luckiest-co 1.0.62 → 1.0.64

package/skills/copywriting-humanize/scripts/ste-lint.py

@@ -0,0 +1,2388 @@
+#!/usr/bin/env python3
+"""Deterministic linter for the structural STE rules in SKILL.md.
+
+Checks only rules verifiable without ASD's dictionary. Deliberately never
+flags hedges or modality (may/might/could): the skill treats confidence as
+content, and a linter that pressures hedges out would rewrite claims.
+
+Usage:
+    ste-lint.py FILE [FILE ...]
+    echo "text" | ste-lint.py [--json]
+    ste-lint.py --baseline 5 FILE      # pass unless hard violations exceed 5
+    ste-lint.py --disable passive-voice,present-perfect FILE
+    ste-lint.py --selftest
+
+Exit 1 when hard ("advisory-free") violations exceed the baseline (default 0).
+Advisory findings (passive voice, compound tenses) never fail the run.
+Exit 2 on a usage error (unknown flag, bad value, unknown rule, unreadable file).
+"""
+import argparse
+import bisect
+import contextlib
+import html.entities
+import io
+import json
+import os
+import re
+import sys
+import tempfile
+
+# ponytail: regex heuristics, not a parser. No noun-cluster rule — needs POS
+# tagging to avoid constant false positives; add spaCy-backed rule if ever needed.
+# No ellipsis rule by owner's choice: technical writing sometimes earns one.
+# Irregular past participles used by the present-perfect heuristic.
+IRREGULAR_PARTICIPLES = "given|taken|made|done|found|seen|known|shown|written|built|sent|set|run|read|kept|held|left|put|cut|hit|let|shut|split|spread|begun|become|come|gone|got|gotten|lost|met|paid|said|sold|told|thought|brought|bought|caught|taught|won|worn|torn|born|drawn|grown|thrown|flown|driven|risen|chosen|broken|spoken|frozen|hidden|ridden|forgotten|fallen|eaten|beaten|understood|stood|struck|stuck|swung|hung|led|fed|bled|fled|sped|bound|wound|dug|spun|slid|bit|lit|quit"
+
+# Preserve the original passive heuristic; perfects also include intransitive verbs.
+PASSIVE_PARTICIPLES = "given|taken|made|done|found|seen|known|shown|written|built|sent|set|run|read|kept|held|left|put"
+# Up to two words may stand between the modal and "have": "may not",
+# "might already", "could not possibly", "may very well".
+MODAL_PERFECT_PREFIX = re.compile(
+    r"\b(?:(?:may|might|could|should|would|must|can|will|shall)(?:n['’]t)?"
+    r"|cannot|can['’]t|won['’]t|shan['’]t)"
+    r"(?:\s+(?:not|already|also|ever|even|just|never|still|very|well|\w+ly)){0,2}\s+$",
+    re.I,
+)
+
+RULES = [
+    ("semicolon", "advisory-free",
+     re.compile(r";"),
+     "STE bans the semicolon (Rule 8.1). Split into separate sentences."),
+    ("phrasal-verb", "advisory-free",
+     re.compile(r"\b(spin(?:ning|s)? up|spun up|reach(?:ing|es|ed)? out|div(?:e|es|ing|ed) into|dove into|kick(?:ing|s|ed)? off|circl(?:e|es|ing|ed) back|touch(?:ing|es|ed)? base)\b", re.I),
+     "Soft phrasal verb. Use the single plain verb (start, contact, read, begin)."),
+    ("marketing-adjective", "advisory-free",
+     re.compile(r"\b(seamless(?:ly)?|robust(?:ly)?|cutting-edge|effortless(?:ly)?|blazing[- ]fast|world-class|state-of-the-art|game-chang(?:ing|er))\b", re.I),
+     "Marketing adjective. Delete, or replace with the measurement that earns the claim."),
+    ("nominalization", "advisory-free",
+     re.compile(r"\b(perform|performs|performed|conduct|conducts|conducted|carry out|carries out|carried out)\s+(?:a|an|the)\s+\w+(?:tion|sion|ment|ance|ence|ysis)\b", re.I),
+     "Action frozen into a noun. Use the verb (analyze, not perform an analysis of)."),
+    ("passive-voice", "advisory",
+     re.compile(r"\b(is|are|was|were|been|being)\s+(\w+ed|" + PASSIVE_PARTICIPLES + r")\b(?!\s+(?:to|for|by)\s+\w+ing)", re.I),
+     "Possible passive voice. Name the actor and use an active verb, unless the actor is unknown or irrelevant."),
+    ("present-perfect", "advisory",
+     # modal + perfect infinitive ("may have failed") is a protected hedge, not present perfect
+     re.compile(r"(?<!\bmay )(?<!\bmight )(?<!\bcould )(?<!\bshould )(?<!\bwould )(?<!\bmust )\b(has|have|had)\s+(?:been\s+)?(?:\w+(?:ed|en)|" + IRREGULAR_PARTICIPLES + r")\b", re.I),
+     "Compound tense. Use simple past/present unless current relevance is the point (then keep and flag)."),
+]
+
+# One word, one meaning: groups of verbs commonly rotated for the same action.
+# Only pairs where the members are genuinely interchangeable — error/fault/failure
+# are distinct concepts and stay out.
+SYNONYM_GROUPS = [
+    ("check", "verify", "confirm", "validate"),
+    ("delete", "remove", "erase"),
+    ("start", "launch", "begin", "initiate"),
+    ("stop", "halt", "terminate"),
+    ("show", "display"),
+    ("use", "utilize", "employ"),
+    ("fix", "repair", "correct"),
+    ("send", "transmit"),
+    ("get", "retrieve", "fetch", "obtain"),
+    ("change", "modify", "alter"),
+]
+
+MAX_WORDS = 25  # descriptions cap; instructions cap is 20 but undetectable without context
+
+CODE_FENCE = re.compile(r"^(`{3,}|~{3,})")
+INLINE_CODE = re.compile(r"`[^`]*`")
+BACKTICK_RUN = re.compile(r"`+")
+# A list item start at any indent, to find the content indent of its item.
+LIST_ITEM_INDENT = re.compile(r"^( *)([-*+]|[0-9]{1,9}[.)])( +|$)")
+# Markup that is not prose: character references (&nbsp; &#39; &#x2192;) and
+# autolinks (<https://...>). Inline link destinations ("](url") are found by
+# _link_destination_end. A reference is markup only when GitHub reads it as one.
+CHAR_REF = re.compile(r"&(?:([A-Za-z][A-Za-z0-9]{1,31})|#([0-9]{1,7})|#[xX]([0-9A-Fa-f]{1,6}));")
+LINK_OPEN = re.compile(r"\]\(")
+BRACKET = re.compile(r"[\[\]]")
+# A list item or a fence that can interrupt a paragraph: a bullet or "1." with text.
+PARAGRAPH_INTERRUPT = re.compile(r" {0,3}(?:[-+*][ \t]+\S|1[.)][ \t]+\S|`{3,}|~{3,})")
+AUTOLINK = re.compile(r"<[A-Za-z][A-Za-z0-9+.-]{1,31}:[^<>\s]*>")
+# An inline HTML open or closing tag on one line (CommonMark 6.6).
+INLINE_TAG = re.compile(
+    r"<(?:(?P<name>[A-Za-z][A-Za-z0-9-]*)(?:[ \t]+[A-Za-z_:][A-Za-z0-9_.:-]*"
+    r"(?:[ \t]*=[ \t]*(?:[^\s\"'=<>`]+|'[^']*'|\"[^\"]*\"))?)*[ \t]*/?"
+    r"|/[A-Za-z][A-Za-z0-9-]*[ \t]*)>"
+)
+# Other inline HTML (CommonMark 6.6): a comment, a processing instruction, a
+# declaration, a CDATA section. Each pair is (opener, closer).
+INLINE_HTML_OTHER = ((re.compile(r"<!--"), "-->"), (re.compile(r"<\?"), "?>"),
+                     (re.compile(r"<![A-Za-z]"), ">"), (re.compile(r"<!\[CDATA\["), "]]>"))
+# GitHub shows these tags as text (GFM 6.11, the tagfilter extension).
+FILTERED_TAGS = {"title", "textarea", "style", "xmp", "iframe", "noembed", "noframes", "script", "plaintext"}
+ATTRIBUTE_VALUE = re.compile(r"=[ \t]*(?:\"[^\"]*\"|'[^']*'|[^\s\"'=<>`]+)")
+# CommonMark 6.3 lets an implementation limit the nesting of parentheses in a
+# link destination. The limit keeps the scan of a long line linear.
+MAX_LINK_PARENS = 32
+ASCII_PUNCTUATION = "!\"#$%&'()*+,-./:;<=>?@[\\]^_`{|}~"
+LIST_ITEM_START = re.compile(
+    r"^(?P<indent> {0,3})(?P<marker>[-*+]|[0-9]+[.)])(?P<gap> +)(?P<body>.*)$"
+)
+CONJUNCTION_END = re.compile(r"\b(?:and|or)\s*$", re.I)
+TABLE_SEPARATOR_CELL = re.compile(r"^:?-+:?$")
+# Headings, thematic breaks, front-matter delimiters, and HTML lines stand alone.
+STANDALONE_LINE = re.compile(r"^ {0,3}(?:#|<|(?:[-*_=] *){3,}$)")
+BLOCKQUOTE_PREFIX = re.compile(r"^ {0,3}(?:> ?)+")
+# A GitHub alert marker, alone on the first line of a quote: "> [!NOTE]".
+GITHUB_ALERT = re.compile(r"\[!(?:note|tip|important|warning|caution)\]", re.I)
+# MkDocs admonition title lines. Four spaces allow a title inside a list item.
+# Only ??? admonitions support the collapsible "+" marker.
+ADMONITION_TITLE = re.compile(
+    r'^ {0,4}(?:!!!|\?\?\?\+?)[ \t]*[\w-]+'
+    r'(?:[ \t]+[\w-]+)*(?:[ \t]+".*")?[ \t]*$'
+)
+# Track ordinary block containers, not inline or void tags such as span or hr.
+HTML_BLOCK_TAG = re.compile(
+    r"<(/?)(address|article|aside|blockquote|body|caption|center|dd|details|"
+    r"dialog|dir|div|dl|dt|fieldset|figcaption|figure|footer|form|h[1-6]|"
+    r"header|hgroup|html|iframe|legend|li|main|menu|nav|noframes|ol|"
+    r"optgroup|option|p|pre|script|search|section|style|summary|table|"
+    r"tbody|td|textarea|tfoot|th|thead|title|tr|ul)(?=[\s/>])[^>]*>", re.I,
+)
+# Any list marker at any depth starts a new prose block.
+ANY_LIST_ITEM = re.compile(r"^\s*(?:[-*+]|[0-9]+[.)])(?:\s|$)")
+# A letter or a digit: a token without one is punctuation, not a word.
+WORD_CHAR = re.compile(r"[^\W_]")
+# A token that ends in terminal punctuation, optionally followed by closing
+# quotes, brackets, or Markdown emphasis markers ("**Lead.**").
+SENTENCE_END = re.compile(r"[.!?][\"'”’)\]*_]*$")
+OPENERS = "\"'“‘([*_"
+# Abbreviations whose period never ends a sentence.
+ABBREVIATIONS = {"vs.", "cf.", "approx.", "mr.", "mrs.", "ms.", "dr.", "fig.", "no."}
+# Dotted initialisms such as "e.g." and "U.S." end a sentence only when the
+# next word starts with a capital letter.
+DOTTED_ABBREVIATION = re.compile(r"(?:[a-z]\.){2,}", re.I)
+
+
+def _word_re(base):
+    return re.compile(r"\b" + base + r"(?:s|es|ed|d|ing)?\b", re.I)
+
+
+def _count_words(text):
+    """Count whitespace-separated tokens that contain a letter or a digit.
+
+    Bare punctuation is not a word: a comma left behind by a masked code
+    span, or a spaced em dash, does not make a sentence longer.
+    """
+    return sum(1 for token in text.split() if WORD_CHAR.search(token))
+
+
+def _mask_inline_code(text):
+    """Blank code spans with equal-length spaces so columns stay true."""
+    return INLINE_CODE.sub(lambda match: " " * len(match.group(0)), text)
+
+
+def _list_item(rest, list_indent, in_paragraph):
+    """Return the match of the list item that starts REST, or None (CommonMark 5.2).
+
+    REST is a line after its quote prefix. A thematic break such as "* * *" is no
+    item. IN_PARAGRAPH: the line would otherwise continue a paragraph in the same
+    quote. Then an item to the left of LIST_INDENT is the next item of the open
+    list, and a new list starts only with a bullet or "1" and some text, and not
+    as indented code.
+    """
+    if NO_PARAGRAPH.match(rest):
+        return None
+    item = LIST_ITEM_INDENT.match(rest)
+    if not item or not in_paragraph or len(item.group(1)) < list_indent:
+        return item
+    marker = item.group(2)
+    if (len(item.group(1)) < list_indent + 4 and rest[item.end():].strip()
+            and (marker in "-*+" or int(marker[:-1]) == 1)):
+        return item
+    return None
+
+
+def _breaks_paragraph(depth, line, list_indent):
+    """True when LINE ends a paragraph at quote DEPTH, as the main loop ends it.
+
+    These lines end it (CommonMark 4-5): a blank or quote-only line, a deeper quote,
+    a heading, a thematic break, a setext underline in the open item, a fence that is
+    no indented code, an HTML block of types 1 to 6, and a list item that can
+    interrupt a paragraph. A text line with fewer ">" is a lazy line and continues
+    the paragraph, but any list item there ends it. A code span ends there too.
