#!/usr/bin/env python3
"""Deterministic per-recipe CVE/advisory scan — an ADDITIVE pre-step for /recipe-upgrade.

WHY THIS EXISTS (2026-08-10): gitea 1.27.1 fixed two CVSS-9.8 RCEs (CVE-2026-60004,
CVE-2026-59774). Our weekly report showed gitea's CVE count as "1", then "none". The upgrade
subagent had scanned the GitHub *release notes*, which mention neither; the two CVEs were announced
only in the vendor's blog security section. The report generator then derived security content from
those notes plus model knowledge — and the model's training predates the CVEs. Nothing in the
pipeline ever queried an advisory source, so a critical CVE that is newer than the model and absent
from the changelog was invisible by construction.

WHAT IT DOES NOT DO: it does not replace or alter any existing security analysis. It is a strictly
ADDITIONAL evidence source whose findings are unioned into the CVE count.

SOURCES (measured against the gitea case before being chosen):
  1. GitHub Security Advisories API — repos/<owner>/<repo>/security-advisories. PRIMARY: carries
     CVE id, GHSA id, severity AND vulnerable/patched version ranges, so "fixed by THIS upgrade" is
     computable rather than guessed. Found both gitea CVEs. Derived from the source-repo URLs the
     per-recipe registry already records — no new per-recipe config needed.
  2. Vendor release/security pages — every URL in cc-ci-plan/upstream/<recipe>.md, fetched and
     regex-scanned for CVE ids. This is what would have caught gitea: the vendor blog names both,
     while the GitHub releases page names neither. Add vendor security/announcement URLs to the
     registry to widen this.
  3. OSV.dev — supplementary, best-effort, only when the recipe declares an ecosystem/package
     mapping below. NOTE: for gitea, OSV returned only Go *dependency* advisories and 404'd on both
     application CVEs; NVD's API had them neither by CPE, CVE id, nor keyword. Advisory databases
     lag the vendor — which is exactly why (1) and (2) lead.

Every source reports its own status, so "checked, none found" is never confused with "not checked".

Usage:
  advisory-scan.py <recipe> [--from <version>] [--to <version>] [--json] [--registry DIR]

  --from/--to are the app versions being upgraded between (e.g. 1.26.2 -> 1.27.1). When given, each
  advisory is classified fixed-by-this-upgrade / still-open / older. Without them everything known
  is listed unclassified. Exits 0 even when sources fail (informational; failures are reported).
"""

from __future__ import annotations

import argparse
import gzip
import json
import os
import re
import sys
import urllib.error
import urllib.request

REGISTRY_DIR = os.environ.get("CCCI_UPSTREAM_REGISTRY", "/srv/cc-ci/cc-ci-plan/upstream")
UA = "cc-ci-advisory-scan (+https://git.autonomic.zone/recipe-maintainers/cc-ci)"
TIMEOUT = int(os.environ.get("ADVISORY_SCAN_TIMEOUT", "45"))
CVE_RE = re.compile(r"CVE-\d{4}-\d{4,7}")
# Leading-version-component jump that means the scheme changed (semver → calver).
SCHEME_JUMP = 100
# A patched_versions field that names no usable version. GitHub carries these verbatim from the
# vendor: redis publishes "TBD" for 11 advisories and "6.2.X, 7.2.X, 7.4.X" for another, and their
# vulnerable_version_range is open-ended ("All", ">= 7.0.0"), so the fix version is NOT recoverable.
# Such an advisory must be reported as INDETERMINATE, never silently counted as "not fixed" — that
# would assert an upgrade did not fix something we simply cannot judge.
PLACEHOLDER_RE = re.compile(r"\bTBD\b|\bunknown\b|\bnone\b|\d+\.[Xx]\b|\?", re.I)

# Optional OSV mappings: recipe -> (ecosystem, package). Supplementary only (see module docstring).
OSV_PACKAGES: dict[str, tuple[str, str]] = {
    "gitea": ("Go", "code.gitea.io/gitea"),
    "n8n": ("npm", "n8n"),
}


def _github_token() -> str | None:
    """Read-only GitHub token, for the API rate limit ONLY (60/hr anonymous → 5000/hr with a token).

    Env `GITHUB_TOKEN` wins; otherwise the file at `GITHUB_TOKEN_FILE` (default
    /srv/cc-ci/.github-token, chmod 600, never in git). Reading PUBLIC security advisories needs NO
    scopes at all — create a classic PAT with every box unticked, or a fine-grained token limited to
    "Public repositories: read". Do NOT grant repo/write scopes: this tool only ever GETs advisories.
    A missing token is not an error — the scan simply runs anonymously and will report sources as
    failed once the 60/hr limit bites, which is visible rather than silent.
    """
    tok = os.environ.get("GITHUB_TOKEN")
    if tok:
        return tok.strip()
    path = os.environ.get("GITHUB_TOKEN_FILE", "/srv/cc-ci/.github-token")
    try:
        with open(path) as f:
            return f.read().strip() or None
    except OSError:
        return None


def _gh_paginate(url: str, hdrs: dict, max_pages: int = 20):
    """Yield every row from a GitHub list endpoint, following Link rel=\"next\" cursors."""
    seen_keys = set()
    for _ in range(max_pages):
        req = urllib.request.Request(url, headers={"User-Agent": UA, **hdrs})
        with urllib.request.urlopen(req, timeout=TIMEOUT) as r:
            rows = json.load(r)
            link = r.headers.get("Link", "") or ""
        fresh = 0
        for a in rows:
            k = a.get("ghsa_id") or json.dumps(a, sort_keys=True)[:120]
            if k not in seen_keys:
                seen_keys.add(k); fresh += 1
                yield a
        nxt = None
        for part in link.split(","):
            if 'rel="next"' in part:
                nxt = part.split(";")[0].strip().strip("<>")
        if not nxt or fresh == 0:
            return
        url = nxt


def _tag_date(owner: str, repo: str, version: str | None) -> str | None:
    """Publish date of a release tag, for DATE-BASED classification (see classify_by_date).

