src/fetch_wikimedia.py

#!/usr/bin/env python3
"""Polish Wikinews, Wikivoyage and Wikibooks from the official dumps, licence per record.

v0.2.0-v0.2.5 labelled every row with one licence (Wikinews CC-BY-2.5, the others
CC-BY-SA-3.0). Each row now carries the label of its page's last revision
(`license_for_revision` in src/rights.py): v0.2.6 rights audit, slayerlabs/polish-dynaword-v2#57.
The dump date is pinned to the one the v0.2.5 text came from (the rebuild changes labels,
ids, authors and dates, not the text); the dump file is checked against the sha1 in
dumpstatus.json and in the dump's sha1sums.txt.

    B=<build dir, never committed>
    python3 src/fetch_wikimedia.py acquire     --out $B --sources wikinews wikivoyage wikibooks
    python3 src/fetch_wikimedia.py prepare     --out $B ...   # $B/plwiki{news,voyage,books}.jsonl.zst
    python3 src/build_dynaword.py --sources wikinews wikivoyage wikibooks --speakleash-dir $B --out .
    python3 src/fetch_wikimedia.py attribution --out $B ...   # data/<source> sidecars + evidence
    python3 src/fetch_wikimedia.py datasheet   --out $B ...   # data/<source>/<source>.md
"""
from __future__ import annotations

import argparse
import bz2
from collections import Counter
from datetime import timedelta
import gzip
import hashlib
import html
import json
from pathlib import Path
import re
import subprocess
import sys
from urllib.parse import quote

sys.path.insert(0, str(Path(__file__).resolve().parent))
from build_plwiki_talk_pl import (  # noqa: E402  shared dump helpers
    UA, digest, load, now, request, request_json, save, sha1_file,
)
from fetch_wikipedia import pages, text_key  # noqa: E402
from language_gate import card_line  # noqa: E402
from rights import (  # noqa: E402
    BY_SA_4_CUTOFF, DROPPED_PAGES, WIKINEWS_BY_4_CUTOFF, check_dropped_reached, license_for_revision,
)
from sources import SOURCES  # noqa: E402

DUMP_DATE = "20260601"  # the dump the v0.2.5 text was cut from (checked: same text set)
MIN_CHARS = 200  # same gate as build_dynaword.py

# WMF Terms of Use in force from 2023-06-07 (CC BY-SA 4.0), and the 2014 revision (CC BY-SA 3.0).
WMF_TOU_NEW = {
    "title": "Wikimedia Terms of Use in force since 7 June 2023 (CC BY-SA 4.0)",
    "url": "https://foundation.wikimedia.org/w/index.php?title=Policy:Terms_of_Use&oldid=554823&action=raw",
    "sha256": "c96ef93b439ff9b8242dc3233f41ede57c0984ef9153fafa51aa5f8cf0678832",
    "effective": "These Terms of Use went into effect on June 7, 2023",
    "statement": "Creative Commons Attribution-ShareAlike 4.0 International License",
}
WMF_TOU_OLD = {
    "title": "Wikimedia Terms of Use of 2014-2023 (CC BY-SA 3.0)",
    "url": "https://foundation.wikimedia.org/w/index.php?title=Policy:Terms_of_Use&oldid=98524&action=raw",
    "sha256": "46617e3576a507b8f83a0cf7b28b60e43817ad93cc5d747a558b503e7d1cc6de",
    "effective": "These Terms of Use went into effect on June 16, 2014",
    "statement": "Creative Commons Attribution-ShareAlike 3.0 Unported License",
}
# pl.wikinews has its own terms page: CC BY, 2.5 until 2024-12-16, 4.0 since.
WIKINEWS_TERMS = {
    "title": "pl.wikinews terms of use (CC BY 2.5 until 2024-12-16, CC BY 4.0 since)",
    "url": "https://pl.wikinews.org/w/index.php?title=Wikinews:Warunki_u%C5%BCycia&oldid=342796&action=raw",
    "sha256": "8b461acb8e0608eb906e59d9669bc003759d7840d82fc934f83d160a72895de3",
    "effective": "udostępnia swoją treść na licencji Creative Commons Uznanie autorstwa 4.0 od 16 grudnia 2024 roku",
    "statement": "a wcześniej na licencji. [http://creativecommons.org/licenses/by/2.5/deed.pl Creative Commons Uznanie autorstwa 2.5]",
}
WIKINEWS_SWITCH = {  # the config change that made 4.0 the site licence (Phabricator T381421)
    "gerrit": "https://gerrit.wikimedia.org/r/changes/operations%2Fmediawiki-config~1101867",
    "phabricator": "https://phabricator.wikimedia.org/T381421",
    "merged_utc": "2024-12-16 08:07",
    "synced_by_utc": "2024-12-16 08:32:24",  # SAL scap sync-world finish; read by the rights audit, not re-read
}