+    """
+    line_depth, rest = _strip_quotes(line.expandtabs(4))
+    in_item = _indent_width(rest) >= list_indent
+    return (not rest.strip() or line_depth > depth
+            or bool(NO_PARAGRAPH.match(rest)
+                    or (_fence_step(None, line)[0] and _indent_width(rest) < list_indent + 4)
+                    or (line_depth == depth and in_item and SETEXT_UNDERLINE.match(rest))
+                    or any(start.match(rest) for start, _ in HTML_BLOCK_KINDS)
+                    or _list_item(rest, list_indent, line_depth == depth)))
+
+
+def _paragraph_runs(lines, start, depth, list_indent):
+    """Return (end, last) for the paragraph that continues at LINES[START].
+
+    END is the index of the line that ends it. LAST maps each backtick run length
+    to the index of its last line before END. One scan serves every run in the
+    paragraph, so the cost stays linear in the line count.
+    """
+    last, end = {}, start
+    while end < len(lines) and not _breaks_paragraph(depth, lines[end], list_indent):
+        for run in BACKTICK_RUN.finditer(lines[end]):
+            last[len(run.group())] = end
+        end += 1
+    return end, last
+
+
+def _mask_code_spans(line, open_run, closes_later, lines=(), next_index=0, depth=0, skip=0, carry=None,
+                     brackets=None):
+    """Blank the code spans in LINE, also a span that wraps across lines.
+
+    OPEN_RUN is the backtick run length of a span still open from an earlier
+    line, or 0. LINES[NEXT_INDEX:] are the lines after LINE. As in CommonMark 6.1, a
+    span ends at the next run of the same length, and a run that no later run
+    in the paragraph closes is literal text. Outside code spans, fill inline
+    link destinations with underscores, blank link titles, and fill the
+    inline HTML that GitHub does not show. What starts first wins: a code
+    span that opens before "](" hides the link, and a backtick in a link, an
+    autolink or inline HTML is text.
+    DEPTH is the quote depth of the paragraph that LINE is in, or None when no
+    paragraph goes on after LINE, so a link must close in it. SKIP is where a
+    link from an earlier line ends in LINE. When a link goes on in a later
+    line, CARRY gets [lines ahead, column after its ")"]. BRACKETS carries the
+    open "[" marks from line to line of a paragraph. Return (masked line, open
+    run).
+    """
+    masked = list(line)
+    start = 0 if open_run else None
+    runs = list(BACKTICK_RUN.finditer(line))
+    last = {len(run.group()): i for i, run in enumerate(runs)}  # last run of each length
+    opens = [m.end() for m in LINK_OPEN.finditer(line) if not _escaped(line, m.start())]
+    # Autolinks and inline HTML tags: a code span or a link in one is text.
+    # An item is (start, end, the parts of it that GitHub does not show).
+    atoms = [(m.start(), m.end(), ()) for m in AUTOLINK.finditer(line)]
+    for m in INLINE_TAG.finditer(line):
+        # GitHub shows no attribute value. A bare word stays, as in the
+        # placeholder "<one of the two>".
+        hide = ()
+        if m.group("name") and m.group("name").lower() not in FILTERED_TAGS:
+            hide = [value.span() for value in ATTRIBUTE_VALUE.finditer(line, m.end("name"), m.end())]
+        atoms.append((m.start(), m.end(), hide))
+    atoms += _inline_html_other(line)
+    atoms.sort(key=lambda atom: atom[:2])
+    atoms = [atom for atom in atoms if not _escaped(line, atom[0])]
+    marks = [m.start() for m in BRACKET.finditer(line) if not _escaped(line, m.start())]
+    if brackets is None:
+        brackets = [[], 0]
+    # BRACKETS[0] holds the "[" marks of the paragraph that no "]" closed yet,
+    # True for an image "![". A "[" below BRACKETS[1] is in a link, so it
+    # opens no link (CommonMark 6.3); an image "![" can still open an image.
+    stack = brackets[0]
+    a = 0
+    b = 0
+    k = 0
+    done = skip  # the end of the last code span or link
+    _blank_title(masked, line, 0, skip)  # a link title from an earlier line
+
+    def pop_bracket():
+        image = stack.pop()
+        active = image or len(stack) >= brackets[1]
+        brackets[1] = min(brackets[1], len(stack))
+        return image, active
+
+    def take_brackets(limit):
+        # Read the "[" and "]" marks before LIMIT, outside code, links and tags.
+        nonlocal b
+        while b < len(marks) and marks[b] < limit:
+            pos = marks[b]
+            b += 1
+            if pos < done:
+                continue
+            if line[pos] == "[":
+                stack.append(pos > 0 and line[pos - 1] == "!" and not _escaped(line, pos - 1))
+            elif line[pos + 1:pos + 2] != "(" and stack:
+                pop_bracket()  # mask_links reads a "](" mark
+
+    def take_atoms(limit):
+        # Let the autolinks and tags that start before LIMIT start first.
+        nonlocal a, done
+        while a < len(atoms) and atoms[a][0] < limit:
+            begin, end, hide = atoms[a]
+            a += 1
+            if begin < done:
+                continue
+            take_brackets(begin)
+            done = end
+            # Keep the spaces, so the token count stays.
+            for left, right in hide:
+                masked[left:right] = [c if c.isspace() else "_" for c in line[left:right]]
+
+    def mask_links(limit):
+        # Fill the destinations after the "](" marks before LIMIT.
+        nonlocal k, done
+        while k < len(opens) and opens[k] <= limit:
+            pos = opens[k]
+            k += 1
+            take_atoms(pos - 2)
+            take_brackets(pos - 2)
+            if pos - 2 < done or not stack:
+                continue  # no "[" opens link text, so "](" is text
+            image, active = pop_bracket()
+            if not active:
+                continue
+            later = (lines[index] for index in range(next_index, len(lines)))
+            link = _link_destination_end(line, pos, later, depth)
+            if link is None or (link[0] == pos and line[pos:link[1]].strip() != ")"):
+                continue  # with no destination, a "title" is the destination: a gap this reads as no link
+            if not image:
+                brackets[1] = len(stack)  # a link holds no link
+            masked[pos:link[0]] = "_" * (link[0] - pos)
+            _blank_title(masked, line, link[0], link[1])
+            done = link[1]
+            if carry is not None and link[2]:
+                carry[:] = link[2]
+
+    for i, run in enumerate(runs):
+        n = len(run.group())
+        if start is not None:
+            if n == open_run:
+                masked[start:run.end()] = " " * (run.end() - start)
+                start, open_run = None, 0
+                done = run.end()
+            continue
+        mask_links(run.start())
+        take_atoms(run.start())
+        if run.start() < done:
+            continue  # a backtick in a link, an autolink or a tag is text
+        first = run.start()
+        if _escaped(line, first):
+            first, n = first + 1, n - 1  # an escaped backtick is text
+        if n and (last.get(n, -1) > i or closes_later(n)):
+            take_brackets(first)
+            start, open_run = first, n
+    if start is not None:
+        masked[start:] = " " * (len(line) - start)
+    else:
+        mask_links(len(line))
+        take_atoms(len(line))
+        take_brackets(len(line))  # a "[" that a later line can close
+    return "".join(masked), open_run
+
+
+def _inline_html_other(line):
+    """Return (start, end, hidden parts) for each comment, processing
+    instruction, declaration and CDATA section that starts and ends in LINE.
+
+    GitHub shows none of a comment or a declaration. Its HTML sanitizer reads
+    a processing instruction and CDATA as a comment that ends at the first ">",
+    so it shows the text after that ">".
+    """
+    found = []
+    for opener, closer in INLINE_HTML_OTHER:
+        for m in opener.finditer(line):
+            if closer == "-->" and line.startswith(("<!-->", "<!--->"), m.start()):
+                end = line.index(">", m.end()) + 1
+                found.append((m.start(), end, [(m.start(), end)]))
+                continue
+            end = line.find(closer, m.end())
+            if end < 0:
+                break  # no later opener closes either
+            end += len(closer)
+            shown = end if closer in ("-->", ">") else line.index(">", m.start()) + 1
+            found.append((m.start(), end, [(m.start(), shown)]))
+    return found
+
+
+def _escaped(line, pos):
+    """True when an odd number of backslashes comes right before LINE[POS]."""
+    start = pos
+    while start and line[start - 1] == "\\":
+        start -= 1
+    return (pos - start) % 2 == 1
+
+
+def _blank_title(masked, line, start, end):
+    """Blank the link title in MASKED[START:END] with spaces, but keep its ")".
+
+    GitHub shows a title in no text, so it is not prose. The ")" stays, so the
+    link stays one token for the word count, as it is when it has no title.
+    """
+    if start < end:
+        masked[start:end] = " " * (end - start)
+        if line[end - 1] == ")":
+            masked[end - 1] = ")"
+
+
+def _char_ref_is_markup(match):
+    """True when the character reference MATCH renders as one character.
+
+    A named reference counts only when its name is in the HTML5 list. A
+    reference that renders as a semicolon is prose: GitHub shows a semicolon.
+    """
+    name, decimal, hexadecimal = match.groups()
+    if name is not None:
+        return name + ";" in html.entities.html5 and name != "semi"
+    value = int(decimal) if decimal is not None else int(hexadecimal, 16)
+    return value != ord(";")
+
+
+def _paragraph_lines(later_lines, depth):
+    """Yield (text, column, lines ahead) for the lines that continue the paragraph.
+
+    TEXT is the line without its quote prefix, which ends at COLUMN.
+
+    DEPTH is the quote depth of the paragraph, or None when no paragraph goes
+    on. A blank line, a deeper quote and a line that starts a block end the
+    paragraph (CommonMark 4.1 to 4.6, 5.1 and 5.2). A line with fewer ">" is a
+    lazy line and continues it.
+    """
+    if depth is None:
+        return
+    for ahead, line in enumerate(later_lines, 1):
+        levels, rest = _strip_quotes(line)
+        if (levels > depth or not rest.strip() or NO_PARAGRAPH.match(rest)
+                or SETEXT_UNDERLINE.match(rest) or PARAGRAPH_INTERRUPT.match(rest)
+                or any(start.match(rest) for start, _ in HTML_BLOCK_KINDS)):
+            return
+        yield rest, len(line) - len(rest), ahead
+
+
+def _skip_link_gap(item, pos, rest):
+    """Skip spaces and tabs from POS, with at most one line ending.
+
+    ITEM is (text, column, lines ahead), and REST yields the next items of the
+    paragraph. Return (item, pos, crossed), or None when the paragraph ends.
+    """
+    while pos < len(item[0]) and item[0][pos] in " \t":
+        pos += 1
+    if pos < len(item[0]):
+        return item, pos, False
+    item = next(rest, None)
+    if item is None:
+        return None
+    pos = 0
+    while pos < len(item[0]) and item[0][pos] in " \t":
+        pos += 1
+    return item, pos, True
+
+
+def _skip_link_title(item, pos, rest):
+    """Return (item, pos) after a link title that starts at POS, or None.
+
+    ITEM is (text, column, lines ahead). The title can go on in the next lines
+    of the paragraph (REST), but it must close before the paragraph ends.
+    """
+    line = item[0]
+    closer = {'"': '"', "'": "'", "(": ")"}[line[pos]]
+    pos += 1
+    while True:
+        while pos < len(line):
+            if line[pos] == "\\" and pos + 1 < len(line) and line[pos + 1] in ASCII_PUNCTUATION:
+                pos += 2
+                continue
+            if line[pos] == closer:
+                return item, pos + 1
+            if closer == ")" and line[pos] == "(":
+                return None
+            pos += 1
+        item, pos = next(rest, None), 0
+        if item is None:
+            return None
+        line = item[0]
+
+
+def _link_destination_end(text, start, later_lines=(), depth=None):
+    """Return (destination end, link end, carry) for the destination at START, or None.
+
+    START is after "](". CommonMark 6.3: the destination is in <...>, or it has
+    no space or control character and its parentheses are balanced. An
+    optional title and ")" must follow, in the same paragraph. LATER_LINES are
+    the lines after TEXT, and DEPTH is the quote depth of the paragraph (see
+    _paragraph_lines). Each gap of spaces and tabs may hold one line ending,
+    and the title may span lines. The link end is after ")" when the link
+    closes in TEXT, else the end of TEXT, and CARRY is then (lines ahead,
+    column after ")") of the line where it closes, else None.
+    """
+    pos = start
+    if pos < len(text) and text[pos] == "<":
+        pos += 1
+        while pos < len(text) and text[pos] not in "<>":
+            pos += 2 if text[pos] == "\\" and pos + 1 < len(text) else 1
+        if pos >= len(text) or text[pos] != ">":
+            return None
+        end = pos + 1
+    else:
+        parens = 0
+        while pos < len(text):
+            char = text[pos]
+            if char == "\\" and pos + 1 < len(text) and text[pos + 1] in ASCII_PUNCTUATION:
+                pos += 2
+                continue
+            if char <= " " or char == "\x7f":
+                break
+            if char == "(":
+                parens += 1
+                if parens > MAX_LINK_PARENS:
+                    return None
+            elif char == ")":
+                if parens == 0:
+                    break
+                parens -= 1
+            pos += 1
+        if parens:
+            return None
+        end = pos
+    rest = _paragraph_lines(later_lines, depth)
+    gap = _skip_link_gap((text, 0, 0), end, rest)
+    if gap is None:
+        return None
+    item, pos, crossed = gap
+    if (crossed or pos > end) and item[0][pos] in "\"'(":
+        title = _skip_link_title(item, pos, rest)
+        if title is None:
+            return None
+        gap = _skip_link_gap(*title, rest)
+        if gap is None:
+            return None
+        item, pos, _ = gap
+    if item[0][pos] != ")":
+        return None
+    if item[2] == 0:
+        return end, pos + 1, None
+    return end, len(text), (item[2], item[1] + pos + 1)
+
+
+def _mask_non_prose(text):
+    """Fill character references and autolinks.
