array", "items": {"type": "object"}},
            "metadata": {"type": "object"},
            "observations": {
                "type": "array",
                "items": {
                    "type": "object",
                    "required": ["observation_id", "text", "verbatim", "evidence", "confidence"],
                    "properties": {"evidence": {"type": "array", "items": evidence_schema}},
                    "additionalProperties": True,
                },
            },
            "recommendations": {
                "type": "array",
                "items": {
                    "type": "object",
                    "required": [
                        "recommendation_id", "text", "verbatim", "evidence", "confidence",
                        "recommendation_type", "explicitness", "confidence_tier",
                        "in_conservative_set",
                    ],
                    "properties": {"evidence": {"type": "array", "items": evidence_schema}},
                    "additionalProperties": True,
                },
            },
            "observation_recommendation_links": {
                "type": "array",
                "items": {
                    "type": "object",
                    "required": [
                        "observation_id", "recommendation_id", "relation", "link_basis",
                        "evidence", "confidence",
                    ],
                    "additionalProperties": True,
                },
            },
            "figures": {"type": "array", "items": {"type": "object"}},
            "extraction": {"type": "object"},
        },
        "additionalProperties": False,
    }


def validate_evidence_grounding(record: dict[str, Any], pages: list[dict[str, Any]]) -> list[str]:
    errors: list[str] = []
    page_text = {page["page"]: normalize_evidence(page["text"] + " " + page["markdown"]) for page in pages}
    page_count = len(pages)
    for kind in ("observations", "recommendations"):
        for item in record[kind]:
            if not item.get("evidence"):
                errors.append(f"{kind}:{item.get(kind[:-1] + '_id')}: missing evidence")
            for entry in item.get("evidence", []):
                page = entry.get("page")
                quote = normalize_evidence(entry.get("quote", ""))
                if not isinstance(page, int) or not 1 <= page <= page_count:
                    errors.append(f"{kind}: invalid page {page}")
                elif quote and quote not in page_text.get(page, ""):
                    # Layout table extraction can normalize hyphenation or collapse
                    # spaces between adjacent PDF text spans differently.
                    compact_quote = re.sub(r"\s+", "", quote)
                    compact_page = re.sub(r"\s+", "", page_text.get(page, ""))
                    quote_tokens = token_set(quote)
                    page_tokens = token_set(page_text.get(page, ""))
                    token_coverage = len(quote_tokens & page_tokens) / max(1, len(quote_tokens))
                    if compact_quote not in compact_page and token_coverage < 0.9:
                        errors.append(f"{kind}: evidence not grounded on page {page}: {entry.get('quote','')[:80]}")
    observation_ids = {item["observation_id"] for item in record["observations"]}
    recommendation_ids = {item["recommendation_id"] for item in record["recommendations"]}
    for link in record["observation_recommendation_links"]:
        if link["observation_id"] not in observation_ids:
            errors.append(f"link unknown observation {link['observation_id']}")
        if link["recommendation_id"] not in recommendation_ids:
            errors.append(f"link unknown recommendation {link['recommendation_id']}")
    return errors


def global_figures_for_report(interim_dir: Path, report_id: str) -> list[dict[str, Any]]:
    cache_key = interim_dir.resolve().as_posix()
    if cache_key not in _GLOBAL_FIGURES_CACHE:
        grouped: dict[str, list[dict[str, Any]]] = defaultdict(list)
        for figure in load_jsonl(interim_dir / "figures.jsonl"):
            grouped[figure["report_id"]].append(figure)
        _GLOBAL_FIGURES_CACHE[cache_key] = dict(grouped)
    return _GLOBAL_FIGURES_CACHE[cache_key].get(report_id, [])


def extract_one(job: dict[str, Any]) -> dict[str, Any]:
    report = job["report"]
    source = job["manifest"]
    report_id = report["report_id"]
    interim_report_dir = Path(job["interim_dir"]) / report_id
    processed_dir = Path(job["processed_dir"])
    final_path = processed_dir / "reports" / f"{report_id}.json"
    refresh = bool(job.get("refresh"))

    if final_path.exists() and not refresh:
        existing = load_json(final_path)
        if (
            existing.get("source_sha256") == source["sha256"]
            and existing.get("extraction", {}).get("extractor_version") == EXTRACTOR_VERSION
        ):
            return {
                "report_id": report_id,
                "status": "existing",
                "record": existing,
                "validation_errors": [],
            }