    Version strings cannot be ordered across a scheme change (semver → calver), but tag dates
    always can. Tries the common tag spellings; returns an ISO timestamp or None."""
    if not version:
        return None
    hdrs = {"Accept": "application/vnd.github+json"}
    tok = _github_token()
    if tok:
        hdrs["Authorization"] = f"Bearer {tok}"
    for tag in (f"v{version}", version):
        try:
            ref = json.loads(_fetch(f"https://api.github.com/repos/{owner}/{repo}/git/ref/tags/{tag}", hdrs))
            obj = ref.get("object", {})
            sha, typ = obj.get("sha"), obj.get("type")
            if typ == "tag":
                t = json.loads(_fetch(f"https://api.github.com/repos/{owner}/{repo}/git/tags/{sha}", hdrs))
                if t.get("tagger", {}).get("date"):
                    return t["tagger"]["date"]
                sha = t.get("object", {}).get("sha")
            c = json.loads(_fetch(f"https://api.github.com/repos/{owner}/{repo}/commits/{sha}", hdrs))
            return c["commit"]["committer"]["date"]
        except Exception:  # noqa: BLE001 — try the next spelling
            continue
    return None


def _fetch(url: str, headers: dict | None = None) -> str:
    h = {"User-Agent": UA, "Accept-Encoding": "gzip"}
    h.update(headers or {})
    with urllib.request.urlopen(urllib.request.Request(url, headers=h), timeout=TIMEOUT) as r:
        raw = r.read()
        if r.headers.get("Content-Encoding") == "gzip":
            raw = gzip.decompress(raw)
        return raw.decode(errors="replace")


def _vkey(v: str | None) -> tuple:
    """Loose version ordering key: leading integers of each dot-part ('1.27.1-rootless' -> (1,27,1))."""
    if not v:
        return ()
    v = v.strip().lstrip("vV").split("+")[0]
    out = []
    for part in re.split(r"[.\-_]", v):
        m = re.match(r"^\d+", part)
        if m:
            out.append(int(m.group()))
        elif out:
            break
    return tuple(out)


def _within(kf: tuple, kt: tuple, c: tuple) -> bool:
    """Is patched-version `c` inside the window (kf, kt] — exclusive lower, inclusive upper?

    Compares ZERO-PADDED to equal length, so "18" == "18.0" == "18.0.0" the way semver means it.
    Without the padding, plain tuple order says (18,) < (18,0), i.e. a CVE patched in 18.0 falls
    OUTSIDE a window ending at 18 — and bare major tags are the norm for sidecars (postgres:18,
    redis:8-alpine), so that silently dropped real fixes. Padding also keeps the upper bound
    conservative: a fix in 18.5 is still outside a window ending at "18", because nothing proves
    which 18.x a floating tag resolved to.
    """
    n = max(len(kf), len(kt), len(c))
    pad = lambda t: t + (0,) * (n - len(t))
    return pad(kf) < pad(c) <= pad(kt)


def registry_urls(recipe: str, registry_dir: str) -> tuple[list[str], str | None]:
    path = os.path.join(registry_dir, f"{recipe}.md")
    try:
        with open(path) as f:
            text = f.read()
    except OSError:
        return [], None
    urls = []
    for u in re.findall(r"https?://[^\s)|\]]+", text):
        # The registry is MARKDOWN: urls appear inside `backticks`, 'quotes', **bold**, and at the
        # end of sentences. Trailing punctuation captured into the url makes the fetch 404 and the
        # recipe render '?' for no real reason — that is what put immich and n8n in the unknown
        # column on 2026-08-07 (https://docs.n8n.io/release-notes/` ← note the backtick).
        u = u.rstrip("`'\"*.,;:>)")
        if u and u not in urls:
            urls.append(u)
    return urls, path


def github_advisories(urls: list[str]) -> list[dict]:
    """Query GitHub Security Advisories for every github.com/<owner>/<repo> in the registry."""
    seen, results = set(), []
    for u in urls:
        m = re.match(r"https?://github\.com/([^/]+)/([^/#?]+)", u)
        if not m:
            continue
        owner, repo = m.group(1), m.group(2).removesuffix(".git")
        if (owner, repo) in seen:
            continue
        seen.add((owner, repo))
        api = f"https://api.github.com/repos/{owner}/{repo}/security-advisories?per_page=100"
        hdrs = {"Accept": "application/vnd.github+json"}
        tok = _github_token()
        if tok:
            hdrs["Authorization"] = f"Bearer {tok}"
        entry = {"source": f"github-advisories:{owner}/{repo}", "status": "ok", "advisories": []}
        try:
            # PAGINATE. This endpoint caps at 100 per response and IGNORES ?page= — it returns the
            # same rows again, which silently truncates busy projects (discourse has 286; a hand
            # count on 2026-08-10 found 123 CVEs in one upgrade window that a single page missed).
            # Follow the Link rel="next" cursor to exhaustion instead.
            for a in _gh_paginate(api, hdrs):
                # An advisory carries ONE ENTRY PER PATCHED RELEASE LINE. n8n patches three
                # (1.123.32, 2.17.4, 2.18.1); reading only vulnerabilities[0] silently dropped the
                # line our deployment is actually on, so CVE-2026-42231/42232 classified as
                # out-of-window. Keep them ALL and let the classifier match any of them.
                vulns = a.get("vulnerabilities") or []
                entry["advisories"].append(
                    {
                        "cve": a.get("cve_id"),
                        "ghsa": a.get("ghsa_id"),
                        "severity": a.get("severity"),
                        "summary": (a.get("summary") or "")[:200],
                        "vulnerable_range": "; ".join(
                            filter(None, (v.get("vulnerable_version_range") for v in vulns))
                        ) or None,
                        "patched": "; ".join(
                            filter(None, (v.get("patched_versions") for v in vulns))
                        ) or None,
                        "url": a.get("html_url"),
                        "published_at": a.get("published_at"),
                        # The list response ALREADY carries the prose. Keep it: the adjudication
                        # pass needs it, and re-fetching per advisory costs a request each.
                        "description": (a.get("description") or "")[:4000],
                        "cvss": ((a.get("cvss") or {}).get("vector_string")),
                    }
                )
        except urllib.error.HTTPError as e:
            # 404 = this repo simply publishes no security advisories (e.g. sidecar images like
            # pgautoupgrade). That is a BENIGN ABSENCE, not a failed check — conflating the two
            # would push nearly every recipe to "unknown" and make the ? signal meaningless again.
            entry["status"] = "no-advisories-published" if e.code == 404 else f"error: HTTP {e.code}"
        except Exception as e:  # noqa: BLE001 — a genuinely dead source must be REPORTED, never silent
            entry["status"] = f"error: {type(e).__name__}: {e}"
        results.append(entry)
    return results