+
+    The fill is underscores, not spaces, so a token stays one token and word
+    counts do not change. An underscore is not a word character for _count_words.
+    Code spans and link destinations are blanked by _mask_code_spans.
+    """
+    out = []
+    pos = done = 0
+    while pos < len(text):
+        end = None
+        if text[pos] == "&":
+            ref = CHAR_REF.match(text, pos)
+            if ref and _char_ref_is_markup(ref):
+                end = ref.end()
+        if end is None and text[pos] == "<":
+            link = AUTOLINK.match(text, pos)
+            if link:
+                end = link.end()
+        if end is None:
+            pos += 1
+            continue
+        out.append(text[done:pos])
+        out.append("_" * (end - pos))
+        pos = done = end
+    out.append(text[done:])
+    return "".join(out)
+
+
+def _sentence_starts(text):
+    """Return the offset of each sentence start in text.
+
+    A sentence ends at a word that ends in terminal punctuation, after any
+    closing quotes or brackets. Known abbreviations do not end a sentence.
+    One pass over the words keeps the cost linear in the text length.
+    """
+    words = list(re.finditer(r"\S+", text))
+    if not words:
+        return [0]
+    starts = [words[0].start()]
+    for word, next_word in zip(words, words[1:]):
+        end = SENTENCE_END.search(word.group())
+        if not end:
+            continue
+        core = word.group()[:end.start() + 1].lstrip(OPENERS)
+        if core.lower() in ABBREVIATIONS:
+            continue
+        if (DOTTED_ABBREVIATION.fullmatch(core)
+                and not next_word.group().lstrip(OPENERS)[:1].isupper()):
+            continue
+        starts.append(next_word.start())
+    return starts
+
+
+def _long_sentence_findings(pieces, filename):
+    """Flag long sentences in one prose block.
+
+    pieces is a list of (lineno, source_column, text) tuples. The texts are
+    joined with spaces, so a sentence wrapped across source lines is counted
+    once, and each finding points at the line and column where it starts.
+    """
+    offsets, position = [], 0
+    for _, _, text in pieces:
+        offsets.append(position)
+        position += len(text) + 1
+    joined = " ".join(text for _, _, text in pieces)
+    findings = []
+    starts = _sentence_starts(joined)
+    for start, end in zip(starts, starts[1:] + [len(joined)]):
+        n = _count_words(joined[start:end])
+        if n <= MAX_WORDS:
+            continue
+        index = bisect.bisect_right(offsets, start) - 1
+        offset = offsets[index]
+        lineno, source_column, _ = pieces[index]
+        findings.append({"file": filename, "line": lineno,
+                         "col": source_column + start - offset + 1,
+                         "rule": "long-sentence", "level": "advisory-free",
+                         "match": f"{n} words",
+                         "message": f"Sentence has {n} words (cap {MAX_WORDS}). Split it."})
+    return findings
+
+
+# One blockquote level: up to three spaces, ">", and one optional space.
+QUOTE_LEVEL = re.compile(r" {0,3}> ?")
+
+
+def _strip_quotes(line, depth=None):
+    """Return (levels, rest): the blockquote levels that start LINE and the text after them.
+
+    With DEPTH, remove at most DEPTH levels: a deeper ">" is content.
+    """
+    levels, pos = 0, 0
+    while depth is None or levels < depth:
+        level = QUOTE_LEVEL.match(line, pos)
+        if not level:
+            break
+        levels, pos = levels + 1, level.end()
+    return levels, line[pos:]
+
+
+def _indent_width(text):
+    """Columns of the spaces and tabs that start TEXT, with a tab stop of 4."""
+    return len(text[:len(text) - len(text.lstrip(" \t"))].expandtabs(4))
+
+
+def _fence_ends_with_quote(fence, line):
+    """True when LINE has fewer blockquote levels than the open FENCE."""
+    return fence is not None and _strip_quotes(line.expandtabs(4), fence[1])[0] < fence[1]
+
+
+def _fence_step(fence, line):
+    """Return (is_fence_line, open_fence) after LINE.
+
+    FENCE is None, or the open fence as (run, quote depth, indent). This follows
+    CommonMark 4.5 for blockquotes and for the fence itself: a fence in a blockquote
+    ends when its quote ends. Only a run of the same character, at least as long,
+    with only spaces or tabs after it, closes the block. A backtick fence has no
+    backtick in its info string. List items are not tracked, so the indent of a
+    fence is not limited to three columns. A closing fence may have at most three
+    columns more indent than the opening fence, also at the top level.
+    """
+    # Tabs become spaces from the real column, so a tab after ">" counts correctly.
+    line = line.expandtabs(4)
+    if fence is None:
+        depth, rest = _strip_quotes(line)
+        content = rest.lstrip(" \t")
+        marker = CODE_FENCE.match(content)
+        if not marker:
+            return False, None
+        run = marker.group(1)
+        if run[0] == "`" and "`" in content[marker.end():]:
+            return False, None
+        return True, (run, depth, _indent_width(rest))
+    if _fence_ends_with_quote(fence, line):
+        # The quote that holds the fence ended, so the fence ended too.
+        return _fence_step(None, line)
+    run, depth, indent = fence
+    rest = _strip_quotes(line, depth)[1]
+    content = rest.lstrip(" \t")
+    marker = CODE_FENCE.match(content)
+    if (marker and marker.group(1)[0] == run[0] and len(marker.group(1)) >= len(run)
+            and _indent_width(rest) <= indent + 3
+            and not content[marker.end():].strip(" \t")):
+        return True, None
+    return False, fence
+
+
+# CommonMark 4.6: how each kind of HTML block starts, and the text that ends it.
+# None: the block ends at a blank line (types 6 and 7).
+_BLOCK_TAGS = (
+    "address|article|aside|base|basefont|blockquote|body|caption|center|col|colgroup|dd|"
+    "details|dialog|dir|div|dl|dt|fieldset|figcaption|figure|footer|form|frame|frameset|"
+    "h[1-6]|head|header|hr|html|iframe|legend|li|link|main|menu|menuitem|nav|noframes|ol|"
+    "optgroup|option|p|param|search|section|summary|table|tbody|td|tfoot|th|thead|title|"
+    "tr|track|ul"
+)
+HTML_BLOCK_KINDS = [
+    (re.compile(r"^ {0,3}<(?:script|pre|style|textarea)(?:[ \t]|>|$)", re.I),
+     re.compile(r"</(?:script|pre|style|textarea)>", re.I)),
+    (re.compile(r"^ {0,3}<!--"), re.compile(r"-->")),
+    (re.compile(r"^ {0,3}<\?"), re.compile(r"\?>")),
+    (re.compile(r"^ {0,3}<!\[CDATA\["), re.compile(r"\]\]>")),
+    (re.compile(r"^ {0,3}<![A-Za-z]"), re.compile(r">")),
+    (re.compile(rf"^ {{0,3}}</?(?:{_BLOCK_TAGS})(?:[ \t]|/?>|$)", re.I), None),
+]
+HTML_COMMENT_END = HTML_BLOCK_KINDS[1][1]
+# Type 7: one complete open or closing tag alone on its line (CommonMark 6.6 grammar).
+# It cannot interrupt a paragraph. In one line, CommonMark whitespace is a space or a
+# tab: a no-break space or another Unicode space is text, so the line is no tag.
+_ATTRIBUTE = (r"[ \t]+[A-Za-z_:][A-Za-z0-9_.:-]*"
+              r"(?:[ \t]*=[ \t]*(?:[^ \t\"'=<>`]+|'[^']*'|\"[^\"]*\"))?")
+# The names of type 1 (pre, script, style, textarea) never start a type 7 block,
+# as an open tag or as a closing tag. A custom name such as pre-x is no type 1 name.
+_TYPE1_NAME = r"(?!(?i:pre|script|style|textarea)(?![A-Za-z0-9-]))"
+HTML_TYPE7 = re.compile(
+    rf"^ {{0,3}}(?:<{_TYPE1_NAME}[A-Za-z][A-Za-z0-9-]*(?:{_ATTRIBUTE})*[ \t]*/?>"
+    rf"|</{_TYPE1_NAME}[A-Za-z][A-Za-z0-9-]*[ \t]*>)[ \t]*$")
+# Lines that end a paragraph and are no paragraph line, after the quote prefix: an ATX
+# heading and a thematic break of one repeated character. A quote line with no text
+# also ends a paragraph.
+NO_PARAGRAPH = re.compile(
+    r"^ {0,3}(?:#{1,6}(?:[ \t]|$)|(?:-[ \t]*){3,}$|(?:\*[ \t]*){3,}$|(?:_[ \t]*){3,}$)")
+# A setext underline. It ends a paragraph only when a paragraph is open at the same
+# quote depth. Else it is text, also as a lazy line after a quote (CommonMark 4.3).
+SETEXT_UNDERLINE = re.compile(r"^ {0,3}(?:=+|-+)[ \t]*$")
+
+
+def _html_block_step(block, line, in_paragraph):
+    """Return (in_html_block, open_block) for LINE.
+
+    BLOCK is None, False (an HTML block that ends at a blank line), or the pattern
+    that ends the open block. IN_PARAGRAPH: a paragraph is open before LINE. A ">"
+    in an HTML block is text, so it opens no fence.
+    """
+    if block is None:
+        for start, end in HTML_BLOCK_KINDS:
+            if start.match(line):
+                block = False if end is None else end
+                break
+        else:
+            if in_paragraph or not HTML_TYPE7.match(line):
+                return False, None
+            block = False
+    if block is False:
+        return bool(line.strip()), (block if line.strip() else None)
+    return True, (None if block.search(line) else block)
+
+
+# GitHub shows the text of an HTML block of type 6 or 7. It also shows the text of a
+# <script>, <style> or <textarea> block, because it escapes these tags (tagfilter). It
+# shows a <pre> block as code, and it hides a comment and a "<?", "<!X" or CDATA block.
+SHOWN_TYPE1 = re.compile(r"^ {0,3}<(?i:script|style|textarea)(?:[ \t]|>|$)")
+
+
+def _block_shows_text(block, line):
+    """True when GitHub shows the text of the HTML BLOCK that starts on LINE as prose."""
+    return block is False or bool(SHOWN_TYPE1.match(line))
+
+
+def _is_list_continuation(line, content_indent):
+    if not line.strip():
+        return True
+    if LIST_ITEM_START.match(line):
+        return False
+    return len(line[:len(line) - len(line.lstrip(" \t"))].expandtabs(4)) >= content_indent
+
+
+def _split_table_row(line):
+    """Return trimmed table cells and their zero-based source columns.
+
+    A row needs a leading or trailing pipe, or a pipe between two cells.
+    Escaped pipes stay in their cell. This deliberately implements only the
+    ordinary Markdown table shape; it is enough to distinguish a table from
+    prose that happens to contain a pipe.
+    """
+    left = len(line) - len(line.lstrip())
+    right = len(line.rstrip())
+    content = line[left:right]
+    if "|" not in content:
+        return None
+    edge_pipe = False
+    if content.startswith("|"):
+        content = content[1:]
+        left += 1
+        edge_pipe = True
+    if content.endswith("|"):
+        content = content[:-1]
+        edge_pipe = True
+    raw_cells = re.split(r"(?<!\\)\|", content)
+    if len(raw_cells) < 2 and not edge_pipe:
+        return None
+
+    cells = []
+    column = left
+    for raw_cell in raw_cells:
+        leading = len(raw_cell) - len(raw_cell.lstrip())
+        cells.append((raw_cell.strip(), column + leading))
+        column += len(raw_cell) + 1
+    return cells
+
+
+def _markdown_table_cells(lines):
+    """Map ordinary Markdown table rows to their prose cells.
+
+    The separator row anchors detection, so pipe-containing prose is not
+    treated as a table. Both leading-pipe and no-leading-pipe table styles are
+    accepted. As in GFM, the header and separator must have the same number of
+    cells, but a body row may have fewer or more. A line without a pipe ends
+    the table.
+    """
+    table_cells = {}
+    index = 1
+    while index < len(lines):
+        separator = _split_table_row(lines[index])
+        header = _split_table_row(lines[index - 1])
+        if (not separator or not header or len(separator) != len(header)
+                or not all(TABLE_SEPARATOR_CELL.fullmatch(cell)
+                           for cell, _ in separator)):
+            index += 1
+            continue
+
+        table_cells[index - 1] = header
+        table_cells[index] = []
+        index += 1
+        while index < len(lines):
+            row = _split_table_row(lines[index])
+            if not row:
+                break
+            table_cells[index] = row
+            index += 1
+    return table_cells
+
+
+def _dangling_conjunction_findings(text, filename):
+    lines = text.splitlines()
+    findings = []
+    fence = None
+    # The HTML block open outside the item loop below, as _html_block_step returns it,
+    # the content indent of the list item that holds it (0: no item), and whether
+    # GitHub shows its text.
+    html_block, html_container, html_shows = None, 0, False
+    # The content indents of the list items that can still hold a line, outermost first.