GFDL_ERA_END = "2009-06-15"  # the Wikimedia sites began dual licensing under CC BY-SA 3.0 and the GFDL
# What the label rule rests on: each licence's own clause on adaptations (read from the legal code).
LEGAL_BY_SA_3 = {"url": "https://creativecommons.org/licenses/by-sa/3.0/legalcode", "section": "4(b)(ii)",
                 "statement": "a later version of this License with the same License Elements as this License"}
LEGAL_BY_SA_4 = {"url": "https://creativecommons.org/licenses/by-sa/4.0/legalcode", "section": "3(b)(1)",
                 "statement": "this version or later, or a BY-SA Compatible License"}
LEGAL_BY_2_5 = {"url": "https://creativecommons.org/licenses/by/2.5/legalcode", "section": "4(a)",
                "statement": "You may not sublicense the Work", "absent": "later version"}  # no later-version clause

SITES = {
    "wikinews": {
        "wiki": "plwikinews", "host": "pl.wikinews.org", "pretty": "Polish Wikinews",
        "author": "Wikinews contributors", "family": "CC BY",
        "cutoff": WIKINEWS_BY_4_CUTOFF, "before": "CC-BY-2.5", "after": "CC-BY-4.0",
        "site_rights": "creativecommons.org/licenses/by/4.0/", "terms": [WIKINEWS_TERMS], "legal": [LEGAL_BY_2_5],
        "clock": WIKINEWS_SWITCH["synced_by_utc"], "era_end": "2005-09-25", "domain": "news",
        "v025": (24386, 32524646, 12141355),
    },
    "wikivoyage": {
        "wiki": "plwikivoyage", "host": "pl.wikivoyage.org", "pretty": "Polish Wikivoyage",
        "author": "Wikivoyage contributors", "family": "CC BY-SA",
        "cutoff": BY_SA_4_CUTOFF, "before": "CC-BY-SA-3.0", "after": "CC-BY-SA-4.0",
        "site_rights": "creativecommons.org/licenses/by-sa/4.0/", "terms": [WMF_TOU_NEW, WMF_TOU_OLD], "legal": [LEGAL_BY_SA_3, LEGAL_BY_SA_4],
        "era_end": GFDL_ERA_END, "domain": "travel", "v025": (13645, 45184300, 17128200),
    },
    "wikibooks": {
        "wiki": "plwikibooks", "host": "pl.wikibooks.org", "pretty": "Polish Wikibooks",
        "author": "Wikibooks contributors", "family": "CC BY-SA",
        "cutoff": BY_SA_4_CUTOFF, "before": "CC-BY-SA-3.0", "after": "CC-BY-SA-4.0",
        "site_rights": "creativecommons.org/licenses/by-sa/4.0/", "terms": [WMF_TOU_NEW, WMF_TOU_OLD], "legal": [LEGAL_BY_SA_3, LEGAL_BY_SA_4],
        "era_end": GFDL_ERA_END, "domain": "instructional", "v025": (9112, 38892589, 15571295),
    },
}


def label(key: str, timestamp: str) -> str:
    s = SITES[key]
    return license_for_revision(timestamp, s["cutoff"], s["before"], s["after"])


def page_url(key: str, title: str) -> str:
    return f"https://{SITES[key]['host']}/wiki/" + quote((title or "").replace(" ", "_"), safe="/:")