def vendor_pages(urls: list[str]) -> list[dict]:
    """Fetch each registry URL and regex out CVE ids, with a little surrounding context."""
    out = []
    for u in urls:
        if u.startswith("https://api.github.com"):
            continue
        if re.search(r"[<>{}]|\bVERSION\b|\bvX\.Y\.Z\b", u):
            # Registry entries sometimes carry TEMPLATE urls for humans
            # (…/changelog/v<VERSION>/). They are documentation, not fetchable — skipping them is
            # correct; counting them as failures would wrongly mark the recipe's count unreliable.
            out.append({"source": u, "status": "skipped: template URL (not fetchable)", "cves": [], "context": {}})
            continue
        entry = {"source": u, "status": "ok", "cves": [], "context": {}}
        try:
            text = _fetch(u)
            plain = re.sub(r"<[^>]+>", " ", text)
            for cve in sorted(set(CVE_RE.findall(plain))):
                entry["cves"].append(cve)
                i = plain.find(cve)
                entry["context"][cve] = re.sub(r"\s+", " ", plain[max(0, i - 160) : i + 200]).strip()
        except Exception as e:  # noqa: BLE001
            entry["status"] = f"error: {type(e).__name__}: {e}"
        out.append(entry)
    return out


def osv(recipe: str, version: str | None) -> dict | None:
    pkg = OSV_PACKAGES.get(recipe)
    if not pkg or not version:
        return None
    eco, name = pkg
    entry = {"source": f"osv:{eco}/{name}@{version}", "status": "ok", "cves": []}
    try:
        body = json.dumps({"package": {"name": name, "ecosystem": eco}, "version": version}).encode()
        req = urllib.request.Request(
            "https://api.osv.dev/v1/query", data=body,
            headers={"Content-Type": "application/json", "User-Agent": UA}, method="POST",
        )
        with urllib.request.urlopen(req, timeout=TIMEOUT) as r:
            data = json.load(r)
        ids = set()
        for v in data.get("vulns", []):
            for a in [v.get("id")] + (v.get("aliases") or []):
                if a and a.startswith("CVE"):
                    ids.add(a)
        entry["cves"] = sorted(ids)
    except Exception as e:  # noqa: BLE001
        entry["status"] = f"error: {type(e).__name__}: {e}"
    return entry


_RELEASE_CACHE: dict[str, list[tuple[str, str]]] = {}


def _releases(owner: str, repo: str, max_pages: int = 4) -> list[tuple[str, str]]:
    """[(tag, body)] for a repo's GitHub releases, cached per repo for the process."""
    key = f"{owner}/{repo}"
    if key in _RELEASE_CACHE:
        return _RELEASE_CACHE[key]
    hdrs = {"Accept": "application/vnd.github+json"}
    tok = _github_token()
    if tok:
        hdrs["Authorization"] = f"Bearer {tok}"
    out: list[tuple[str, str]] = []
    try:
        for rel in _gh_paginate(
            f"https://api.github.com/repos/{owner}/{repo}/releases?per_page=100", hdrs, max_pages
        ):
            out.append((rel.get("tag_name") or "",
                        f"{rel.get('name') or ''}\n{rel.get('body') or ''}"))
    except Exception:  # noqa: BLE001 — best effort; absence just leaves advisories undetermined
        pass
    _RELEASE_CACHE[key] = out
    return out


def _source_repo(source: str) -> tuple[str, str] | None:
    """(owner, repo) for a source, whether it is an advisory feed or a vendor page on GitHub.

    A CVE that appears ONLY on a vendor page still deserves the release-note method when that page
    lives on GitHub — mailu announces its Roundcube CVEs on github.com/Mailu/Mailu/releases and
    nowhere structured, so requiring an advisory feed sent a deterministic case to pass 2."""
    if source.startswith("github-advisories:"):
        owner, _, repo = source.split(":", 1)[1].partition("/")
        return (owner, repo) if owner and repo else None
    m = re.match(r"https?://github\.com/([^/]+)/([^/#?]+)", source)
    return (m.group(1), m.group(2).removesuffix(".git")) if m else None


def release_fix_versions(source: str, cve: str) -> list[str]:
    """Release tags whose notes NAME this CVE — a deterministic fix version when the advisory has none.