+    open_items = []
+    # A paragraph is open before the line: a text line with less indent continues it
+    # (a lazy line). PARAGRAPH_INDENT is the content indent of the list item that holds
+    # the paragraph (0: no item), or None in a quote. As in cmark-gfm, type 7 cannot
+    # interrupt the paragraph on a line that is in that item too.
+    in_paragraph, paragraph_indent = False, None
+
+    def close_items(indent):
+        """A line with INDENT ends the list items with more content indent."""
+        while open_items and open_items[-1] > indent:
+            open_items.pop()
+
+    index = 0
+    while index < len(lines):
+        line = lines[index]
+        if _fence_ends_with_quote(fence, line):
+            # The quote of the fence ended: judge this line as any other line.
+            fence = None
+        if fence is None:
+            # A line with less indent than the item that holds the block ends the item,
+            # and the block in it. GitHub still hides the text after an open comment up to
+            # "-->", as in the item loop below.
+            if (html_block is not None and html_block is not HTML_COMMENT_END
+                    and line.strip(" \t") and _indent_width(line) < html_container):
+                html_block = None
+            open_block = html_block
+            continues = (in_paragraph and paragraph_indent is not None
+                         and _indent_width(line) >= paragraph_indent)
+            # A block starts at the indent of the list item that holds the line, not at
+            # the start of the line.
+            line_indent = _indent_width(line)
+            block_indent = next((indent for indent in reversed(open_items)
+                                 if indent <= line_indent), 0)
+            block_line = line.expandtabs(4)[block_indent:]
+            in_html, html_block = _html_block_step(html_block, block_line, continues)
+            if in_html and open_block is None:
+                close_items(line_indent)
+                html_container = open_items[-1] if open_items else 0
+                html_shows = _block_shows_text(html_block, block_line)
+            # An HTML block holds literal text: no fence opens in it, and outside list items
+            # no list item starts in it. In a list item, the text that GitHub shows in the
+            # block is item text, so a list marker there starts an item, as before. The
+            # block stays open after that item, so a fence line in it is still text. The
+            # line that ends the block is in the block, so no item starts on it.
+            if (in_html and open_block is not None and html_block is not None
+                    and html_shows and html_container and LIST_ITEM_START.match(line)):
+                in_html = False
+            if in_html:
+                in_paragraph = False
+                index += 1
+                continue
+        is_fence, fence = _fence_step(fence, line)
+        if is_fence or fence:
+            if is_fence and fence is not None:
+                close_items(_indent_width(line))
+            in_paragraph = False
+            index += 1
+            continue
+        start = LIST_ITEM_START.match(line)
+        if not start:
+            # A blank line has only spaces and tabs: a no-break space is text. A line
+            # with only quote markers is blank in its quote.
+            depth, content = _strip_quotes(line.expandtabs(4))
+            # A setext underline in the container of the paragraph makes it a heading,
+            # so the paragraph ends. As a lazy line it is text (CommonMark 4.3).
+            underline = (in_paragraph and paragraph_indent is not None and not depth
+                         and _indent_width(line) >= paragraph_indent
+                         and SETEXT_UNDERLINE.match(content))
+            text_line = (bool(content.strip(" \t")) and not NO_PARAGRAPH.match(content)
+                         and not underline)
+            if line.strip(" \t") and not (in_paragraph and text_line and not depth):
+                close_items(_indent_width(line))
+                paragraph_indent = None if depth else (open_items[-1] if open_items else 0)
+            in_paragraph = text_line
+            index += 1
+            continue
+
+        # A list item that starts with a number other than 1, and an empty list item,
+        # cannot interrupt a paragraph (CommonMark 5.2): GitHub reads the line as text of
+        # the paragraph. An empty "-" item there is a setext underline, so the paragraph
+        # ends. The pass still reads the lines of the item, but with no HTML block in them
+        # (they are paragraph text), and after the item the paragraph stays open in its
+        # container. The item is not open for later lines.
+        marker, body = start.group("marker"), start.group("body")
+        outer_paragraph = (in_paragraph and paragraph_indent is not None
+                           and _indent_width(line) >= paragraph_indent
+                           and ((marker[:-1].isdigit() and marker[:-1].lstrip("0") != "1")
+                                or (not body.strip(" \t") and marker != "-")))
+        outer_indent = paragraph_indent
+        close_items(len(start.group("indent")))
+        # After five or more spaces the text is indented code, and the content column
+        # is one space after the marker (CommonMark 5.2).
+        gap = len(start.group("gap"))
+        if gap > 4 and body.strip(" \t"):
+            gap = 1
+        content_indent = len(start.group("indent")) + len(start.group("marker")) + gap
+        # A list marker in an open HTML block around the item is text of the block on
+        # GitHub, so the item is not open for later lines (see the item loop below).
+        if not outer_paragraph and html_block is None:
+            open_items.append(content_indent)
+        # An HTML block can start on the line of the marker. It is in the item, and it
+        # ends when the item ends. After five or more spaces the text is indented code.
+        body_html, item_html, item_shows = False, None, False
+        # In an open HTML block around the item, the body is text of that block.
+        if html_block is None and not outer_paragraph and len(start.group("gap")) <= 4:
+            body_html, item_html = _html_block_step(None, body, False)
+            item_shows = _block_shows_text(item_html, body)
+        item_paragraph = outer_paragraph or (bool(body.strip(" \t")) and not body_html
+                                             and not NO_PARAGRAPH.match(body))
+        # A block of code or hidden text that stays open shows no item text (see below).
+        item_lines = [] if item_html and not item_shows else [(index, body)]
+        next_index = index + 1
+        item_fence = None
+        while next_index < len(lines):
+            candidate = lines[next_index]
+            # An open HTML block around this item (a list marker in it started the item,
+            # see above) goes on over the lines of the item, as raw item text, up to a
+            # blank line, its end pattern or a line with less indent than its container.
+            if html_block is not None:
+                if candidate.strip(" \t") and _indent_width(candidate) < html_container:
+                    html_block = None
+                else:
+                    in_outer_html, html_block = _html_block_step(html_block, candidate, False)
+                    if html_block is None:
+                        # The block ended. GitHub reads the next lines in its container.
+                        content_indent = html_container
+                    if in_outer_html:
+                        # The line that ends the block is in it, also after a list marker.
+                        if (html_block is not None
+                                and not _is_list_continuation(candidate, content_indent)):
+                            break
+                        item_lines.append((next_index, candidate))
+                        item_paragraph = False
+                        next_index += 1
+                        continue
+            if _fence_ends_with_quote(item_fence, candidate):
+                # The quote of the fence ended: judge this line as any other line.
+                item_fence = None
+            if item_html is HTML_COMMENT_END:
+                # GitHub hides the text from "<!--" up to "-->", also over the end of the
+                # item and over the next items. The text after "-->" joins this item. No
+                # item, block or fence starts in the comment.
+                end = HTML_COMMENT_END.search(candidate)
+                if end:
+                    item_html = None
+                    # spaces keep the columns
+                    item_lines.append((next_index, " " * end.end() + candidate[end.end():]))
+                item_paragraph = False
+                next_index += 1
+                continue
+            in_html = hidden = False
+            if item_fence is None and not outer_paragraph:
+                if candidate.strip() and _indent_width(candidate) < content_indent:
+                    item_html = None  # the line ends the item, and the HTML block in it
+                else:
+                    # The item holds its lines from the content indent on.
+                    open_block = item_html
+                    rest = candidate.expandtabs(4)[content_indent:]
+                    in_html, item_html = _html_block_step(open_block, rest, item_paragraph)
+                    if in_html and open_block is None:
+                        item_shows = _block_shows_text(item_html, rest)
+                    # A <pre> block holds code, and GitHub hides a comment and a "<?", "<!X"
+                    # or CDATA block. As in a fence, the lines of such a block, from its
+                    # first line up to its last line, are no item text, and no list item
+                    # starts there. The last line can show text after the end of the block.
+                    # (An open end pattern is not None or False.)
+                    hidden = bool(item_html) and not item_shows
+            if hidden:
+                item_paragraph = False
+                next_index += 1
+                continue
+            # A line outside the item ends it, also when it opens a fence. The line that ends
+            # an HTML block of the item is in the block, also after a list marker.
+            if (item_fence is None and not _is_list_continuation(candidate, content_indent)
+                    and not (in_html and item_html is None)):
+                break
+            # Fence delimiters are state markers, not meaningful item lines.
+            # In an HTML block, a fence line is text.
+            if not in_html:
+                is_fence, item_fence = _fence_step(item_fence, candidate)
+                if is_fence or item_fence:
+                    item_paragraph = False
+                    next_index += 1
+                    continue
+            item_lines.append((next_index, candidate))
+            item_paragraph = (not in_html and bool(candidate.strip(" \t"))
+                              and not NO_PARAGRAPH.match(candidate.expandtabs(4)[content_indent:]))
+            next_index += 1
+        # A list marker in a block of the item whose text GitHub shows ends the item, as
+        # before, but on GitHub the block goes on. Hand it on, so that the lines after the
+        # next item are still in it.
+        if (item_html is not None and next_index < len(lines)
+                and LIST_ITEM_START.match(lines[next_index])
+                and _indent_width(lines[next_index]) >= content_indent):
+            html_block, html_container, html_shows = item_html, content_indent, item_shows
+        # The line after the item can continue the last paragraph of the item.
+        in_paragraph = item_paragraph
+        paragraph_indent = outer_indent if outer_paragraph else content_indent
+
+        meaningful = []
+        for line_index, item_line in item_lines:
+            # Preserve code spans as neutral operands while ignoring their contents.
+            cleaned = INLINE_CODE.sub(" CODE ", item_line).strip()
+            if cleaned:
+                meaningful.append((line_index, cleaned))
+        if meaningful:
+            end_line_index, end_line = meaningful[-1]
+            conjunction = CONJUNCTION_END.search(end_line)
+        else:
+            end_line_index, end_line, conjunction = None, None, None
+        if conjunction:
+            if end_line_index == index:
+                finding_line = index + 1
+                finding_col = start.start("marker") + 1
+            else:
+                raw_end_line = next(
+                    raw for line_index, raw in item_lines
+                    if line_index == end_line_index
+                )
+                masked_end_line = INLINE_CODE.sub(
+                    lambda match: " " * len(match.group(0)), raw_end_line
+                )
+                raw_conjunction = CONJUNCTION_END.search(masked_end_line)
+                finding_line = end_line_index + 1
+                finding_col = raw_conjunction.start() + 1 if raw_conjunction else 1
+            findings.append({
+                "file": filename,
+                "line": finding_line,
+                "col": finding_col,
+                "rule": "dangling-conjunction",
+                "level": "advisory-free",
+                "match": end_line,
+                "message": "List item ends with a coordinating conjunction. Complete the item or join it with the next item.",
+            })
+        index = next_index
+    return findings
+
+
+# A line of YAML front matter: a key, a list item, an indented line, or a comment.
+# A key is quoted, or a plain word that starts with no YAML indicator; ": "
+# or ":" at the end of the line ends it. A plain key with a space is valid
+# YAML, but prose such as "Step 1: run it" looks the same, so it is Markdown.
+_YAML_KEY = (r"""(?:"(?:[^"\\]|\\.)*"|'(?:[^']|'')*'|[^\s#'"\[\]{},&*!|>%@`-]\S*?)"""
+             r"[ \t]*:(?:[ \t]|$)")
+FRONT_MATTER_KEY = re.compile(_YAML_KEY)
+FRONT_MATTER_LINE = re.compile(_YAML_KEY + r"|-(?:[ \t]|$)|[ \t]+\S|#")
+# The key "description", plain or in quotes.
+DESCRIPTION_KEY = re.compile(r"""(?:description|"description"|'description')[ \t]*:(?:[ \t]+|$)""")
+# A block scalar indicator ("|", ">-", "|2+"), with an optional comment.
+BLOCK_SCALAR = re.compile(r"[|>][1-9+-]{0,2}(?:[ \t]+#.*)?")
+# A YAML comment in a plain scalar: "#" at the start or after a space or tab.
+YAML_COMMENT = re.compile(r"(?:^|[ \t])#")
+
+
+def _closing_quote(text, quote):
+    """Return the index of the QUOTE that closes a quoted YAML scalar in TEXT, or None.
+
+    In single quotes, '' is a quote. In double quotes, a backslash escapes the
+    next character.
+    """
+    pos = 0
+    while pos < len(text):
+        if quote == '"' and text[pos] == "\\":
+            pos += 2
+            continue
+        if text[pos] == quote:
+            if quote == "'" and text[pos + 1:pos + 2] == "'":
+                pos += 2
+                continue
+            return pos
+        pos += 1
+    return None
+
+
+def _scalar_pieces(value, parts):
+    """Return the prose pieces of a YAML scalar as (line index, column, text).
+
+    VALUE is the text after the key. PARTS are the lines of the scalar with
+    their line index and the column of their text, the key line first when
+    VALUE is not empty. A block scalar ("|", ">-") keeps each later line up to
+    one with less indent than its first, also a "#". A quoted scalar ends at
+    its closing quote. A plain scalar ends at a comment. A comment after the
+    scalar is not prose.
+    """
+    pieces = []
+    if BLOCK_SCALAR.fullmatch(value):
+        # The first line sets the indent. A line with less indent, such as a
+        # comment, ends the scalar.