def dump_base(key: str) -> str:
    return f"https://dumps.wikimedia.org/{SITES[key]['wiki']}/{DUMP_DATE}"


def legal_pin(pin: dict) -> dict:
    """Fetch a legal-code page, check the clause the label rests on, record the page hash."""
    body = request(pin["url"]).content
    text = re.sub(r"\s+", " ", html.unescape(re.sub(r"<[^>]+>", " ", body.decode("utf-8", "replace"))))
    if pin["statement"] not in text or (pin.get("absent") and pin["absent"] in text):
        raise ValueError(f"legal code changed: {pin['url']}")
    return {**pin, "sha256": hashlib.sha256(body).hexdigest()}


def strip(wikitext: str) -> str | None:
    """Wikitext -> plain text, exactly as the v0.2.5 builder did. None when the parser fails."""
    import mwparserfromhell
    try:
        return mwparserfromhell.parse(wikitext).strip_code(normalize=True, collapse=True).strip()
    except Exception:  # noqa: BLE001  parser failure drops the page, counted
        return None


def acquire(out: Path, key: str) -> None:
    site, wiki, base = SITES[key], SITES[key]["wiki"], dump_base(key)
    status = request_json(f"{base}/dumpstatus.json")["jobs"]["articlesdump"]
    if status.get("status") != "done":
        raise ValueError(f"{wiki} articlesdump not done: {status.get('status')}")
    name = f"{wiki}-{DUMP_DATE}-pages-articles.xml.bz2"
    meta = status["files"][name]
    sums = {n: sha for sha, n in (ln.split() for ln in request(f"{base}/{wiki}-{DUMP_DATE}-sha1sums.txt").text.splitlines() if ln.strip())}
    if sums.get(name) != meta["sha1"]:
        raise ValueError(f"sha1sums.txt disagrees with dumpstatus.json for {name}")
    rights = request_json(f"https://{site['host']}/w/api.php?action=query&meta=siteinfo&siprop=rightsinfo&format=json")["query"]["rightsinfo"]
    if site["site_rights"] not in rights.get("url", ""):
        raise ValueError(f"{wiki} rightsinfo changed: {rights}")
    terms = []
    for t in site["terms"]:
        body = request(t["url"]).content
        if hashlib.sha256(body).hexdigest() != t["sha256"] or t["effective"].encode() not in body:
            raise ValueError(f"pinned terms page changed: {t['url']}")
        terms.append(t)
    legal = [legal_pin(p) for p in site["legal"]]
    switch = None
    if key == "wikinews":  # Gerrit prefixes its JSON with ")]}'"
        change = json.loads(request(WIKINEWS_SWITCH["gerrit"]).text.split("\n", 1)[1])
        task = request(WIKINEWS_SWITCH["phabricator"]).text
        if change["status"] != "MERGED" or not change["submitted"].startswith("2024-12-16") or "T381421" not in task:
            raise ValueError(f"Wikinews licence switch evidence changed: {change['status']} {change['submitted']}")
        switch = {**WIKINEWS_SWITCH, "gerrit_status": change["status"], "gerrit_submitted": change["submitted"]}
    raw = out / "raw_dump"
    raw.mkdir(parents=True, exist_ok=True)
    dst = raw / name
    if not dst.exists() or dst.stat().st_size != meta["size"]:
        subprocess.run(["curl", "-fsSL", "--retry", "8", "-C", "-", "-A", UA["User-Agent"],
                        "-o", str(dst), f"{base}/{name}"], check=True)
    if sha1_file(dst) != meta["sha1"]:
        raise ValueError(f"sha1 mismatch: {name}")
    save(out / f"{key}.selection.json", {
        "observed_at": now(), "dump_date": DUMP_DATE, "dump_base": base, "sha1sums_url": f"{base}/{wiki}-{DUMP_DATE}-sha1sums.txt",
        "file": {"name": name, "url": f"{base}/{name}", "size": meta["size"], "sha1": meta["sha1"]},
        "rightsinfo": {**rights, "url_queried": f"https://{site['host']}/w/api.php?action=query&meta=siteinfo&siprop=rightsinfo"},
        "terms": terms, "legal": legal, "switch": switch,
        "cutoff": site["cutoff"].isoformat(), "before": site["before"], "after": site["after"],
    })
    print(f"{key}: {name} sha1 OK")