    Vendors routinely publish an advisory with `patched_versions: "TBD"` and then name the CVE in the
    release notes of every branch that got the fix. redis does exactly this: all 12 of its advisories
    that discourse's redis bump crosses carry TBD, yet each is named in concrete releases
    (CVE-2025-32023 → 6.2.19, 7.2.10, 7.4.5, 8.0.3, 8.2.0). Ignoring that evidence undercounted
    discourse by 12 CVEs, so this is checked BEFORE giving up on an advisory.
    """
    ref = _source_repo(source)
    if not ref:
        return []
    return [tag for tag, body in _releases(*ref) if cve in body]


def advisory_text(ghsa: str, source: str | None = None) -> dict:
    """Full text of one advisory, for the ADJUDICATION pass (see adjudication_block).

    The structured `patched_versions` field is often "TBD" while the prose description and the
    linked references DO state where the fix landed. That prose is not machine-parseable in general
    — which is the point: it is collected here for a MODEL to judge, not for a regex."""
    hdrs = {"Accept": "application/vnd.github+json"}
    tok = _github_token()
    if tok:
        hdrs["Authorization"] = f"Bearer {tok}"
    out = {"ghsa": ghsa, "status": "ok"}
    # Repo-scoped FIRST. Many repository advisories are never mirrored into the global GitHub
    # Advisory Database, so /advisories/<ghsa> 404s for them (all 12 redis ones, for instance)
    # while /repos/<owner>/<repo>/security-advisories/<ghsa> returns the full record.
    cands = []
    if source and source.startswith("github-advisories:"):
        cands.append(f"https://api.github.com/repos/{source.split(':',1)[1]}/security-advisories/{ghsa}")
    cands.append(f"https://api.github.com/advisories/{ghsa}")
    try:
        a, last = None, None
        for u in cands:
            try:
                a = json.loads(_fetch(u, hdrs)); break
            except Exception as ex:  # noqa: BLE001 — try the next endpoint
                last = ex
        if a is None:
            raise last or RuntimeError("no advisory endpoint responded")
        out.update({
            "summary": a.get("summary"),
            "description": (a.get("description") or "")[:4000],
            "severity": a.get("severity"),
            "published_at": a.get("published_at"),
            "references": [r for r in (a.get("references") or [])][:12] or (
                [a.get("html_url")] if a.get("html_url") else []),
            "cvss": (a.get("cvss") or {}).get("vector_string"),
            "vulnerabilities": [
                {"package": (v.get("package") or {}).get("name"),
                 "vulnerable_version_range": v.get("vulnerable_version_range"),
                 "first_patched_version": v.get("first_patched_version")}
                for v in (a.get("vulnerabilities") or [])
            ],
        })
    except Exception as e:  # noqa: BLE001
        out["status"] = f"error: {type(e).__name__}: {e}"
    return out


MAX_ADJUDICATE = int(os.environ.get("ADVISORY_SCAN_MAX_ADJUDICATE", "25"))
# Compact review rows for advisories pass 1 DID decide. Pass 2 sees these too, so a wrong
# deterministic verdict can be caught rather than inherited.
MAX_REVIEW_ROWS = int(os.environ.get("ADVISORY_SCAN_MAX_REVIEW", "400"))


def needs_judgement(rep: dict) -> list[str]:
    """CVEs the deterministic pass could not decide — the input set for the adjudication pass.

    Two kinds, both real gaps rather than noise:
      1. INDETERMINATE — from an image WITH a window, but no fix version is knowable (advisory says
         `TBD`/`7.4.X`, range is open-ended, and no release note names it).
      2. VENDOR-PAGE-ONLY — a CVE seen only on a vendor security page, with no structured advisory
         behind it at all. gitea's two CVSS-9.8 RCEs are this shape. They carry no version data, so
         no arithmetic can place them, but the page's prose usually states the fixed release.
    """
    windows = rep.get("windows") or {}
    out = list(rep.get("indeterminate") or [])
    for cve in rep.get("unclassified") or []:
        srcs = rep["cves"][cve]["sources"]
        if not any(s.startswith("github-advisories:") for s in srcs) and cve not in out:
            out.append(cve)
    return sorted(out)


def evidence_bundle(rep: dict, cve: str) -> dict:
    """EVERY deterministic signal held about one CVE, gathered for a model to weigh.

    Pass 1 collects; pass 2 judges. Nothing here interprets — it assembles what was measured, so the
    judgement is made against evidence rather than recollection (the exact failure that let two
    CVSS-9.8 gitea RCEs be published as "none": the report leaned on model knowledge that predated
    them, and no source had been queried at all).
    """
    e = rep["cves"][cve]
    src = e["sources"][0]
    windows = rep.get("windows") or {}
    win = next((windows[s] for s in e["sources"] if s in windows), None)
    ev = {
        "cve": cve,
        "severity": e.get("severity"),
        "cvss": e.get("cvss"),
        "sources": e["sources"],
        "window": win,
        "why_undecided": ("no fix version published and no release note names it"
                          if cve in (rep.get("indeterminate") or [])
                          else "seen only on a vendor page — no structured advisory, no version data"),
        "patched_as_published": e.get("patched"),
        "vulnerable_range_as_published": e.get("vulnerable_range"),
        "summary": e.get("context"),
        "description": e.get("description"),
        "advisory_url": e.get("url"),
        # Release tags NAMING this CVE, whether or not they fall in the window — the model may
        # reason about branch lines the arithmetic deliberately would not.
        "releases_naming_it": release_fix_versions(src, cve) if src.startswith("github-advisories:") else [],
        "references": [],
    }
    if e.get("ghsa"):
        t = advisory_text(e["ghsa"], src)
        if t.get("status") == "ok":
            ev["references"] = t.get("references") or []
            ev["description"] = t.get("description") or ev["description"]
            ev["summary"] = t.get("summary") or ev["summary"]
            ev["affected"] = t.get("vulnerabilities") or []
        else:
            ev["detail_fetch"] = t.get("status")
    return ev


def adjudication_block(rep: dict) -> str:
    """SECOND PASS: present the collected evidence and ask for a judgement on each open case.