    pages = load_jsonl(interim_report_dir / "pages.jsonl")
    document = load_json(interim_report_dir / "document.json")
    figures = global_figures_for_report(Path(job["interim_dir"]), report_id)
    if not pages:
        raise RuntimeError(f"missing interim pages for {report_id}")

    recommendations: list[dict[str, Any]] = []
    observations: list[dict[str, Any]] = []
    explicit_links = extract_tables(report_id, pages, recommendations, observations)
    extract_body_candidates(report_id, pages, recommendations, observations)
    # Keep a conservative upper bound and favor higher-confidence candidates.
    if len(recommendations) > 250:
        recommendations = sorted(recommendations, key=lambda item: -item["confidence"])[:250]
    if len(observations) > 300:
        observations = sorted(observations, key=lambda item: -item["confidence"])[:300]
    recommendations, observations, links = finalize_entities(
        report_id, recommendations, observations, explicit_links
    )

    countries = []
    formal = report.get("formal_countries", [])
    iso_codes = report.get("iso_codes", [])
    names = report.get("countries", []) or formal
    country_source = "imf_index_metadata"
    country_confidence = 1.0
    if not names and ":" in (report.get("title") or ""):
        names = [(report["title"].split(":", 1)[0]).strip()]
        country_source = "title_prefix_fallback"
        country_confidence = 0.9
    fallback_iso = {
        "armenia": "ARM",
        "republic of armenia": "ARM",
        "kosovo": "XKX",
        "republic of kosovo": "XKX",
        "democratic republic of the congo": "COD",
    }
    for index, name in enumerate(names):
        countries.append(
            {
                "name": name,
                "formal_name": formal[index] if index < len(formal) else None,
                "iso3": (
                    iso_codes[index]
                    if index < len(iso_codes)
                    else fallback_iso.get(name.lower())
                ),
                "evidence": [{"source": country_source}],
                "confidence": country_confidence,
            }
        )

    record = {
        "schema_version": SCHEMA_VERSION,
        "report_id": report_id,
        "source_sha256": source["sha256"],
        "title": report.get("title") or "",
        "language": report.get("language", []),
        "authors": extract_authors(report_id, report, document, pages),
        "countries": countries,
        "dates": extract_date_mentions(
            report_id,
            report.get("title") or "",
            pages,
            report.get("publication_date"),
            report.get("source_page_url"),
        ),
        "metadata": {
            **extract_identifiers(pages, report),
            "authoring_departments": extract_authoring_departments(report_id, pages),
            "source_page_url": report.get("source_page_url"),
            "source_pdf_url": source.get("source_pdf_url"),
            "page_count": document["page_count"],
            "interim_extraction_method": document["extraction_method"],
            "needs_ocr_pages": document.get("needs_ocr_pages", []),
        },
        "observations": observations,
        "recommendations": recommendations,
        "observation_recommendation_links": links,
        "figures": figures,
        "extraction": {
            "method": EXTRACTION_METHOD,
            "extractor_version": EXTRACTOR_VERSION,
            "generated_at": utc_now(),
            "review_status": "unreviewed",
            "evidence_requirement": "page-grounded verbatim source span",
            "limitations": [
                "Automated deterministic baseline; not a human-annotated gold record.",
                "The recommendations array is recall-oriented and includes classified candidates; use in_conservative_set=true for the higher-precision subset.",
                "Contextual links are proximity/lexical associations unless link_basis is explicit_table_row.",
                "Priority, timeframe, and actor are null when not explicit in a recommendation table or sentence.",
            ],
        },
    }
    schema_errors = [error.message for error in jsonschema.Draft202012Validator(report_schema()).iter_errors(record)]
    grounding_errors = validate_evidence_grounding(record, pages)
    errors = schema_errors + grounding_errors
    final_path.parent.mkdir(parents=True, exist_ok=True)
    write_json(final_path, record)
    return {
        "report_id": report_id,
        "status": "extracted",
        "record": record,
        "validation_errors": errors,
    }