def prepare(out: Path, key: str) -> None:
    """Label each page, drop exact duplicates keeping the lowest page id and the pages in
    DROPPED_PAGES[key], write the build input. A dropped page's text stays in `seen`, so a later
    page with the same text is dropped as its duplicate."""
    site, dropped_ids = SITES[key], DROPPED_PAGES[key]
    src = out / "raw_dump" / load(out / f"{key}.selection.json")["file"]["name"]
    seen, counts, duplicates, dropped, last_id = set(), Counter(), [], [], 0
    zst = subprocess.Popen(["zstd", "-q", "-f", "-3", "-o", str(out / f"{site['wiki']}.jsonl.zst")], stdin=subprocess.PIPE)
    try:
        with bz2.open(src, "rb") as fh, gzip.open(out / f"{key}.manifest.jsonl.gz", "wt", encoding="utf-8") as manifest:
            for pageid, revid, ns, title, timestamp, redirect, wikitext in pages(fh):
                counts["pages"] += 1
                if int(pageid) <= last_id:  # "lowest page id wins" below relies on dump order
                    raise ValueError(f"dump not in page-id order at {pageid}")
                last_id = int(pageid)
                if ns != "0":
                    continue
                counts["ns0"] += 1
                if redirect:
                    counts["ns0_redirect"] += 1
                elif not wikitext.strip():
                    counts["ns0_empty"] += 1
                elif (text := strip(wikitext)) is None:
                    counts["ns0_strip_error"] += 1
                elif len(text) < MIN_CHARS:
                    counts["ns0_short"] += 1
                elif text_key(text) in seen:
                    counts["exact_duplicate"] += 1
                    duplicates.append(pageid)
                else:
                    seen.add(text_key(text))
                    if int(pageid) in dropped_ids:
                        dropped.append({"pageid": pageid, "title": title, "reason": dropped_ids[int(pageid)]})
                        continue
                    lic = label(key, timestamp)
                    counts["written"] += 1
                    counts[lic] += 1
                    row_id, url = f"{key}_{pageid}", page_url(key, title)
                    zst.stdin.write((json.dumps({
                        "id": row_id, "text": text, "created": timestamp[:10],
                        "license": lic, "author": f"{site['author']}, {url}",
                    }, ensure_ascii=False) + "\n").encode("utf-8"))
                    manifest.write(json.dumps({
                        "id": row_id, "pageid": int(pageid), "revid": int(revid), "title": title, "url": url,
                        "last_revision_at": timestamp, "license": lic,
                    }, ensure_ascii=False) + "\n")
    finally:  # a failed gate must not leave zstd waiting on stdin
        zst.stdin.close()
        status = zst.wait()
    if status:
        raise RuntimeError("zstd failed")
    check_dropped_reached(key, [p["pageid"] for p in dropped])
    counts["dropped_page"] = len(dropped)
    save(out / f"{key}.prepare.json", {"counts": dict(counts), "exact_duplicate_pageids": duplicates,
                                       "dropped_pages": dropped})
    print(key, dict(counts))