    This block decides nothing. The deterministic count stands as a FLOOR; a verdict here may only
    ADD to it, matching the rule that this scan raises a count on evidence but never lowers one.
    """
    todo = needs_judgement(rep)
    if not todo:
        return ""
    shown, dropped = todo[:MAX_ADJUDICATE], max(0, len(todo) - MAX_ADJUDICATE)
    L = ["", "---", "",
         f"## Adjudication pass — {len(todo)} advisory/advisories need judgement", "",
         "Pass 1 collected the evidence below deterministically and could NOT decide these cases. "
         "Weigh the evidence and decide each one.", "",
         "**For each: did the version move in its window fix it?** Answer **FIXED** / **NOT-FIXED** / "
         "**STILL-UNKNOWN**, each with a one-line reason **citing the evidence shown** — a fixed "
         "release named in the text, a branch line, an affected range. Add every FIXED to the "
         "recipe's CVE count; the deterministic number is a floor, not a total. If the evidence does "
         "not settle it, say STILL-UNKNOWN: do NOT infer from memory of the project, and never "
         "record an undecided CVE as unaffected.", ""]
    for src, win in (rep.get("windows") or {}).items():
        L.append(f"- window: `{src.split(':',1)[-1]}` {win['from']} → {win['to']}")
    if dropped:
        L += ["", f"⚠ Showing the first {MAX_ADJUDICATE} of {len(todo)}; **{dropped} not shown** "
                  f"(raise ADVISORY_SCAN_MAX_ADJUDICATE). The unshown remain undetermined — do not "
                  f"treat them as absent."]
    L.append("")
    for cve in shown:
        ev = evidence_bundle(rep, cve)
        L.append(f"### {cve} — {ev['severity'] or '?'}")
        L.append(f"- undecided because: {ev['why_undecided']}")
        L.append(f"- source: `{ev['sources'][0]}`"
                 + (f" · window {ev['window']['from']} → {ev['window']['to']}" if ev["window"] else
                    " · **no version window** for this image"))
        L.append(f"- patched_versions as published: `{ev['patched_as_published']}`")
        L.append(f"- vulnerable_range as published: `{ev['vulnerable_range_as_published']}`")
        if ev["releases_naming_it"]:
            L.append(f"- **releases naming this CVE**: {', '.join(ev['releases_naming_it'][:14])}")
        for v in ev.get("affected") or []:
            L.append(f"- affects `{v.get('package')}` {v.get('vulnerable_version_range')} — "
                     f"first_patched_version: {v.get('first_patched_version')}")
        if ev["references"]:
            L.append(f"- references: {', '.join(r.strip() for r in ev['references'][:6])}")
        if ev.get("detail_fetch"):
            L.append(f"- ⚠ detail fetch failed: {ev['detail_fetch']} (evidence below is from pass 1)")
        if ev["summary"]:
            L += ["", f"> {ev['summary']}"]
        if ev["description"]:
            L += ["", "```", (ev["description"] or "").strip()[:2000], "```"]
        L.append("")

    # ── everything pass 1 DID decide, with the evidence behind each verdict ──────────────────────
    # Pass 2 must see the whole picture, not only the leftovers: a deterministic verdict can still
    # be wrong (a mis-parsed range, a release note that names a CVE without fixing it), and only a
    # reader with the evidence in front of it can catch that.
    decided = []
    for cve in rep.get("fixed_by_this_upgrade") or []:
        e = rep["cves"][cve]
        decided.append((cve, "COUNTED", e))
    for cve, e in sorted(rep.get("cves", {}).items()):
        if e.get("classification") == "outside-window":
            decided.append((cve, "excluded (outside window)", e))
    if decided:
        shown_rows, dropped_rows = decided[:MAX_REVIEW_ROWS], max(0, len(decided) - MAX_REVIEW_ROWS)
        L += ["---", "",
              f"## Pass 1 decisions — {len(decided)} already judged deterministically", "",
              "Review these too. If any verdict looks wrong given its evidence, say so and explain; "
              "a correction here changes the count. Silence means you agree.", ""]
        if dropped_rows:
            L.append(f"⚠ Showing {MAX_REVIEW_ROWS} of {len(decided)}; **{dropped_rows} not shown** "
                     f"(raise ADVISORY_SCAN_MAX_REVIEW).")
            L.append("")
        L += ["| CVE | verdict | severity | patched as published | releases naming it | source |",
              "|---|---|---|---|---|---|"]
        for cve, verdict, e in shown_rows:
            rel = e.get("fix_versions_from_release_notes") or []
            L.append(f"| {cve} | {verdict} | {e.get('severity') or '?'} | "
                     f"{(e.get('patched') or '—')[:60]} | {', '.join(rel[:5]) or '—'} | "
                     f"{e['sources'][0].split(':',1)[-1]} |")
        L.append("")
    return "\n".join(L)


def scan(recipe: str, v_from: str | None, v_to: str | None, registry_dir: str,
         images: list[tuple[str, str, str]] | None = None) -> dict:
    urls, reg_path = registry_urls(recipe, registry_dir)
    report: dict = {
        "recipe": recipe,
        "from": v_from,
        "to": v_to,
        "registry": reg_path,
        "registry_urls": len(urls),
        "sources": [],
        "cves": {},
    }
    if reg_path is None:
        report["sources"].append(
            {"source": f"registry:{recipe}.md", "status": "error: registry file not found"}
        )