def flatten_outputs(
    results: list[dict[str, Any]], processed_dir: Path, failures: list[dict[str, str]]
) -> dict[str, Any]:
    records = [result["record"] for result in results]
    records.sort(key=lambda record: record["report_id"])
    write_jsonl(processed_dir / "reports.jsonl", records)
    write_jsonl(
        processed_dir / "observations.jsonl",
        (
            {"report_id": record["report_id"], **item}
            for record in records
            for item in record["observations"]
        ),
    )
    write_jsonl(
        processed_dir / "recommendations.jsonl",
        (
            {"report_id": record["report_id"], **item}
            for record in records
            for item in record["recommendations"]
        ),
    )
    write_jsonl(
        processed_dir / "recommendations_conservative.jsonl",
        (
            {"report_id": record["report_id"], **item}
            for record in records
            for item in record["recommendations"]
            if item["in_conservative_set"]
        ),
    )
    write_jsonl(
        processed_dir / "observation_recommendation_links.jsonl",
        (
            {"report_id": record["report_id"], **item}
            for record in records
            for item in record["observation_recommendation_links"]
        ),
    )
    write_jsonl(
        processed_dir / "observation_recommendation_links_conservative.jsonl",
        (
            {"report_id": record["report_id"], **item}
            for record in records
            for item in record["observation_recommendation_links"]
            if item.get("conservative_recommendation")
        ),
    )
    write_jsonl(
        processed_dir / "figures.jsonl",
        (
            item for record in records for item in record["figures"]
        ),
    )
    validation_errors = [
        {"report_id": result["report_id"], "errors": result["validation_errors"]}
        for result in results
        if result["validation_errors"]
    ]
    recommendation_type_counts = Counter(
        item["recommendation_type"]
        for record in records
        for item in record["recommendations"]
    )
    recommendation_tier_counts = Counter(
        item["confidence_tier"]
        for record in records
        for item in record["recommendations"]
    )
    conservative_recommendation_count = sum(
        item["in_conservative_set"]
        for record in records
        for item in record["recommendations"]
    )
    conservative_link_count = sum(
        item.get("conservative_recommendation", False)
        for record in records
        for item in record["observation_recommendation_links"]
    )
    reports_with_no_conservative_recommendations = [
        record["report_id"]
        for record in records
        if not any(item["in_conservative_set"] for item in record["recommendations"])
    ]
    summary = {
        "updated_at": utc_now(),
        "schema_version": SCHEMA_VERSION,
        "extractor_version": EXTRACTOR_VERSION,
        "method": EXTRACTION_METHOD,
        "report_count": len(records),
        "observation_count": sum(len(record["observations"]) for record in records),
        "recommendation_count": sum(len(record["recommendations"]) for record in records),
        "recommendation_candidate_count": sum(
            len(record["recommendations"]) for record in records
        ),
        "conservative_recommendation_count": conservative_recommendation_count,
        "conservative_recommendation_report_count": (
            len(records) - len(reports_with_no_conservative_recommendations)
        ),
        "reports_with_no_conservative_recommendations": (
            reports_with_no_conservative_recommendations
        ),
        "recommendation_type_counts": dict(sorted(recommendation_type_counts.items())),
        "recommendation_confidence_tier_counts": dict(
            sorted(recommendation_tier_counts.items())
        ),
        "link_count": sum(len(record["observation_recommendation_links"]) for record in records),
        "conservative_link_count": conservative_link_count,
        "explicit_table_link_count": sum(
            link["link_basis"] == "explicit_table_row"
            for record in records
            for link in record["observation_recommendation_links"]
        ),
        "figure_count": sum(len(record["figures"]) for record in records),
        "reports_with_no_observations": [
            record["report_id"] for record in records if not record["observations"]
        ],
        "reports_with_no_recommendations": [
            record["report_id"] for record in records if not record["recommendations"]
        ],
        "validation_error_report_count": len(validation_errors),
        "validation_errors": validation_errors,
        "failed_count": len(failures),
        "failures": failures,
        "review_status": "unreviewed_automated_baseline",
        "recommendation_taxonomy": {
            "explicit_table": "A recommendation/action row in a report table explicitly designated for recommendations or an action plan.",
            "explicit_attributed_statement": "Text explicitly attributed as a recommendation (for example, 'the mission recommends' or 'is recommended').",
            "direct_action_in_recommendation_section": "An imperative or normative action inside a recommendation section.",
            "normative_modal_candidate": "A should/must/need-to statement outside a recommendation section; retained for recall but not in the conservative set.",
            "recommendation_section_context_candidate": "Context in a recommendation section without a direct action signal; low-confidence candidate.",
        },
    }
    write_json(processed_dir / "validation_summary.json", summary)
    write_json(
        processed_dir / "recommendation_taxonomy.json",
        {
            "schema_version": SCHEMA_VERSION,
            "definitions": summary["recommendation_taxonomy"],
            "type_counts": summary["recommendation_type_counts"],
            "confidence_tier_counts": summary[
                "recommendation_confidence_tier_counts"
            ],
            "candidate_count": summary["recommendation_candidate_count"],
            "conservative_count": summary["conservative_recommendation_count"],
        },
    )
    write_json(processed_dir / "schemas" / "report.schema.json", report_schema())
    return summary