def attribution(out: Path, data_dir: Path, key: str) -> None:
    """Sidecars for the rows build_dynaword kept: per-page attribution, build record, licence evidence."""
    import pyarrow.parquet as pq
    site = SITES[key]
    table = pq.read_table(data_dir / f"{key}.parquet", columns=["id", "created", "license", "token_count"]).to_pydict()
    tokens_of = dict(zip(table["id"], table["token_count"]))
    rows = []
    with gzip.open(out / f"{key}.manifest.jsonl.gz", "rt", encoding="utf-8") as fh:
        for line in fh:
            if (rec := json.loads(line))["id"] in tokens_of:
                rows.append((line, rec))
    if len(rows) != len(tokens_of):
        raise ValueError(f"attribution rows {len(rows)} != parquet rows {len(tokens_of)}")
    sidecar = data_dir / f"{key}.attribution.jsonl"
    sidecar.write_text("".join(line for line, _ in rows), encoding="utf-8")
    sel, prep = load(out / f"{key}.selection.json"), load(out / f"{key}.prepare.json")
    tokens, docs = Counter(), Counter()
    for lic, tok in zip(table["license"], table["token_count"]):
        docs[lic] += 1
        tokens[lic] += tok
    day = site["cutoff"].date().isoformat()  # edits on the switch day are labelled `after`
    on_day = [rec["last_revision_at"] for _, rec in rows if rec["last_revision_at"].startswith(day)]
    boundary = {"day": day, "docs_last_edited_that_day": len(on_day)}
    if clock := site.get("clock"):  # known deployment time of the switch, when there is one
        boundary.update({"known_switch_time_utc": clock,
                         "of_which_before_that_time": sum(t.replace("T", " ").rstrip("Z") < clock for t in on_day)})
    scope = {"labels": dict(docs), "tokens": dict(tokens), "created_range": [min(table["created"]), max(table["created"])]}
    old = [rec["id"] for _, rec in rows if rec["last_revision_at"] < site["era_end"]]  # recorded, not acted on
    scope["last_edited_before_era_end"] = {"date": site["era_end"], "docs": len(old), "tokens": sum(tokens_of[i] for i in old)}
    save(data_dir / f"{key}.build.json", {
        "dump_date": sel["dump_date"], "dump_base": sel["dump_base"], "sha1sums_url": sel["sha1sums_url"],
        "file": sel["file"], "rightsinfo": sel["rightsinfo"], "terms": sel["terms"], "switch": sel["switch"],
        "cutoff": sel["cutoff"], "labels": {"before": sel["before"], "after": sel["after"]},
        "discovered_at": sel["observed_at"], "prepare": prep,
        "attribution_rows": len(rows), "attribution_sha256": digest(sidecar.read_bytes()),
    })
    save(data_dir / f"{key}.license-evidence.json", {
        "license": f"{sel['before']} or {sel['after']} per record, by the page's last revision",
        "spdx": [f"https://spdx.org/licenses/{sel['before']}.html", f"https://spdx.org/licenses/{sel['after']}.html"],
        "status": "documented-open-cc-by-sa" if "SA" in sel["after"] else "documented-open-cc-by",
        "review_scope": f"licence of contributors' text on {site['host']} and its version per page; not independent legal advice",
        "checked_at": sel["observed_at"][:10],
        "cutoff": {"at": sel["cutoff"], "before": sel["before"], "after": sel["after"], "boundary_day": boundary},
        "terms": sel["terms"], "legal": sel["legal"], "switch": sel["switch"],
        "site_licence": {"observed_at": sel["observed_at"], "rightsinfo": sel["rightsinfo"]},
        "release_scope": scope,
    })
    print(f"{key}: attribution rows {len(rows)}, labels {dict(docs)}")


def share(part: int, whole: int) -> str:
    return f"{100 * part / whole:.1f}%"