+        for piece in parts[1:]:
+            if piece[1] < parts[1][1]:
+                break
+            pieces.append(piece)
+    elif value[:1] in ("'", '"'):
+        index, column, text = parts[0]
+        parts = [(index, column + 1, text[1:])] + parts[1:]
+        for index, column, text in parts:
+            close = _closing_quote(text, value[0])
+            pieces.append((index, column, text if close is None else text[:close]))
+            if close is not None:
+                break
+    else:
+        for index, column, text in parts:
+            comment = YAML_COMMENT.search(text)
+            pieces.append((index, column, text if comment is None else text[:comment.start()]))
+            if comment is not None:
+                break
+    return [piece for piece in pieces if piece[2].strip()]
+
+
+def _front_matter(text):
+    """Find a leading YAML front-matter block.
+
+    Return (text, description): TEXT with the block blanked, so line numbers
+    stay true, and the value of its "description:" key as (line index,
+    column, text) pieces. The block counts only when its first line that is
+    not a comment is a key, and every line is a key, a list item, an indented
+    line, a comment or a blank line. Else it is Markdown: a thematic break,
+    prose and a second thematic break. Only the description is prose, because
+    agents act on it.
+    """
+    lines = text.splitlines()
+    if not lines or lines[0].rstrip() != "---":
+        return text, []
+    for end in range(1, len(lines)):
+        if lines[end].rstrip() == "---":
+            break
+    else:
+        return text, []
+    body = [line for line in lines[1:end] if line.strip()]
+    content = [line for line in body if not line.startswith("#")]
+    if (body and (not content or not FRONT_MATTER_KEY.match(content[0]))
+            or not all(FRONT_MATTER_LINE.match(line) for line in body)):
+        return text, []
+    pieces = []
+    # "description" in quotes is the same key. The first key counts: a later line that
+    # starts with the key can be in a quoted value that goes on over lines.
+    for index in range(1, end):
+        key = DESCRIPTION_KEY.match(lines[index])
+        if key:
+            break
+    else:
+        index = end
+    if index < end:
+        value = lines[index][key.end():].rstrip()
+        parts = [(index, key.end(), value)] if value else []
+        for later in range(index + 1, end):
+            line = lines[later]
+            if line.strip() and line[0] not in " \t":
+                break
+            lead = len(line) - len(line.lstrip(" \t"))
+            if line[lead:].rstrip():
+                parts.append((later, lead, line[lead:].rstrip()))
+        pieces = _scalar_pieces(value, parts)
+    return "\n" * (end + 1) + "\n".join(lines[end + 1:]), pieces
+
+
+def lint(text, filename="<stdin>"):
+    text, description = _front_matter(text)
+    findings = []
+    words_total = 0
+    fence = None
+    html_block = None
+    html_depth = 0
+    in_paragraph = False
+    # content indent of the open list item, so its indented lines are not code
+    list_indent = 0
+    # backtick run length of a code span that is open across lines, or 0
+    open_span = 0
+    paragraph_depth = 0
+    # (line number, column) where a link from an earlier line ends, or None
+    link_end = None
+    html_containers = []
+    lines = text.splitlines()
+    table_cells = _markdown_table_cells(lines)
+    # first occurrence of each synonym-group member: (group_idx, base) -> (line, col, match)
+    seen_synonyms = {}
+    # consecutive prose lines form one block, so wrapped sentences count once
+    block = []
+    in_quote = False
+    after_boundary = True
+
+    # the end index and backtick run index of the paragraph that holds the current line
+    span_end, span_runs = -1, {}
+
+    def flush():
+        findings.extend(_long_sentence_findings(block, filename))
+        block.clear()
+
+    def check(lineno, segment, source_column, line, skip, is_table,
+              add_to_block=True):
+        nonlocal words_total
+        words_total += len(line.split())
+        for rule_id, level, pattern, msg in RULES:
+            for m in pattern.finditer(line):
+                # the modal may end the line before: "may" / "have failed"
+                before = (block[-1][2] + " " if block else "") + line[skip:m.start()]
+                if rule_id == "present-perfect" and MODAL_PERFECT_PREFIX.search(before):
+                    continue
+                findings.append({"file": filename, "line": lineno,
+                                 "col": source_column + m.start() + 1,
+                                 "rule": rule_id, "level": level,
+                                 "match": m.group(0), "message": msg})
+        for gi, group in enumerate(SYNONYM_GROUPS):
+            for base in group:
+                if (gi, base) in seen_synonyms:
+                    continue
+                m = _word_re(base).search(line)
+                if m:
+                    seen_synonyms[(gi, base)] = (
+                        lineno, source_column + m.start() + 1, m.group(0)
+                    )
+        # This flag affects sentence blocks only; rule checks still scan the line.
+        if add_to_block and line[skip:].strip():
+            block.append((lineno, source_column + skip, line[skip:]))
+
+    # The description in the front matter is one paragraph of prose.
+    span = 0
+    values = [piece[2] for piece in description]
+    last_piece = {}  # the last piece with a run of each length
+    for piece_index, piece_value in enumerate(values):
+        for m in BACKTICK_RUN.finditer(piece_value):
+            last_piece[len(m.group())] = piece_index
+    piece_end = None  # (piece, column) where a link from an earlier piece ends
+    piece_brackets = [[], 0]
+    for position, (index, column, value) in enumerate(description):
+        skip = 0
+        if piece_end and piece_end[0] > position:
+            skip = len(value)  # the whole piece is in the title of a link
+        elif piece_end and piece_end[0] == position:
+            skip = piece_end[1]
+        carry = []
+        masked, span = _mask_code_spans(value, span, lambda run: last_piece.get(run, -1) > position, values, position + 1,
+                                        0, skip, carry, piece_brackets)
+        if carry:
+            piece_end = (position + carry[0], carry[1])
+        elif piece_end and piece_end[0] <= position:
+            piece_end = None
+        check(index + 1, value, column, _mask_non_prose(masked), 0, False)
+    flush()
+    def closes_later(run):
+        """True when a run of length RUN closes a span on a later line of this paragraph."""
+        nonlocal span_end, span_runs
+        if not in_paragraph:
+            return False
+        if lineno > span_end:
+            span_end, span_runs = _paragraph_runs(lines, lineno, paragraph_depth, list_indent)
+        return span_runs.get(run, -1) >= lineno
+
+    for lineno, raw_line in enumerate(lines, 1):
+        was_paragraph = in_paragraph
+        if not was_paragraph or _breaks_paragraph(paragraph_depth, raw_line, list_indent):
+            open_span = 0
+            open_brackets = [[], 0]  # the "[" marks that can open link text
+        if _fence_ends_with_quote(fence, raw_line):
+            # The quote that holds the fence ended: judge this line as any other line.
+            fence = None
+        in_html = False
+        if fence is None:
+            # An HTML block can sit in a block quote. It ends when its quote ends. An
+            # open block strips only its own ">" levels: a deeper ">" is its text.
+            if html_block is not None:
+                line_depth, line_unquoted = _strip_quotes(raw_line.expandtabs(4), html_depth)
+                if line_depth < html_depth:
+                    html_block = None
+            if html_block is None:
+                line_depth, line_unquoted = _strip_quotes(raw_line.expandtabs(4))
+            # A deeper quote starts a block quote, which a paragraph outside it does not continue.
+            continues = in_paragraph and line_depth <= paragraph_depth
+            was_html = html_block is not None
+            in_html, html_block = _html_block_step(html_block, line_unquoted, continues)
+            if in_html and not was_html:
+                html_depth = line_depth
+        if in_html:
+            # An HTML block holds literal text until it ends, so no fence opens in it.
+            is_fence = False
+        else:
+            is_fence, fence = _fence_step(fence, raw_line)
+        # A paragraph is open after a line of text that no other block takes.
+        # The paragraph state reads the line after its quote prefix, so a line with only
+        # quote markers is blank inside its quote.
+        depth, unquoted = _strip_quotes(raw_line.expandtabs(4))
+        prev_depth = paragraph_depth
+        # an underline left of the open item's content is a lazy line, not an underline
+        underline = (in_paragraph and depth == paragraph_depth
+                     and _indent_width(unquoted) >= list_indent
+                     and SETEXT_UNDERLINE.match(unquoted))
+        in_paragraph = bool(unquoted.strip() and not in_html and not is_fence and not fence
+                            and not NO_PARAGRAPH.match(unquoted) and not underline)
+        if not in_paragraph:
+            paragraph_depth = 0
+        elif (not was_paragraph or depth > paragraph_depth
+              or (depth < paragraph_depth
+                  and _breaks_paragraph(paragraph_depth, raw_line, list_indent))):
+            paragraph_depth = depth
+        # A lazy line with fewer ">" keeps the depth of the paragraph. A list item
+        # with fewer ">" is no lazy line: it ends the quote and starts a paragraph.
+        if is_fence:
+            flush()
+            # A fence inside a quote does not start the quote: an alert marker after it is text.
+            after_boundary = _strip_quotes(raw_line)[0] == 0
+            continue
+        if fence:
+            continue
+        # CommonMark 4.4: an indented code block. It cannot interrupt a paragraph.
+        rest = _strip_quotes(raw_line.expandtabs(4))[1]
+        indent = _indent_width(rest)
+        if rest.strip() and not in_html:
+            if indent >= list_indent + 4:
+                if not was_paragraph:
+                    in_paragraph = False
+                    continue
+            else:
+                # An item line sets the content indent also after a text line or
+                # in a tight list. A lazy line does not end the item.
+                item = _list_item(rest, list_indent, was_paragraph and depth == prev_depth)
+                if item:
+                    list_indent = len(item.group(0)) if item.group(3) else len(item.group(0)) + 1
+                elif indent < list_indent and not (was_paragraph and in_paragraph):
+                    # only a lazy text line keeps the item open, not a heading or a break
+                    list_indent = 0
+        if HTML_BLOCK_TAG.match(raw_line.lstrip()):
+            for tag in HTML_BLOCK_TAG.finditer(_mask_inline_code(raw_line)):
+                name = tag.group(2).lower()
+                if tag.group(1):
+                    if name in html_containers:
+                        while html_containers:
+                            if html_containers.pop() == name:
+                                break
+                elif not tag.group().endswith("/>"):
+                    html_containers.append(name)
+        segments = table_cells.get(lineno - 1, [(raw_line, 0)])
+        is_table = (lineno - 1) in table_cells
+        quote = BLOCKQUOTE_PREFIX.match(raw_line)
+        is_quote = bool(quote)
+        # judge block boundaries on the text inside any blockquote prefix
+        content = raw_line[quote.end():] if quote else raw_line
+        # a GitHub alert marker is a label that stands alone, as a heading does.
+        # Accept it only in an unindented quote of one level after a blank line,
+        # a fence, or at the start of the text. Elsewhere a list item or a lazy
+        # continuation line can hold the quote, and GitHub shows the marker as
+        # text. Markers inside HTML containers are also plain text.
+        alert = (not html_containers and raw_line.startswith(">") and after_boundary
+                 and quote.group().count(">") == 1
+                 and GITHUB_ALERT.fullmatch(content.strip()))
+        admonition_title = bool(ADMONITION_TITLE.match(content))
+        standalone = STANDALONE_LINE.match(content) or admonition_title or alert
+        if (is_table or not content.strip() or standalone
+                or ANY_LIST_ITEM.match(content) or is_quote != in_quote
+                # a deeper quote starts a new paragraph, so the text before it is no context
+                or depth > prev_depth):
+            flush()
+        in_quote = is_quote
+        after_boundary = not raw_line.strip()
+        for segment, source_column in segments:
+            if in_html:
+                # an HTML block has no Markdown links; backtick pairs stay blanked
+                line = _mask_non_prose(_mask_inline_code(segment))
+            elif is_table:
+                # a table cell holds its own spans and links
+                line = _mask_non_prose(_mask_code_spans(segment, 0, lambda run: False)[0])
+            else:
+                skip = 0
+                if link_end and link_end[0] > lineno:
+                    skip = len(segment)  # the whole line is in the title of a link
+                elif link_end and link_end[0] == lineno:
+                    skip = link_end[1]
+                carry = []
+                masked, open_span = _mask_code_spans(segment, open_span, closes_later, lines, lineno,
+                                                     paragraph_depth if in_paragraph else None,
+                                                     skip, carry, open_brackets)
+                if carry:
+                    link_end = (lineno + carry[0], carry[1])
+                elif link_end and link_end[0] <= lineno:
+                    link_end = None
+                line = _mask_non_prose(masked)
+            # The blockquote prefix and list marker are markup.
+            skip = quote.end() if quote and not is_table else 0
+            marker = None if is_table else ANY_LIST_ITEM.match(segment[skip:])
+            if marker:
+                skip += marker.end()
+            check(
+                lineno, segment, source_column, line, skip, is_table,
+                add_to_block=not admonition_title,
+            )
+            if is_table:
+                flush()  # each table cell is its own block
+        if standalone:
+            flush()
+    flush()
+    # synonym rotation: flag each member after the first, at its first occurrence
+    for gi, group in enumerate(SYNONYM_GROUPS):
+        present = [(seen_synonyms[(gi, b)], b) for b in group if (gi, b) in seen_synonyms]
+        if len(present) > 1:
+            present.sort()  # document order
+            first_base = present[0][1]
+            for (lineno, col, match), base in present[1:]:
+                findings.append({"file": filename, "line": lineno, "col": col,
+                                 "rule": "synonym-rotation", "level": "advisory-free",
+                                 "match": match,
+                                 "message": f"'{base}' and '{first_base}' name the same action. Pick one and use it every time."})
+    findings.extend(_dangling_conjunction_findings(text, filename))
+    findings.sort(key=lambda f: (f["line"], f["col"]))
+    return findings, words_total
+
+
+def report(findings, words_total, as_json, hard_count, baseline):
+    rate = round(len(findings) * 100 / words_total, 1) if words_total else 0.0
+    if as_json:
+        print(json.dumps({"violations": findings, "count": len(findings),
+                          "hard_count": hard_count, "baseline": baseline,
+                          "words": words_total, "per_100_words": rate}, indent=2))
+        return
+    for f in findings:
+        print(f"{f['file']}:{f['line']}:{f['col']} {f['rule']}: {f['message']} [{f['match']}]")
+    print(f"\n{len(findings)} violations ({hard_count} hard, baseline {baseline}), "
+          f"{words_total} words, {rate} per 100 words")
+    print("Hedges/modality (may, might, could) are never flagged: confidence is content.")
