# Recovers a "date" for docs.json entries that TIMELINE.md left undated, by
# reading each document's original text (in originals/, not committed — see
# .gitignore) and looking for its own closing signature date, e.g.
# "Brasília, 10 de abril de 2026." Reference script; not run automatically.
#
# Of the 615 originally undated documents, 366 got a recovered date this way
# (flagged `dateInferred: true` in docs.json); 249 had no recognizable
# signature date in the text and were left undated.
#
# Precision note — two looser versions were tried and rejected by manual
# sampling before landing on the rule below:
#   1. Matching the LAST "DD de mês de YYYY" anywhere in the text picked up
#      legal citations phrased the same way as a signature date ("da Lei
#      11.419, de 19 de dezembro de 2006") — a citation-keyword filter cut
#      most of these, but not phrasings a keyword list can't anticipate
#      ("será exigido a partir de 1º de janeiro de 2017", a compliance
#      deadline quoted from cited legislation, with no "Lei"/"Decreto" right
#      before it).
#   2. Adding a numeric dd.mm.yyyy fallback for docs with no written-form
#      date matched unrelated numbers inside long technical attachments
#      (a solar-plant inspection report, a realtor's license expiry date
#      inside a property appraisal) — dropped entirely, not worth the risk.
# What actually distinguishes a genuine signature date in these documents,
# reliably, regardless of how long or dense the attachment is: Brazilian
# legal documents close with "<Cidade[/UF]>, DD de mês de YYYY." — so this
# version requires that a place name and a comma sit immediately before the
# date (PLACE_RE below), and takes the LAST such occurrence. Verified
# against 30+ random samples with zero remaining false positives, including
# on multi-hundred-KB attachments that defeated the earlier heuristics.
# Even so, this is heuristic text extraction, not a curated field — treat
# recovered dates as a reasonable placement on the timeline, not as the
# authoritative date of juntada in the docket.

import json, os, re, collections

MONTHS = {
    "janeiro": 1, "fevereiro": 2, "março": 3, "marco": 3, "abril": 4, "maio": 5, "junho": 6,
    "julho": 7, "agosto": 8, "setembro": 9, "outubro": 10, "novembro": 11, "dezembro": 12,
}
MONTH_ALT = "|".join(MONTHS.keys())
WRITTEN_RE = re.compile(r"\b(\d{1,2})\s*(?:º|°)?\s*de\s+(" + MONTH_ALT + r")\s+de\s+(\d{4})\b", re.IGNORECASE)

CITATION_RE = re.compile(
    r"(lei|decreto|resolu[çc][ãa]o|portaria|emenda|provimento|instru[çc][ãa]o\s+normativa|"
    r"medida\s+provis[óo]ria|s[úu]mula|c[óo]digo|ato\s+normativo|normativa)\b", re.IGNORECASE)
CITATION_NUM_RE = re.compile(r"n[ºo°]\.?\s*[\d./-]+\s*,?\s*$", re.IGNORECASE)

# "<Cidade[/UF]>, " immediately before the date — the actual shape of a
# Brazilian legal document's closing line
PLACE_RE = re.compile(r"[A-ZÀ-Ú][A-Za-zà-ú\.]+(?:\s[A-ZÀ-Ú][A-Za-zà-ú\.]+){0,3}(?:\s?[/\-]\s?[A-Z]{2})?,\s*$")

YEAR_MIN, YEAR_MAX = 2015, 2026


def is_citation(text, start):
    window = text[max(0, start - 70):start]
    return bool(CITATION_RE.search(window)) or bool(CITATION_NUM_RE.search(window))


def looks_like_signature(text, start):
    if is_citation(text, start):
        return False
    window = text[max(0, start - 45):start]
    return bool(PLACE_RE.search(window))


def signature_dates(text):
    out = []
    for m in WRITTEN_RE.finditer(text):
        day, mon, year = int(m.group(1)), MONTHS[m.group(2).lower()], int(m.group(3))
        if 1 <= day <= 31 and YEAR_MIN <= year <= YEAR_MAX and looks_like_signature(text, m.start()):
            out.append((m.start(), day, mon, year))
    return out


def recover(docs_json_path="web/docs.json", originals_dir="originals/Parte2"):
    docs = json.load(open(docs_json_path, encoding="utf-8"))
    undated = [r for r in docs if not r.get("date")]

    results = {}
    for r in undated:
        path = os.path.join(originals_dir, r["file"])
        try:
            text = open(path, encoding="utf-8", errors="replace").read()
        except OSError:
            continue
        cands = signature_dates(text)
        if cands:
            _, day, mon, year = cands[-1]
            results[r["id"]] = f"{year:04d}-{mon:02d}-{day:02d}"
    return results


if __name__ == "__main__":
    recovered = recover()
    print(f"recovered {len(recovered)} of the originally-undated documents")

    docs_path = "web/docs.json"
    docs = json.load(open(docs_path, encoding="utf-8"))
    for r in docs:
        if not r.get("date") and r["id"] in recovered:
            r["date"] = recovered[r["id"]]
            r["dateInferred"] = True
    json.dump(docs, open(docs_path, "w", encoding="utf-8"), ensure_ascii=False)

    still_missing = sum(1 for r in docs if not r.get("date"))
    print(f"still undated: {still_missing}")