    def record(cve: str, src: str, **extra):
        e = report["cves"].setdefault(cve, {"sources": [], "severity": None, "ghsa": None,
                                            "vulnerable_range": None, "patched": None,
                                            "context": None, "published_at": None,
                                            "description": None, "url": None, "cvss": None})
        if src not in e["sources"]:
            e["sources"].append(src)
        for k, v in extra.items():
            if v and not e.get(k):
                e[k] = v

    for entry in github_advisories(urls):
        report["sources"].append({"source": entry["source"], "status": entry["status"],
                                  "found": len(entry.get("advisories", []))})
        for a in entry.get("advisories", []):
            if a.get("cve"):
                record(a["cve"], entry["source"], severity=a.get("severity"), ghsa=a.get("ghsa"),
                       vulnerable_range=a.get("vulnerable_range"), patched=a.get("patched"),
                       context=a.get("summary"), published_at=a.get("published_at"),
                       description=a.get("description"), url=a.get("url"), cvss=a.get("cvss"))

    for entry in vendor_pages(urls):
        report["sources"].append({"source": entry["source"], "status": entry["status"],
                                  "found": len(entry.get("cves", []))})
        for cve in entry.get("cves", []):
            record(cve, entry["source"], context=entry["context"].get(cve))

    for version in filter(None, (v_from, v_to)):
        o = osv(recipe, version)
        if o:
            report["sources"].append({"source": o["source"], "status": o["status"],
                                      "found": len(o.get("cves", []))})
            for cve in o.get("cves", []):
                record(cve, o["source"])

    # Classify against the upgrade window when we know it: an advisory is "fixed by this upgrade"
    # when its patched version is newer than `from` and no newer than `to`.
    #
    # TWO HARD-WON CONSTRAINTS (2026-08-10, discourse reported a false 133):
    #  a) The window belongs to ONE image. Advisories from OTHER repos in the registry (redis,
    #     postgres, nginx sidecars) must NOT be judged by it — redis CVE-2021-21309, patched in
    #     redis 6.0.11, scored as "fixed" because 6.0.11 sits numerically inside discourse's
    #     3.5.3 → 2026.7.1 window. Only the PRIMARY app repo is classified; every other source is
    #     reported as unclassified so a human/agent still sees it but it never inflates the count.
    #  b) A version-SCHEME change (semver → calver, 3.5.3 → 2026.7.1) makes numeric ordering
    #     meaningless: 2025.12.2 compares "newer" than 3.5.3 while shipping earlier. When the
    #     leading component jumps by more than SCHEME_JUMP we refuse to classify and say so,
    #     rather than emitting a confident wrong number.
    # ── Classification ────────────────────────────────────────────────────────────────────────
    # A recipe upgrades SEVERAL images (app + redis/postgres/nginx sidecars), each with its OWN
    # version window. Judging every advisory by the app's window is how discourse once reported a
    # false 133 (34 of them redis CVEs, incl. one patched in redis 6.0.11 in 2021). So each source
    # is classified against ITS OWN window, and the count is the union across windows.
    #
    #   --from/--to            → the PRIMARY app repo (first github source in the registry)
    #   --image NAME=FROM:TO   → any other source whose name contains NAME (repeatable),
    #                            e.g. --image redis=7.4:8.10
    #
    # A source with no window is not classified: its advisories are listed as unclassified so they
    # stay visible without inflating the count.
    gh_sources = [x["source"] for x in report["sources"] if x["source"].startswith("github-advisories:")]
    primary = gh_sources[0] if (gh_sources and (v_from or v_to)) else None
    report["primary_source"] = primary

    windows = {}                                    # source name -> (from, to)
    if primary:
        windows[primary] = (v_from, v_to)
    for key, wf, wt in (images or []):
        for src in gh_sources:
            if key.lower() in src.lower() and src not in windows:
                windows[src] = (wf, wt)
    report["windows"] = {k: {"from": f, "to": t} for k, (f, t) in windows.items()}

    def _classify_window(src, wf, wt):
        """Return (fixed, method, date_window|None, unresolved, indeterminate) for one source.

        `indeterminate` = advisories from this source that the method COULD NOT JUDGE (no usable
        patched version, or no publish date). They are neither counted nor dismissed."""
        kf, kt = _vkey(wf), _vkey(wt)
        # A version-SCHEME change (semver 3.5.3 → calver 2026.7.1) makes numeric ordering
        # meaningless: 2025.12.2 compares "newer" than 3.5.3 while shipping earlier.
        scheme = bool(kf and kt and abs(kt[0] - kf[0]) >= SCHEME_JUMP)
        if not scheme:
            got, undecidable = set(), set()
            for cve, e in report["cves"].items():
                if src not in e["sources"]:
                    continue
                patched = e.get("patched") or ""
                cands = [_vkey(t) for t in re.findall(r"\d+(?:\.\d+)*", patched)]
                if kf and kt and any(_within(kf, kt, c) for c in cands):
                    got.add(cve)
                elif not patched or PLACEHOLDER_RE.search(patched):
                    # No fix version published ("TBD") or only a placeholder ("7.4.X" — which could
                    # be 7.4.1, inside the window). We cannot say either way, so say so.
                    undecidable.add(cve)
            return got, "patched version ranges", None, False, undecidable
        # DATE FALLBACK: release DATES always order, even across a scheme change. Resolve both
        # versions to git tag dates and count advisories PUBLISHED in that window — the method a
        # hand count used to establish discourse 3.5.3 (2025-12-30) → 2026.7.1 (2026-07-31) = 123.
        owner, _, repo = src.split(":", 1)[1].partition("/")
        d_from, d_to = _tag_date(owner, repo, wf), _tag_date(owner, repo, wt)
        if d_from and d_to and d_from < d_to:
            got = {cve for cve, e in report["cves"].items()
                   if src in e["sources"] and e.get("published_at")
                   and d_from < e["published_at"] <= d_to}
            undecidable = {cve for cve, e in report["cves"].items()
                           if src in e["sources"] and not e.get("published_at")}
            return got, "advisory publish date (version scheme changed)", (d_from, d_to), False, undecidable
        return set(), "unresolved", None, True, set()