+
+
+def selftest():
+    bad = ("The panel is removed; spin up the job. "
+           "Perform an analysis of the seamless log. "
+           "We have received the report.")
+    bad_findings, _ = lint(bad)
+    rules = {f["rule"] for f in bad_findings}
+    for expected in ("semicolon", "phrasal-verb", "nominalization",
+                     "marketing-adjective", "passive-voice", "present-perfect"):
+        assert expected in rules, (expected, bad_findings)
+
+    # MkDocs admonition title variants, including multiple classes and no space.
+    for title in (
+        "!!! note",
+        "!!!note",
+        "!!! info inline end \"Title\"",
+        "!!! danger highlight blink \"Title\"",
+        "??? note",
+        "???+ note",
+        '!!! note "Quoted title"',
+    ):
+        assert ADMONITION_TITLE.fullmatch(title), title
+    for title in ("!!!+ note", "!!!+note"):
+        assert not ADMONITION_TITLE.fullmatch(title), title
+
+    # Long admonition titles are headings, not prose sentences.
+    long_title = '!!! note "' + " ".join(["title"] * 30) + '"'
+    findings, _ = lint(long_title)
+    assert not any(f["rule"] == "long-sentence" for f in findings), findings
+
+    # A standalone title does not join the four-space-indented body sentence.
+    body = " ".join(["word"] * 26) + "."
+    findings, _ = lint('!!! note "Title"\n    ' + body)
+    long_sentences = [
+        f for f in findings if f["rule"] == "long-sentence"
+    ]
+    assert len(long_sentences) == 1, findings
+    assert long_sentences[0]["match"] == "26 words", findings
+
+    # A four-space-indented title inside a list item is also standalone.
+    list_title = '!!! info inline end "' + " ".join(["title"] * 30) + '"'
+    findings, _ = lint("- item\n    " + list_title + "\n        " + body)
+    long_sentences = [
+        f for f in findings if f["rule"] == "long-sentence"
+    ]
+    assert len(long_sentences) == 1, findings
+    assert long_sentences[0]["line"] == 3, findings
+    assert long_sentences[0]["match"] == "26 words", findings
+
+    # A long whitespace-only title candidate must fail quickly.
+    assert not ADMONITION_TITLE.fullmatch("!!!" + " " * 32000)
+    # Many class words followed by a broken title must also fail without backtracking.
+    assert not ADMONITION_TITLE.fullmatch("!!! " + "info " * 16000 + '"')
+    # hedges must never be flagged, including modal + perfect infinitive
+    findings, _ = lint("The request may have failed. It could be a timeout. "
+                       "The disk might have filled.")
+    assert findings == [], findings
+    # irregular participles: "has run" is a compound tense as much as "has failed"
+    findings, _ = lint("The task has run. The job has set the flag. We have begun.")
+    assert sum(1 for f in findings if f["rule"] == "present-perfect") == 3, findings
+    findings, _ = lint("The job may have run.")
+    assert not any(f["rule"] == "present-perfect" for f in findings), findings
+    # Modal perfects remain protected across negation and variable whitespace.
+    for modal in ("may", "might", "could", "should", "would", "must"):
+        for gap in (" ", "  ", "\t", " not "):
+            findings, _ = lint(f"The task {modal}{gap}have run.")
+            assert not any(f["rule"] == "present-perfect" for f in findings), findings
+    findings, _ = lint("The task couldn't have run. The task MAY NOT HAVE RUN.")
+    assert not any(f["rule"] == "present-perfect" for f in findings), findings
+    findings, _ = lint("It might already have finished. It could not possibly have run. "
+                       "It may never have started. It should probably have stopped.")
+    assert not any(f["rule"] == "present-perfect" for f in findings), findings
+    findings, _ = lint("It cannot have run. It won't have finished. It can't have stopped. "
+                       "It may well have failed. It may very well have failed.\n\n"
+                       "The request may\nhave failed.")
+    assert not any(f["rule"] == "present-perfect" for f in findings), findings
+    # a modal in an earlier paragraph does not protect the next one
+    findings, _ = lint("It may\n\nhave failed.")
+    assert sum(f["rule"] == "present-perfect" for f in findings) == 1, findings
+    # the quote prefix and the list marker are markup: a modal at the end of the
+    # line before still protects the perfect, also in a quote and in a list
+    for text in ("> It may\n> have failed.", "> > It may\n> > have failed.",
+                 "It may\nhave failed.", "- It may\n  have run.", "> - It may\n>   have run."):
+        findings, _ = lint(text)
+        assert not any(f["rule"] == "present-perfect" for f in findings), (text, findings)
+    for text in ("> It may\n>\n> have failed.", "> It may\n>> have failed."):
+        findings, _ = lint(text)
+        perfect_lines = [f["line"] for f in findings if f["rule"] == "present-perfect"]
+        assert perfect_lines == [text.count("\n") + 1], (text, findings)
+    # an adverb without a modal is not a hedge
+    findings, _ = lint("The task already has run. They probably have finished.")
+    assert sum(f["rule"] == "present-perfect" for f in findings) == 2, findings
+    findings, _ = lint("The task has run. We have begun. The flag is set.")
+    assert sum(f["rule"] == "present-perfect" for f in findings) == 2, findings
+    assert any(f["rule"] == "passive-voice" for f in findings), findings
+    findings, _ = lint("The task is gone.")
+    assert not any(f["rule"] == "passive-voice" for f in findings), findings
+    # code blocks skipped
+    findings, _ = lint("```\nx = a; y = b\n```")
+    assert findings == []
+    # a fence closes only on the same character, at least as long, with nothing after it
+    for text in ("````md\n```\n````\n\nProse; here.",
+                 "```\n~~~\n```\n\nProse; here.",
+                 "~~~~\n~~~\n~~~~\n\nProse; here."):
+        findings, _ = lint(text)
+        assert [f["rule"] for f in findings] == ["semicolon"], (text, findings)
+    findings, _ = lint("```\n```sh\nx = a; y = b\n```")
+    assert findings == [], findings
+    # a fence inside a blockquote is a fence
+    findings, _ = lint("Text.\n\n> ```\n> x = a; y = b\n> ```\n\nProse; here.")
+    assert [(f["rule"], f["line"]) for f in findings] == [("semicolon", 7)], findings
+    # the dangling-conjunction rule uses the same fence state
+    findings, _ = lint("````md\n```\n- Do this and\n````")
+    assert findings == [], findings
+    findings, _ = lint("- Do this\n  ````md\n  ```\n  code and\n  ````\n- Do that")
+    assert findings == [], findings
+    # it tracks HTML blocks: there no fence opens and no list item starts
+    for text, lines in (("<div>\n```\n</div>\n\n- Open the valve and\n", [5]),
+                        ("<!--\n```\n-->\n- Open the valve and\n", [4]),
+                        ("<div>\n- Open the valve and\n</div>\n", []),
+                        ("- x\n<div>\n- y and\n", []),
+                        # an HTML block in an item, also on the line of the marker
+                        ("- Open the valve\n\n  <div>\n  ```\n  </div>\n\n- Close it and\n", [7]),
+                        ("- x\n  - y\n  <div>\n  ```\n  </div>\n\n- z and\n", [7]),
+                        ("- <pre>\n  ```\n  </pre>\n- y and\n", [4]),
+                        # a line with less indent ends the item, and the HTML block in it
+                        ("- x\n  <div>\n- y and\n", [3]),
+                        ("- <pre>\n```\n- y and\n```\n", []),
+                        ("- a\n  - b\n  <div>\n- c and\n", [4]),
+                        # the lines in a <pre> block or a comment are no item text, and no
+                        # item starts there; GitHub hides the text after an open comment
+                        ("- Close it\n  <pre>\n  ```\n  - Step and\n- Next\n", []),
+                        ("- a\n  <!--\n  x\n- b and\n", []),
+                        ("- a\n  <?php\n  x and\n- b\n", []),
+                        # an open comment joins the next items up to "-->"
+                        ("- Run tests and\n  <!-- pending\n- Next:\n  <!-- closes -->\n  finish the checklist.\n", []),
+                        ("- Run tests and\n  <!-- pending\n- Next\n", [1]),
+                        # also the first line of such a block, when the block stays open
+                        ("- Change summary:\n  <!-- Explain the change and\n  include test results.\n", []),
+                        ("- <pre>Expected an expression after and\n  at parser.py:12\n", []),
+                        ("- Fix it\n  <!-- old -->\n  <pre>a</pre> and\n", [3]),
+                        # GitHub shows the text of a <script> block: it escapes the tag
+                        ("- a\n  <script>\n  x and\n- b\n", [3]),
+                        ("- a\n  <!--\n- b and\n\n<div>\n-->\n</div>\n\n- c and\n", [9]),
+                        # in an item, the text of a <div> block is item text: a list marker
+                        # there starts an item, as before
+                        ("- <div>\n  ```\n  - Step and\n- Next\n", [3]),
+                        ("- a\n  - b\n  <div>\n  - c and\n  </div>\n", [4]),
+                        ("  <!--\n  ```\n- Open it and\n", []),
+                        # one tag alone after a text line continues the paragraph, so the
+                        # fence after it opens
+                        ("Text\n<span>\n```\n```\n- a and\n", [5]),
+                        ("- x\n  <span>\n  ```\n  - y and\n  ```\n", []),
+                        # a no-break space is no whitespace in a tag
+                        ("<div\u00a0class=\"note\">\n- Next and\n", [2]),
+                        ("<br/>\u00a0\n- Next and\n", [2]),
+                        # a line with a no-break space is text, also after "#"
+                        ("#\u00a0Notes\n<img src=\"a.png\">\n- Run checks and\n", [3]),
+                        ("Text\n\u00a0\n<img src=\"a.png\">\n- Run checks and\n", [4]),
+                        # a line with only ">" ends the paragraph, and a heading in a quote too
+                        (">\n<img src=\"a.png\">\n- Run checks and\n", []),
+                        ("> # Head\n<img src=\"a.png\">\n- Run checks and\n", []),
+                        # a tag alone interrupts a paragraph of a quote or of an item that
+                        # the line is not in, as in cmark-gfm
+                        ("> Note\n<img src=\"a.png\">\n- Run checks and\n", []),
+                        ("- a\n<span>\n- b and\n", []),
+                        ("- a\nlazy\n<span>\n- b and\n", []),
+                        ("- a\n  - b\n  text\n  <span>\n  ```\n  ```\n- c and\n", [7]),
+                        ("- a\n  - b\n\n  text\n  <span>\n  ```\n  ```\n- c and\n", [8]),
+                        # "2." cannot interrupt a paragraph, so the tag line is paragraph text
+                        ("Text\n2. Example\n<br>\n- Open the valve and\n", [4]),
+                        ("- a\n  Text\n  2. Example\n  <br>\n- Open the valve and\n", [5]),
+                        ("Text\n1. Example\n<br>\n- Open the valve and\n", []),
+                        ("Screenshot:\n2. <img src=\"example.png\">\n<br>\n- Run tests and\n", [4]),
+                        ("Text\n2. # Heading\n<br>\n- Run tests and\n", [4]),
+                        # an empty item cannot interrupt a paragraph; an empty "-" is a setext underline
+                        ("Text\n1. \n<br>\n- Run tests and\n", [4]),
+                        ("Text\n* \n<br>\n- Run tests and\n", [4]),
+                        ("Text\n- \n<br>\n- Run tests and\n", []),
+                        # its lines are paragraph text: no HTML block starts in them
+                        ("Example:\n2. Steps\n   <details>\n   ```markdown\n- Run tests and\n   ```\n"
+                         "   </details>\n", []),
+                        ("Text\n2. Steps and\n   <details>\n", []),
+                        ("Text\n2. Steps\n    <details>\n- Run tests and\n", [4]),
+                        # a setext underline ends the paragraph, so "2." starts a list
+                        ("Setup\n=====\n2. Install\n   - Linux\n   <div>Note</div>\n- Run tests and\n", [6]),
+                        # a lazy underline is text, and "<br>" after the "2." item is a block
+                        ("- Setup\n=====\n2. Install\n<br>\n- Run tests and\n", []),
+                        # an HTML block in an item goes on after a list marker in it
+                        ("- <div>\n  ~~~\n  - Note\n  ~~~\n  </div>\n\n- Open the valve and\n", [7]),
+                        ("-     code\n  <div>Note</div>\n- Open the valve and\n", [3]),
+                        ("1. <details>\n   ~~~\n   - Linux\n   ~~~\n   </details>\n\n2. Run tests and\n", [7]),
+                        ("- <div>\n  - Note\n\n  ~~~\n  - x and\n  ~~~\n- y and\n", [7]),
+                        ("- <div>\n  ~~~\n  - Note\n    ~~~\n  </div>\n\n- Open the valve and\n", [7]),
+                        # a block starts at the indent of the item that holds the line
+                        ("- Example:\n  - Linux\n  <div>\n  ~~~\n  </div>\n\n    <div>\n    ~~~\n    </div>\n\n"
+                         "- Run tests and\n", [11]),
+                        # after the block, the lines are in the item that held the block
+                        ("- <div>\n  ~~~\n  - Note\n  </div>\n\n    <div>\n  ~~~\n    </div>\n\n- Run tests and\n",
+                         [10]),
+                        ("- <div>\n  ~~~\n  - Note\n\n    <div>\n  ~~~\n    </div>\n\n- Run tests and\n", [9]),
+                        # a list marker on the line that ends the block is text of the block
+                        ("- <textarea>\n  ~~~\n  - Note </textarea>\n\n    <div>\n  ~~~\n    </div>\n\n"
+                         "- Run tests and\n", [9]),
+                        ("- <pre>\n  ~~~\n  - Note </pre>\n\n    <div>\n  ~~~\n    </div>\n\n- Run tests and\n", [9])):
+        findings, _ = lint(text)
+        assert [f["line"] for f in findings if f["rule"] == "dangling-conjunction"] == lines, (text, findings)
+    # a quoted fence ends with its quote, and a deeper ">" in a fence is code
+    for text, lines in (("> ```\n> code\n\nProse; here.", [4]),
+                        ("```\n> ```\nx; y\n```\n\nProse; here.", [6]),
+                        (">  > ```\n>  > x; y\n>  > ```\n\nProse; here.", [5]),
+                        ("> ```\n>     ```\n> x; y\n> ```\n\nProse; here.", [6]),
+                        ("```foo`bar\nProse; here.", [2]),
+                        ("```\nx; y\n```` \t\n\nProse; here.", [5])):
+        findings, _ = lint(text)
+        assert [f["line"] for f in findings if f["rule"] == "semicolon"] == lines, (text, findings)
+    # the end of a quoted fence in a list item does not join the next line to the item
+    findings, _ = lint("- Item\n  > ```\n  > code\nProse and")
+    assert findings == [], findings
+    # a tab is indent up to the next tab stop of 4, from its real column, also after ">"
+    for text, line in (("\t```\n\tx = 1\n\t```\n\nProse; here.", 5),
+                       ("```\nx\n\t```\ny; z\n```\n\nProse; here.", 7),
+                       ("> ```\n> x\n>\t```\n> Use a; b.", 4),
+                       ("> ```\n> x\n>  \t```\n> Use a; b.", 4),
+                       (">\t> ```\n>\t> x; y\n>\t> ```\n\nProse; here.", 5),
+                       ("- Item\n  ```\n\t```\n\nProse; here.", 5)):
+        findings, _ = lint(text)
+        assert [(f["rule"], f["line"]) for f in findings] == [("semicolon", line)], (text, findings)
+    # a tab-indented fence stays in its list item
+    findings, _ = lint("- Run the build and\n\t```\n\tmake\n\t```\n  check the log.")