def extract_corpus(
    inventory: list[dict[str, Any]],
    manifest: list[dict[str, Any]],
    *,
    interim_dir: Path,
    processed_dir: Path,
    workers: int,
    refresh: bool,
) -> dict[str, Any]:
    source_by_id = {row["report_id"]: row for row in manifest}
    jobs = [
        {
            "report": report,
            "manifest": source_by_id[report["report_id"]],
            "interim_dir": interim_dir.as_posix(),
            "processed_dir": processed_dir.as_posix(),
            "refresh": refresh,
        }
        for report in inventory
    ]
    results: list[dict[str, Any]] = []
    failures: list[dict[str, str]] = []
    with concurrent.futures.ProcessPoolExecutor(max_workers=max(1, workers)) as pool:
        futures = {pool.submit(extract_one, job): job for job in jobs}
        for future in concurrent.futures.as_completed(futures):
            job = futures[future]
            try:
                result = future.result()
                results.append(result)
                status = result["status"]
            except Exception as exc:
                failures.append(
                    {
                        "report_id": job["report"]["report_id"],
                        "error": f"{type(exc).__name__}: {exc}",
                    }
                )
                status = "failed"
            print(
                f"extract: {len(results) + len(failures)}/{len(jobs)} {status}: "
                f"{job['report']['report_id']}",
                file=sys.stderr,
            )
    summary = flatten_outputs(results, processed_dir, failures)
    return summary


def select_reports(
    inventory: list[dict[str, Any]], limit: int | None, ids: str | None
) -> list[dict[str, Any]]:
    if ids:
        selected_ids = {value.strip() for value in ids.split(",") if value.strip()}
        missing = selected_ids - {row["report_id"] for row in inventory}
        if missing:
            raise SystemExit(f"unknown report IDs: {sorted(missing)}")
        inventory = [row for row in inventory if row["report_id"] in selected_ids]
    if limit is not None:
        inventory = inventory[:limit]
    return inventory


def build_parser() -> argparse.ArgumentParser:
    parser = argparse.ArgumentParser(description="Extract structured IMF recommendations and observations")
    parser.add_argument("--raw-dir", type=Path, default=Path("data/raw"))
    parser.add_argument("--interim-dir", type=Path, default=Path("data/interim"))
    parser.add_argument("--processed-dir", type=Path, default=Path("data/processed"))
    parser.add_argument("--workers", type=int, default=min(8, os.cpu_count() or 1))
    parser.add_argument("--limit", type=int)
    parser.add_argument("--ids", help="comma-separated report IDs")
    parser.add_argument("--refresh", action="store_true")
    return parser


def main(argv: Sequence[str] | None = None) -> int:
    args = build_parser().parse_args(argv)
    inventory = load_jsonl(args.raw_dir / "manifests" / "inventory.jsonl")
    manifest = load_jsonl(args.raw_dir / "manifests" / "download_manifest.jsonl")
    if not inventory or not manifest:
        raise SystemExit("raw inventory/download manifest is missing")
    inventory = select_reports(inventory, args.limit, args.ids)
    missing_interim = [
        report["report_id"]
        for report in inventory
        if not (args.interim_dir / report["report_id"] / "document.json").exists()
    ]
    if missing_interim:
        raise SystemExit(
            f"interim conversion missing for {len(missing_interim)} report(s); "
            "run imf-process convert first"
        )
    summary = extract_corpus(
        inventory,
        manifest,
        interim_dir=args.interim_dir,
        processed_dir=args.processed_dir,
        workers=args.workers,
        refresh=args.refresh,
    )
    print(json.dumps(summary, ensure_ascii=False, indent=2), file=sys.stderr)
    return 1 if summary["failed_count"] else 0


if __name__ == "__main__":
    raise SystemExit(main())