    fixed_set, methods, date_windows, unresolved_any = set(), {}, {}, False
    indeterminate: set = set()
    for src, (wf, wt) in windows.items():
        got, method, dw, unresolved, undecidable = _classify_window(src, wf, wt)
        indeterminate |= undecidable
        methods[src] = method
        if dw:
            date_windows[src] = {"from": dw[0], "to": dw[1]}
        if unresolved:
            unresolved_any = True
        for cve in got:
            report["cves"][cve]["classification"] = f"fixed-by-this-upgrade ({method}) via {src}"
            fixed_set.add(cve)

    unknown = []
    for cve, e in report["cves"].items():
        if cve in fixed_set:
            continue
        if not any(src in e["sources"] for src in windows):
            e["classification"] = "unclassified: no versions given for this image"
            unknown.append(cve)
        else:
            e.setdefault("classification", "outside-window")
            if e["classification"] == "outside-window":
                pass
            else:
                unknown.append(cve)

    report["classified_by"] = methods
    if date_windows:
        report["date_window"] = date_windows
    # THIRD METHOD: before declaring an advisory undecidable, look for the CVE id in the project's
    # own release notes. A tag that names it, inside the window, IS the fix version the advisory
    # failed to publish. Deterministic and citable — not a judgement call.
    # A window is keyed by advisory-feed source; map it to its repo so a vendor page on the SAME
    # repo can be judged by the same window.
    win_by_repo = {}
    for wsrc, wv in windows.items():
        ref = _source_repo(wsrc)
        if ref:
            win_by_repo[ref] = wv
    resolved_by_release, already_fixed = {}, set()
    candidates = set(indeterminate) | {
        c for c in unknown
        if not any(s in windows for s in report["cves"][c]["sources"])
        and any(_source_repo(s) in win_by_repo for s in report["cves"][c]["sources"])
    }
    for cve in sorted(candidates - fixed_set):
        e = report["cves"][cve]
        for src in e["sources"]:
            ref = _source_repo(src)
            if src not in windows and ref not in win_by_repo:
                continue
            wf, wt = windows[src] if src in windows else win_by_repo[ref]
            kf, kt = _vkey(wf), _vkey(wt)
            if not (kf and kt):
                continue
            naming = release_fix_versions(src, cve)
            hits = [t for t in naming if _within(kf, kt, _vkey(t))]
            if hits:
                fixed_set.add(cve)
                resolved_by_release[cve] = sorted(hits)
                e["classification"] = (f"fixed-by-this-upgrade (named in release notes "
                                       f"{', '.join(sorted(hits))}) via {src}")
                e["fix_versions_from_release_notes"] = sorted(hits)
                break
            # Naming releases exist but ALL predate the version we were already on: the fix shipped
            # before this upgrade, so the upgrade did not deliver it. That is a DECISION, not an
            # unknown — mailu's redis 8.8.0 → 8.10.0 crosses 12 advisories all fixed by 8.6.3 or
            # earlier, and reporting them as "could not judge" overstates the uncertainty.
            if naming and all(_vkey(t) and not _within(kf, kt, _vkey(t)) for t in naming) \
                    and max(_vkey(t) for t in naming if _vkey(t)) <= kf:
                e["classification"] = ("outside-window: fixed in "
                                       f"{', '.join(sorted(naming))}, all at or before {wf}")
                e["fix_versions_from_release_notes"] = sorted(naming)
                already_fixed.add(cve)
                break
    if resolved_by_release:
        report["resolved_by_release_notes"] = resolved_by_release

    indeterminate -= fixed_set | already_fixed
    if already_fixed:
        report["already_fixed_before_upgrade"] = sorted(already_fixed)
    for cve in indeterminate:
        report["cves"][cve]["classification"] = "indeterminate: no fix version published"
    unknown = [c for c in unknown if c not in fixed_set]
    report["fixed_by_this_upgrade"] = sorted(fixed_set)
    report["indeterminate"] = sorted(indeterminate)
    report["cve_count_indeterminate"] = len(indeterminate)
    report["unclassified"] = sorted(unknown)
    # NEVER report 0 for something we could not determine — a 0 asserts safety. If ANY requested
    # window could not be ordered at all, the total is UNKNOWN rather than a partial number.
    report["count_known"] = not unresolved_any
    report["cve_count_fixed"] = len(fixed_set) if not unresolved_any else None
    report["cve_count_total_seen"] = len(report["cves"])
    # Only GENUINE failures make a count unreliable. "no-advisories-published" (404: the repo has
    # no advisory feed) and "skipped: template URL" are benign and must not degrade the verdict.
    report["sources_failed"] = [
        s["source"]
        for s in report["sources"]
        if not (s["status"] == "ok" or s["status"].startswith(("no-advisories-published", "skipped:")))
    ]
    report["sources_benign"] = [
        s["source"]
        for s in report["sources"]
        if s["status"].startswith(("no-advisories-published", "skipped:"))
    ]
    return report