+    assert findings == [], findings
+    # a ">" line in an HTML block is text, not a quoted fence. A blank line ends a
+    # block that starts with a tag, and a quoted fence after it is a fence.
+    for text, lines in (("<div>\n> ```\n> Prose; here.\n> ```\n</div>", [3]),
+                        ("<custom>\n> ```\n> Prose; here.\n> ```\n</custom>", [3]),
+                        ("<!--\n> ```\n> Prose; here.\n> ```\n-->", [3]),
+                        ("<div>\n\n> ```\n> x; y\n> ```\n\nProse; here.", [7]),
+                        ("<details>\n\n> ```\n> a; b\n> ```\n\n</details>\n\nProse; here.", [9]),
+                        ("<pre>\n> ```\n> Prose; here.\n> ```\n</pre>", [3]),
+                        ("<?php\n> ```\n> Prose; here.\n> ```\n?>", [3]),
+                        ("<![CDATA[\n> ```\n> Prose; here.\n> ```\n]]>", [3]),
+                        # inline HTML is no HTML block: the quoted fence after it is a fence
+                        ("<span>Label</span>\n> ```\n> a; b\n> ```", []),
+                        ("<kbd>Enter</kbd> starts the job.\n> ```\n> a; b\n> ```", []),
+                        ("Text.\n<custom>\n> ```\n> a; b\n> ```", []),
+                        # a type 7 tag after a heading or a fence starts a block
+                        ("# Title\n<custom>\n> ```\n> Prose; here.\n> ```\n</custom>", [4]),
+                        ("```\nx\n```\n<img src=\"a.png\">\n> ```\n> Prose; here.\n> ```", [6]),
+                        # attribute values may hold "<" and ">", an invalid attribute is no tag
+                        ('<span title="a > b">\n> ```\n> Prose; here.\n> ```', [3]),
+                        ("<custom a=>\n> ```\n> a; b\n> ```", []),
+                        # a setext underline, an empty quote line and a thematic break of one
+                        # character end a paragraph, so a type 7 tag after them starts a block
+                        ("Title\n=====\n<custom>\n> ```\n> Prose; here.\n> ```\n</custom>", [5]),
+                        ("> Quote.\n>\n<custom>\n> ```\n> Prose; here.\n> ```\n</custom>", [5]),
+                        ("-*-\n<custom>\n> ```\n> a; b\n> ```", []),
+                        ("Title\n-\n<custom>\n> ```\n> Prose; here.\n> ```\n</custom>", [5]),
+                        ("> Quote.\n> >\n<custom>\n> ```\n> Prose; here.\n> ```\n</custom>", [5]),
+                        # an underline with no paragraph, or as a lazy line, is text
+                        ("Intro.\n\n==\n<custom>\n> ```\n> a; b\n> ```", []),
+                        ("> Title\n===\n<custom>\n> ```\n> a; b\n> ```", []),
+                        ("> > Title\n> lazy\n> ===\n<custom>\n> ```\n> a; b\n> ```", []),
+                        # a quoted heading or thematic break is no paragraph line
+                        ("> # Note\n> ===\n<custom>\n> ```\n> a; b\n> ```", []),
+                        ("> Note\n> ***\n<custom>\n> ```\n> Prose; here.\n> ```\n</custom>", [5]),
+                        # the line that ends a quoted fence can start an HTML block
+                        ("> ```\n> x\n<custom>\n> ```\n> Prose; here.", [5]),
+                        # a type 1 name is no type 7 tag
+                        ("<pre/>\n> ```\n> a; b\n> ```", []),
+                        ("</pre>\n> ```\n> a; b\n> ```", []),
+                        ("<pre-x>\n> ```\n> Prose; here.\n> ```", [3]),
+                        # no fence opens in an HTML block, also without ">"
+                        ("<pre>\n```\nsample\n</pre>\n\nThe paragraph is here; it must be checked.", [6]),
+                        ("<div>\n```\n\nProse; here.", [4]),
+                        ("<DIV class=x>text\n> ```\n> Prose; here.\n> ```", [3])):
+        findings, _ = lint(text)
+        assert [f["line"] for f in findings if f["rule"] == "semicolon"] == lines, (text, findings)
+    # a fence at column 0 ends a list item
+    findings, _ = lint("- Do this and\n```\nx\n```\n  done")
+    assert [(f["rule"], f["line"]) for f in findings] == [("dangling-conjunction", 1)], findings
+    # a quoted fence also ends on a line with fewer quote levels that is not blank
+    for text, line in (("> ```\n> x\nProse; here.", 3), ("> > ```\n> x; y", 2)):
+        findings, _ = lint(text)
+        assert [(f["rule"], f["line"]) for f in findings] == [("semicolon", line)], (text, findings)
+    # a closing fence up to three columns deeper than the opening fence closes it
+    findings, _ = lint("```\nx; y\n   ```\n\nProse; here.")
+    assert [(f["rule"], f["line"]) for f in findings] == [("semicolon", 5)], findings
+    # a tilde fence opens with a backtick in its info string
+    findings, _ = lint("~~~ a`b\nx; y\n~~~\n\nProse; here.")
+    assert [(f["rule"], f["line"]) for f in findings] == [("semicolon", 5)], findings
+    # an alert marker after a fence in the same quote is text, as GitHub shows it
+    body = " ".join(["word"] * 25) + "."
+    findings, _ = lint(f"> ```\n> x\n> ```\n> [!NOTE]\n> {body}")
+    assert [f["rule"] for f in findings] == ["long-sentence"], findings
+    # a quoted fence after a list item does not join the item
+    findings, _ = lint("- Do this and\n\n> ```\n> code\n> ```\n  done")
+    assert [f["rule"] for f in findings] == ["dangling-conjunction"], findings
+    # all supported list markers, case variants, and trailing whitespace
+    findings, _ = lint(
+        "- Confirm the target and\n"
+        "* Record the result OR  \n"
+        "+ Close the panel\n"
+        "1. Start the task and\n"
+        "2) Stop the task OR"
+    )
+    dangling = [f for f in findings if f["rule"] == "dangling-conjunction"]
+    assert len(dangling) == 4, dangling
+    assert [f["line"] for f in dangling] == [1, 2, 4, 5], dangling
+    assert [f["col"] for f in dangling] == [1, 1, 1, 1], dangling
+    assert all(f["level"] == "advisory-free" for f in dangling), dangling
+
+    # valid continuation lines and standalone four-space code are ignored
+    findings, _ = lint("  - Confirm the target and\n    record the result.")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings)
+    findings, _ = lint("- Confirm the target and\n\trecord the result.")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings), findings
+    findings, _ = lint("- Confirm the target\n  and")
+    dangling = [f for f in findings if f["rule"] == "dangling-conjunction"]
+    assert len(dangling) == 1 and dangling[0]["line"] == 2, dangling
+    assert dangling[0]["col"] == 3, dangling
+    findings, _ = lint("    - code and")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings)
+    findings, _ = lint("> - Confirm the target and\n> - Record the result or")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings)
+    findings, _ = lint("- Do this and\n~~~\ncode and\n~~~")
+    dangling = [f for f in findings if f["rule"] == "dangling-conjunction"]
+    assert len(dangling) == 1 and dangling[0]["line"] == 1, dangling
+    findings, _ = lint("```text\n- code and\n```")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings)
+
+    # one- and three-space markers and ordered continuation width
+    findings, _ = lint(" - Start the task and\n   record the result.")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings)
+    findings, _ = lint("-  Start the task and\n   record the result.")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings)
+    findings, _ = lint("-\tStart the task and")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings)
+    findings, _ = lint("   - Start the task and", filename="fixture.md")
+    dangling = [f for f in findings if f["rule"] == "dangling-conjunction"]
+    assert len(dangling) == 1 and dangling[0]["col"] == 4, dangling
+    assert dangling[0]["file"] == "fixture.md"
+    assert dangling[0]["match"].endswith("and")
+    assert "Complete the item" in dangling[0]["message"]
+    findings, _ = lint("100. Start the task and\n  unrelated text")
+    dangling = [f for f in findings if f["rule"] == "dangling-conjunction"]
+    assert len(dangling) == 1, dangling
+    findings, _ = lint("- Start the task and.\n- Stop the task or,")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings)
+    findings, _ = lint("- Start the task and\n\n  record the result.")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings)
+    findings, _ = lint("- Parent item and\n  - Nested item or")
+    dangling = [f for f in findings if f["rule"] == "dangling-conjunction"]
+    assert [f["line"] for f in dangling] == [1, 2], dangling
+
+    # ordinary prose, inline code, and fenced code are ignored
+    findings, _ = lint("The process may include steps and")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings)
+    findings, _ = lint("- Use `and` as a label")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings)
+    findings, _ = lint("- Combine `left` and `right`")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings), findings
+    findings, _ = lint("~~~\n- code and\n~~~")
+    assert not any(f["rule"] == "dangling-conjunction" for f in findings)
+    findings, _ = lint(("word " * 30).strip() + ".")
+    assert any(f["rule"] == "long-sentence" for f in findings)
+    # Markdown table syntax is layout, not prose. Each cell stays lintable.
+    short_cell = " ".join(f"term{number}" for number in range(1, 25)) + "."
+    for table in (
+            "| Label | Detail |\n"
+            "| --- | --- |\n"
+            f"| Clear | {short_cell} |",
+            "Label | Detail\n"
+            "--- | ---\n"
+            f"Clear | {short_cell}"):
+        findings, words_total = lint(table)
+        assert not any(f["rule"] == "long-sentence" for f in findings), findings
+        assert words_total == 27, words_total
+    # GFM also accepts one-dash separators, one column, and ragged body rows.
+    short_row = " ".join(f"term{number}" for number in range(1, 9))
+    for table in (
+            "| Label | Detail |\n|-|-|\n" + f"| {short_row} | {short_row} |\n" * 5,
+            "| Label |\n| --- |\n" + f"| {short_row} |\n" * 8,
+            "| Label | Detail |\n| --- | --- |\n| x |\n"
+            + f"| {short_row} | {short_row} | extra |\n" * 2):
+        findings, _ = lint(table)
+        assert not any(f["rule"] == "long-sentence" for f in findings), (table, findings)
+    long_cell = " ".join(f"term{number}" for number in range(1, 27)) + "."
+    findings, _ = lint(
+        "| Label | Detail |\n"
+        "| --- | --- |\n"
+        f"| Clear | {long_cell} |"
+    )
+    long_sentences = [f for f in findings if f["rule"] == "long-sentence"]
+    assert len(long_sentences) == 1, long_sentences
+    assert long_sentences[0]["match"] == "26 words", long_sentences
+    # synonym rotation: second member flagged, first named as the keeper
+    findings, _ = lint("Check the config file. Then verify the output. Verify twice.")