def licensing_text(key: str, ev: dict) -> str:
    """The reading of the licences behind the labels, with the cutoff-day count."""
    site, cut = SITES[key], ev["cutoff"]
    day, n = cut["boundary_day"]["day"], cut["boundary_day"]["docs_last_edited_that_day"]
    next_day = (site["cutoff"] + timedelta(days=1)).date().isoformat()
    before, after = cut["before"], cut["after"]
    if "SA" in after:
        return (
            f"A page last edited on or after the cutoff can still hold 3.0 text from earlier edits;\n"
            f"[CC BY-SA 3.0](https://creativecommons.org/licenses/by-sa/3.0/legalcode) §4(b)(ii) lets such an adaptation be licensed under \"a later\n"
            f"version of this License with the same License Elements\", so it is labelled 4.0. The Terms\n"
            f"of Use give the day of the switch, not the time. The cutoff is midnight UTC, so the\n"
            f"{n} pages last edited on {day} itself are labelled 4.0. If the switch came later\n"
            f"that day, those last edited before it are 3.0 text under a 4.0 label; a {next_day} cutoff\n"
            f"(the citation pass's fallback) would instead put that day's 4.0 edits under a 3.0 label,\n"
            f"which [4.0](https://creativecommons.org/licenses/by-sa/4.0/legalcode) §3(b)(1) (\"this version or later\") does not allow. Neither choice is\n"
            f"exact without the time of the switch. The rule is `license_for_revision` in `src/rights.py`,\n"
            f"tested at the cutoff second, across time zones, and on a missing timestamp, which stops the\n"
            f"build. v0.2.0-v0.2.5 labelled every row `{before}`."
        )
    sw, k = ev["switch"], cut["boundary_day"]["of_which_before_that_time"]
    return (
        f"pl.wikinews has its own terms page, which names CC BY 2.5 until {day} and CC BY 4.0 from that\n"
        f"day. A page last edited on or after the cutoff can still hold\n"
        f"2.5 text from earlier edits, and the [CC BY 2.5](https://creativecommons.org/licenses/by/2.5/legalcode) legal code (§4(a)) has no clause that lets an\n"
        f"adaptation be licensed under a later version (BY-SA 3.0 §4(b)(ii) has one). The 4.0 label for such\n"
        f"pages rests on the site's terms page alone; whether that is enough is for the lawyer.\n"
        f"The terms page gives the day, not the time. The cutoff is midnight UTC, so the {n} pages last edited\n"
        f"on {day} are labelled 4.0. The configuration change was merged {sw['merged_utc']} UTC and deployed\n"
        f"by {sw['synced_by_utc']} UTC ([Gerrit change]({sw['gerrit']}), [{sw['phabricator'].rsplit('/', 1)[1]}]({sw['phabricator']});\n"
        f"the deployment time was read by the rights audit, not re-read), so {k} of those {n} were last edited before it.\n"
        f"The rule is `license_for_revision` in `src/rights.py`, tested at the cutoff second, across time\n"
        f"zones, and on a missing timestamp, which stops the build. v0.2.0-v0.2.5 labelled every row `{before}`."
    )