def markdown(rep: dict) -> str:
    """Human/agent-readable block for pasting into the per-recipe upgrade log."""
    L = [f"### Advisory scan (deterministic pre-step) — {rep['recipe']} "
         f"{rep.get('from') or '?'} → {rep.get('to') or '?'}"]
    if not rep.get("count_known", True):
        L.append("\n**CVEs fixed by this upgrade: UNKNOWN — the scan could NOT determine a count.**")
        L.append("\n⚠ This is NOT zero. A version-scheme change (e.g. semver → calver) makes numeric "
                 "ordering meaningless across this jump, so no advisory could be classified. Render "
                 "this recipe's cve cell as `?`, never `0`. Read the vendor's release notes for the "
                 "jump and count by hand.")
        if rep["unclassified"]:
            L.append(f"\nAdvisories seen but unclassifiable ({len(rep['unclassified'])}) — includes "
                     f"other images in this recipe: " + ", ".join(rep["unclassified"][:12]))
        if rep["sources_failed"]:
            L.append(f"\n⚠ sources that FAILED: {', '.join(rep['sources_failed'])}")
        L.append(f"\n_Sources checked: {len(rep['sources'])} ({rep['registry_urls']} registry URLs + "
                 f"advisory APIs). This scan is ADDITIVE — it does not replace the release-note "
                 f"reading in the upgrade step._")
        return "\n".join(L)
    ind = rep.get("cve_count_indeterminate") or 0
    if rep["fixed_by_this_upgrade"]:
        floor = " (at least — see undetermined below)" if ind else ""
        L.append(f"\n**CVEs fixed by this upgrade: {rep['cve_count_fixed']}**{floor}\n")
        cb = rep.get("classified_by") or {}
        if isinstance(cb, dict) and cb:
            for src, method in cb.items():
                dw = (rep.get("date_window") or {}).get(src)
                win = (rep.get("windows") or {}).get(src, {})
                span = f"{win.get('from')} → {win.get('to')}"
                extra = (f" (dates {dw['from'][:10]} → {dw['to'][:10]})" if dw else "")
                L.append(f"_{src.split(':',1)[-1]}: {span} — counted by {method}{extra}._")
            L.append("")
        L.append("| CVE | severity | fixed in | advisory | source |")
        L.append("|---|---|---|---|---|")
        for cve in rep["fixed_by_this_upgrade"]:
            e = rep["cves"][cve]
            L.append(f"| {cve} | {e.get('severity') or '?'} | {e.get('patched') or '?'} | "
                     f"{e.get('ghsa') or '-'} | {e['sources'][0]} |")
    else:
        L.append("\n**CVEs fixed by this upgrade: 0 identified by the deterministic scan.**")
    if ind:
        L.append(f"\n⚠ **{ind} advisory/advisories could NOT be judged** — the vendor published no fix "
                 f"version (GitHub carries `TBD` or a placeholder like `7.4.X`) and the vulnerable "
                 f"range is open-ended, so neither method can tell whether this upgrade fixed them. "
                 f"They are NOT included in the count above and must NOT be read as unaffected: "
                 + ", ".join(f"{c} ({rep['cves'][c].get('severity') or '?'})"
                             for c in rep["indeterminate"][:15])
                 + (" …" if len(rep["indeterminate"]) > 15 else "")
                 + "\n\nRe-run with `--adjudicate` for the collected evidence on each, to judge.")
    if rep["unclassified"]:
        L.append(f"\nSeen but not version-classified ({len(rep['unclassified'])}) — includes advisories "
                 f"from OTHER images in this recipe (sidecars), which this window cannot judge: "
                 + ", ".join(rep["unclassified"][:12]))
    if rep["sources_failed"]:
        L.append(f"\n⚠ sources that FAILED (treat counts as incomplete): {', '.join(rep['sources_failed'])}")
    L.append(f"\n_Sources checked: {len(rep['sources'])} "
             f"({rep['registry_urls']} registry URLs + advisory APIs). This scan is ADDITIVE — it does "
             f"not replace the release-note reading in the upgrade step._")
    return "\n".join(L)


def main() -> int:
    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("recipe")
    ap.add_argument("--from", dest="v_from", default=None)
    ap.add_argument("--to", dest="v_to", default=None)
    ap.add_argument("--json", action="store_true", help="emit raw JSON instead of markdown")
    ap.add_argument("--adjudicate", action="store_true",
                    help="SECOND PASS: for advisories the deterministic pass could not judge (no "
                         "fix version published), fetch their full text + references and append a "
                         "block for the agent to judge. Additive: it never changes the count above.")
    ap.add_argument("--registry", default=REGISTRY_DIR)
    ap.add_argument("--image", action="append", default=[], metavar="NAME=FROM:TO",
                    help="a sidecar image and the versions it moved between, e.g. "
                         "--image redis=7.4:8.10 (repeatable). NAME matches a source repo name; "
                         "its advisories are then counted against ITS OWN versions instead of "
                         "being left unclassified.")
    a = ap.parse_args()
    images = []
    for spec in a.image:
        name, _, rng = spec.partition('=')
        vf, _, vt = rng.partition(':')
        if name and vf and vt:
            images.append((name, vf, vt))
        else:
            print(f'ignoring malformed --image {spec!r} (expected NAME=FROM:TO)', file=sys.stderr)
    rep = scan(a.recipe, a.v_from, a.v_to, a.registry, images)
    if a.adjudicate:
        rep["adjudication"] = [evidence_bundle(rep, c)
                               for c in needs_judgement(rep)[:MAX_ADJUDICATE]]
    if a.json:
        print(json.dumps(rep, indent=2))
    else:
        print(markdown(rep))
        if a.adjudicate:
            print(adjudication_block(rep))
    return 0


if __name__ == "__main__":
    sys.exit(main())