def datasheet(out: Path, data_dir: Path, key: str) -> None:
    """data/<key>/<key>.md from the stats, build record and licence evidence (custom_datasheet in sources.py)."""
    site = SITES[key]
    st, build, ev = (load(data_dir / f"{key}.{n}.json") for n in ("stats", "build", "license-evidence"))
    scope, prep, file = ev["release_scope"], build["prepare"]["counts"], build["file"]
    before, after = ev["cutoff"]["before"], ev["cutoff"]["after"]
    kept, tok = st["kept"], st["tokens"]
    table = "\n".join(
        f"| {when} | `{lab}` | {scope['labels'].get(lab, 0):,} | {scope['tokens'].get(lab, 0):,} | {share(scope['tokens'].get(lab, 0), tok)} |"
        for when, lab in ((f"before {ev['cutoff']['boundary_day']['day']}", before), (f"{ev['cutoff']['boundary_day']['day']} or later", after)))
    labels = "\n".join(f"| `{lab}` | {scope['labels'].get(lab, 0):,} |" for lab in (before, after))
    v_docs, v_chars, v_tokens = site["v025"]
    same = (kept, st["chars"], tok) == (v_docs, v_chars, v_tokens)
    n_dropped = prep.get("dropped_page", 0)
    v025 = ("The same text set as v0.2.5: documents, characters and tokens are unchanged; only labels, ids, authors and dates differ."
            if same else f"v0.2.5 had {v_docs:,} documents, {v_chars:,} characters and {v_tokens:,} tokens"
            + (f"; the difference is the {n_dropped:,} pages withheld by the rights review." if n_dropped else "."))
    same_set = "the text set is the same" + (" except for the withheld pages" if n_dropped else "")
    old = scope["last_edited_before_era_end"]
    if "SA" in after:
        era = (f"{old['docs']:,} documents ({old['tokens']:,} tokens, {share(old['tokens'], tok)}) were last edited before {old['date']}, when the Wikimedia sites\n"
               f"began dual licensing under CC BY-SA 3.0 and the GFDL. Text first contributed before the 2009 move was\n"
               f"under the GFDL only. They are labelled `{before}`, so only the CC BY-SA grant is relied on; the count is recorded, not acted on.")
    else:
        era = (f"The terms page puts articles created before {old['date']} in the public domain. The dump has only the date of the last\n"
               f"revision, so how many released articles were created then is not known; {old['docs']:,} ({old['tokens']:,} tokens) were last edited before it.\n"
               f"All rows carry a CC BY label; nothing is dropped or relabelled for this.")
    review = ""
    if (path := data_dir / f"{key}.gfdl-review.json").exists():
        r = load(path)
        review = f"\n\n{r['summary']} Per-document notes: `{path.name}`."
    skipped = [f"{prep.get('ns0_redirect', 0):,} redirects", f"{prep.get('ns0_empty', 0):,} empty pages",
               f"{prep.get('ns0_strip_error', 0):,} parser failures", f"{prep.get('ns0_short', 0):,} shorter than {MIN_CHARS} characters"]
    chain = (f"{prep['pages']:,} pages scanned → {prep['ns0']:,} in namespace 0 → " + ", ".join(skipped) + " skipped"
             + (f", {prep['exact_duplicate']:,} exact duplicates dropped" if prep.get("exact_duplicate") else ", no exact duplicates")
             + (f", {n_dropped:,} pages withheld by the rights review (`DROPPED_PAGES` in `src/rights.py`; ids, titles and reasons in `{key}.build.json`)"
                if n_dropped else "")
             + f" → {prep['written']:,} read by the build → {kept:,} released.")
    terms = "; ".join(f"[{t['title']}]({t['url']})" for t in ev["terms"])
    with (data_dir / f"{key}.attribution.jsonl").open(encoding="utf-8") as fh:
        example = min((json.loads(next(fh))["url"] for _ in range(200)), key=len)  # a short real URL
    created = scope["created_range"]
    when = ev["checked_at"]
    (data_dir / f"{key}.md").write_text(f"""# {key}

{site['pretty']}

## Dataset description
- **Source (upstream):** {build['dump_base']}/
- **Domain:** {site['domain']}
- **Language:** Polish (pl)
- **License:** `{before}` or `{after}` per record (by last-revision date)
- **Created (range):** {created[0]}, {created[1]} (date of the page's last revision in the dump)
- **Added:** {SOURCES[key]['added']}; rebuilt {when} (UTC) from the official Wikimedia dump

## Licensing — traceable basis
Contributors' text on {site['host']} is licensed {site['family']}: {terms} (read {when} UTC). The site licence
(`rightsinfo`) reads "{ev['site_licence']['rightsinfo']['text']}". Each row is labelled by its page's last revision in the dump:

| last revision (UTC) | `license` | documents | tokens | share of tokens |
|---|---|---:|---:|---:|
{table}

{licensing_text(key, ev)}

{era}{review}

Evidence: `{key}.license-evidence.json`.

## Attribution
Each row's `author` is `{site['author']}, <page URL>`, e.g.
`{site['author']}, {example}`. The page history names the contributors; {ATTRIBUTION[key]}
`{key}.attribution.jsonl` has one line per row: `id`, `pageid`, `revid` (the revision
in the dump), `title`, `url`, `last_revision_at`, `license`. The row `id` is
`{key}_<page id>` and `created` is the date of `last_revision_at`.

The text is modified: wiki markup is converted to plain text (see Provenance).

## Provenance
Parsed from the official Wikimedia dump `{file['name'][:-8]}` by
`src/fetch_wikimedia.py`; the file ({file['size']:,} bytes, sha1 `{file['sha1']}`) was matched against the
dump's `dumpstatus.json` and `sha1sums.txt` (`{key}.build.json`). v0.2.0-v0.2.5 used the same dump
date; {same_set} (the rebuild adds the last-revision time).

Articles (namespace 0) only. Redirects, empty pages and pages shorter than {MIN_CHARS} characters are skipped.
Wikitext is converted with mwparserfromhell `strip_code(normalize=True, collapse=True)`, as in v0.2.5.
Templates and comments are dropped. `<ref>` bodies, category links (`Kategoria:...`) and file-link
options (`thumb|...`) are not removed and stay in the text; the wikipedia fetcher's `strip_article`
removes them, this one does not, because that would change the text set. Of two pages with the same
text (sha1 of the stripped text), the lower page id is kept.

{chain}

## Statistics
| documents | characters | tokens (tiktoken proxy) |
|---:|---:|---:|
| {kept:,} | {st['chars']:,} | {tok:,} |

{v025}

## Per-document license metadata
| license | documents |
|---|---:|
{labels}

Author metadata present for **{st['authors_with_value']:,}** documents.

## Filters applied (build_dynaword.py)
- drop documents < 200 chars: **{st['drop_short']:,}** (the fetcher already dropped {prep.get('ns0_short', 0):,})
{card_line(st)}
- exact cross-source dedup (sha1): **{st['drop_dup']:,}**
- OCR alpha-ratio < 0.70 (OCR sources only): **{st['drop_ocr']:,}**
- read {st['read']:,} → kept {kept:,}

Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
Llama-3 count is computed at release.

## Not checked
{NOT_CHECKED[key]}
""", encoding="utf-8")
    print(f"{key}: datasheet written")


ATTRIBUTION = {
    "wikinews": "the terms page asks for a link to the source article, which the URL gives.",
    "wikivoyage": "the Terms of Use (section 7b) accept a hyperlink or URL to the page as attribution.",
    "wikibooks": "the Terms of Use (section 7b) accept a hyperlink or URL to the page as attribution.",
}
NOT_CHECKED = {
    "wikinews": """- pastes from copyrighted sources, including articles copied from press agencies or other publishers (no sample was run)
- the 2.5 → 4.0 label of pages edited after the cutoff (a reading of the site's terms page, not a lawyer's opinion;
  the lawyer's answer to E1, the wiki 3.0/4.0 labelling, is still pending, and whether it extends to 2.5 → 4.0 is not settled),
  and the time of day of the switch
- how many released articles were created before 2005-09-25, the public-domain date in the terms page
- text imported from elsewhere under another licence or with extra attribution requirements
- quotations inside articles (Art. 29 of the Polish copyright act)
- personal data (the NERGAL scrub runs before release)""",
    "wikivoyage": """- pastes from copyrighted sources (no sample was run)
- the 3.0 → 4.0 label of pages edited after the cutoff (a reading of 3.0 §4(b)(ii), not a lawyer's opinion;
  the lawyer's answer to E1, the wiki 3.0/4.0 labelling, is still pending), and the time of day of the switch
- text imported from elsewhere under another licence or with extra attribution requirements
- text first contributed under GFDL before the 2009 move to CC BY-SA, and GFDL dual
  licensing; only the CC BY-SA grant is relied on
- quotations inside articles (Art. 29 of the Polish copyright act)
- personal data (the NERGAL scrub runs before release)""",
}
NOT_CHECKED["wikibooks"] = NOT_CHECKED["wikivoyage"]


def main() -> None:
    ap = argparse.ArgumentParser()
    ap.add_argument("step", choices=["acquire", "prepare", "attribution", "datasheet"])
    ap.add_argument("--out", required=True, type=Path)
    ap.add_argument("--sources", nargs="+", default=list(SITES), choices=list(SITES))
    ap.add_argument("--data-root", type=Path, default=Path(__file__).resolve().parent.parent / "data")
    a = ap.parse_args()
    a.out.mkdir(parents=True, exist_ok=True)
    for key in a.sources:
        data_dir = a.data_root / key
        {"acquire": lambda: acquire(a.out, key), "prepare": lambda: prepare(a.out, key),
         "attribution": lambda: attribution(a.out, data_dir, key),
         "datasheet": lambda: datasheet(a.out, data_dir, key)}[a.step]()


if __name__ == "__main__":
    main()