+    rot = [f for f in findings if f["rule"] == "synonym-rotation"]
+    assert len(rot) == 1 and "'verify' and 'check'" in rot[0]["message"], rot
+    # single consistent term: no flag
+    findings, _ = lint("Check the config. Check the output.")
+    assert not any(f["rule"] == "synonym-rotation" for f in findings)
+    # character references and link targets are markup, not prose
+    findings, _ = lint("Press Ctrl&nbsp;C to stop &rarr; &#8594; &#x2192; now. "
+                       "See [the docs](https://x.io/a;b) or <https://x.io/a;b>.")
+    assert not any(f["rule"] == "semicolon" for f in findings), findings
+    findings, _ = lint('See [a](https://x.io/a;b "Title") and [b](<a b;c>) and [c](a(b;c)).')
+    assert not any(f["rule"] == "semicolon" for f in findings), findings
+    # a reference GitHub does not read as one, or that renders ";", is prose
+    for text in ("We closed the Q&As; then we left.", "Use &unknownentity; here.",
+                 "Read this&#59; then stop.", "Read this&#x3b; then stop.", "Read&semi; this."):
+        findings, _ = lint(text)
+        assert [f["rule"] for f in findings] == ["semicolon"], (text, findings)
+    # text in parentheses that is no link destination is prose
+    findings, _ = lint("Read [this](seamless; now).")
+    assert [f["rule"] for f in findings] == ["marketing-adjective", "semicolon"], findings
+    for text in ("See [a](b;c d).", "See [a](b(c;d).", "See [a](<b;c).", "See [a](b;c \"t\" x)."):
+        findings, _ = lint(text)
+        assert [f["rule"] for f in findings] == ["semicolon"], (text, findings)
+    # the link can go on in the next lines of its paragraph, and it must close there
+    for text, cols in (("See [docs](https://x.io/a;b\n).", []),
+                       ("See [a](b;c \"multi\nline title\") now.", []),
+                       ("See [a](b;c\n\"title\") now.", []),
+                       ("See [a](b;c (paren\ntitle)) now.", []),
+                       ("> See [docs](https://x.io/a;b\n> ).", []),
+                       ("See [docs](https://x.io/a;b\n\n).", [(1, 26)]),
+                       ("See [a](b;c\n\n\"title\") now.", [(1, 10)]),
+                       ("See [a](b;c\nnext line) now.", [(1, 10)]),
+                       ("See [a](b(c);d \"oops).", [(1, 13)]),
+                       ("See [a](b(c);d \"oops).\n\nNext; para.", [(1, 13), (3, 5)]),
+                       ("See [a](b;c (paren (no)) now.", [(1, 10)]),
+                       # a block that interrupts the paragraph ends the link
+                       ("See [a](b;c \"x\n- y\") z.", [(1, 10)]),
+                       ("See [a](b;c \"x\n# y\") z.", [(1, 10)]),
+                       ("See [a](b;c \"x\n===\") z.", []),
+                       ("See [a](b;c \"x\n===\n\") z.", [(1, 10)]),
+                       ("See [a](b;c\n> )", [(1, 10)]),
+                       ("> See [a](b;c\n)", []),
+                       ("See [a](b;c \"x\n2. y\") z.", []),
+                       # a backtick in a destination is text, and what starts first wins
+                       ("See [a](https://x.io/a;b`c`tail).", []),
+                       ("See [x](a`b;c) and `d` e; f.", [(1, 25)]),
+                       ("Run `a [b](c;d` now.", []),
+                       ("A `` [a](b;c `` d) e; f", [(1, 21)]),
+                       ("| A | B |\n|-|-|\n| [a](b;c`d`e) | x |", []),
+                       # a lazy line continues the quoted paragraph of the link
+                       ("> First line\nSee [a](b;c\n> ) now.", []),
+                       # a backtick in a title is text, a code span after the link is a span
+                       ("See [a](b \"x`y\") and `z; w` now.", []),
+                       # in a table cell the link starts first, and an HTML block has no links
+                       ("| A | B |\n|-|-|\n| [a](b`c) d; e` f | x |", [(3, 13)]),
+                       ("<div>\nSee [a](b;c) now.\n</div>", [(2, 10)]),
+                       # a heading does not go on in the next line
+                       ("# See [a](b;c\n) now.", [(1, 12)]),
+                       # a title that goes on in the next line holds its backticks there too
+                       ("See [a](b \"x\ny`z\") and `w; v` q.", []),
+                       ("> See [a](b \"x\n> y\n> `z\") and `w; v` q.", []),
+                       ("See [a](b \"x\ny\") a; `b` c", [(2, 6)]),
+                       # a backtick that no code span holds is text, also after a link
+                       ("See [a](b \"y`z\") ``c ` d`` text; last `.", [(1, 32)]),
+                       ("See [a](b \"x\ny`z\") ``c ` d`` text; last `.", [(2, 21)]),
+                       ("Run ```a; b```` now.", [(1, 9)]),
+                       # "](" and a space start no destination
+                       ("]( ;)", [(1, 4)]),
+                       ("]( `);`", []),
+                       ("`[`]( ;)", [(1, 7)]),
+                       ("<aa:;[]( )>", [(1, 5)]),
+                       # an escaped backtick is text, and the rest of its run can open a span
+                       ("\\```;``", []),
+                       ("\\```;```", [(1, 5)]),
+                       ("> \\```;``", []),
+                       ("| A |\n|-|\n| \\```;`` |", []),
+                       # GitHub shows a link title in no text, so it is not prose
+                       ("[a](b \"`;`\")", []),
+                       ("[a](b \"`\n`;`\")", []),
+                       ("[a](b \"`;\")`", []),
+                       ("See [a](b \"x; y\") now; stop.", [(1, 22)]),
+                       ("See [a](b \"x\ny; z\nw\") now; stop.", [(3, 8)]),
+                       # an escaped "]" starts no link, and an autolink that starts first holds its backticks
+                       ("[label\\](url \"visible; prose\")", [(1, 22)]),
+                       ("[a\\](b;c) d", [(1, 7)]),
+                       ("<https://example.com/\\``> visible; prose `.", [(1, 34)]),
+                       ("<https://x.io/`a> b; c `.", [(1, 20)]),
+                       ("See <https://x.io/a`b> and `c; d` e.", []),
+                       ("`a <https://x` b; c>", [(1, 17)]),
+                       # "](" is a link only after a "[" of its paragraph that no "]" closed
+                       ("](b \";\")", [(1, 6)]),
+                       ("](b;c) d", [(1, 4)]),
+                       ("](b \"`;`\")", []),
+                       ("\\[a](b \";\")", [(1, 9)]),
+                       ("[a] ](b \";\")", [(1, 10)]),
+                       ("[a](x) ](b \";\")", [(1, 13)]),
+                       ("[a <b>](c \";\") d", []),
+                       ("[a\n](b \";\")", []),
+                       ("> [a\n> ](b \";\")", []),
+                       ("- [a\n  ](b \";\")", []),
+                       ("[a\n\n](b \";\")", [(3, 6)]),
+                       # an image can be in a link; a link in a link is text
+                       ("[![x](y)](z \";\")", []),
+                       ("![a [b](c) d](e \";\") f", []),
+                       ("[a [b](c) d](e \";\") f", [(1, 17)]),
+                       # with no destination, the quoted text is the destination: no title
+                       ("[]( \"\n;\")", [(2, 1)]),
+                       # an autolink or an inline tag starts first; an escaped "<" starts neither
+                       ("\\<aa:`>;`", []),
+                       ("<aa:[a](b> \";\")", [(1, 13)]),
+                       ("<span title=\"`\"> text; `", [(1, 22)]),
+                       ("<span title=\"\\``\"> text; `", [(1, 24)]),
+                       ("<b title='`'>`c; d`", []),
+                       # GitHub shows no attribute of a tag
+                       ("x <b a=\"`;`\">", []),
+                       ("<a href=\"](x ';')\"> a; b", [(1, 22)]),
+                       ("x <b a=\";<b a=\"</b> y", [(1, 9)]),
+                       # GitHub shows a tag that its tagfilter escapes as text
+                       ("x <title a=\"x;y\"> b", [(1, 14)]),
+                       # a comment, a processing instruction, a declaration and CDATA hold their brackets
+                       ("[a <!-- ] --> b](c \"d;\")", []),
+                       ("[x <!-- [a](b) --> y](c \"d;\")", []),
+                       ("[a <? ] ?> b](c \"d;\")", []),
+                       ("[a <!X ] > b](c \"d;\")", []),
+                       ("[a <![CDATA[ ] ]]> b](c \"d;\")", []),
+                       ("x <!-- <a b=\" --> c; <d e=\"> f", [(1, 20)]),
+                       ("x <!-- ](a;b) --> y", []),
+                       ("x <? a; b ?> y", []),
+                       ("x <? a> b; c ?> d", [(1, 10)]),
+                       ("x <![CDATA[ a> b; c ]]> d", [(1, 17)])):
+        findings, _ = lint(text)
+        assert [(f["line"], f["col"]) for f in findings if f["rule"] == "semicolon"] == cols, (text, findings)
+    findings, _ = lint("Press [stop](https://x.io); then wait. Use & ; here.")
+    assert [f["col"] for f in findings if f["rule"] == "semicolon"] == [27, 46], findings
+    # an indented code block is code, also in a quote and under a list item
+    def semicolon_lines(text):
+        return [f["line"] for f in lint(text)[0] if f["rule"] == "semicolon"]
+    assert semicolon_lines("Intro text.\n\n    x = 1; y = 2\n\tz; w") == []
+    assert semicolon_lines("> Intro.\n>\n>     x = 1; y = 2") == []
+    assert semicolon_lines("- Item.\n\n      x = 1; y = 2") == []
+    # it cannot interrupt a paragraph, and a list item's own indent is prose
+    assert semicolon_lines("Intro text\n    goes on; here.") == [2]
+    assert semicolon_lines("- Item.\n\n    More; text.") == [3]
+    assert semicolon_lines("1. Item.\n\n   - Sub.\n\n     More; text.") == [5]
+    # the list ends at a line with less indent
+    assert semicolon_lines("- Item.\n\nText.\n\n    x; y") == []
+    # a list item that follows a text line, or a nested item in a tight list, still
+    # sets the content indent, so its indented continuation paragraph is prose
+    assert semicolon_lines("Steps to follow:\n- Open the file.\n\n    Then save it; close it.\n") == [4]
+    assert semicolon_lines("- Open the file.\n  - Pick a row.\n\n      Then save it; close it.\n") == [4]
+    assert semicolon_lines("- Item\ncontinued\n\n    More; text.") == [4]
+    assert semicolon_lines("Intro text.\n\n    x = 1; y = 2\n") == []
+    assert semicolon_lines("Steps:\n- Open the file.\n\n      x = 1; y = 2\n") == []
+    # in a paragraph, only a bullet or "1." with text starts a list, so its indent
+    # does not apply after "2." or an empty bullet
+    assert semicolon_lines("Intro\n2. item\n\n    x = 1; y = 2") == []
+    assert semicolon_lines("Intro\n*\n\n    x = 1; y = 2") == []
+    assert semicolon_lines("Intro\n1. item\n\n    Then save it; close it.") == [4]
+    # after a quote, any number starts a list; a thematic break and a number of ten
+    # digits start none, and a very long number does not stop the run
+    assert semicolon_lines("> Intro\n2. item\n\n    Then save it; close it.") == [4]
+    assert semicolon_lines("Intro text.\n* * *\n    x = 1; y = 2") == []
+    assert semicolon_lines("Text\n* * *\n\n    x = 1; y = 2") == []
+    assert semicolon_lines("* * *\n  cont; `\n2. it; `") == [2]
+    assert semicolon_lines("1234567890. item\n\n    x = 1; y = 2") == []
+    assert semicolon_lines("Intro\n" + "1" * 5000 + ". digits; here.") == [2]
+    # a lazy line keeps the nested item open, so its paragraph after a blank line is prose
+    assert semicolon_lines("- outer\n  - nested\n  outer again\n\n      Save it; close it.") == [5]
+    # a code span may wrap across lines of one paragraph
+    assert semicolon_lines("Run `a;\nb; c` now; then") == [2]
+    assert semicolon_lines("Run ``a`;\nb;`` now.") == []
+    # a run that nothing closes in the paragraph is literal text
+    assert semicolon_lines("A ` mark; here.\n\nAnd ` there; too.") == [1, 3]
+    # a code span never crosses a block boundary: a new list item, a heading,
+    # a fence, a quote-only line, or a change of quote depth
+    assert semicolon_lines("- Press the ` key; it opens.\n- Run `x` now.\n") == [1]
+    assert semicolon_lines("Press the ` key; it opens.\n# Run `x` now\n") == [1]
+    assert semicolon_lines("Type ` to start; then stop.\n```\ncode\n```\nThe ` mark ends it.\n") == [1]
+    assert semicolon_lines("> A ` mark; here.\n>\n> And ` there; too.\n") == [1, 3]
+    assert semicolon_lines("# Title ` x\nText; with ` more\n") == [2]
+    assert semicolon_lines("A ` mark; here.\n---\nAnd ` there; too.") == [1, 3]
+    assert semicolon_lines("> A ` mark; here.\n> > And ` there; too.") == [1, 2]
+    # a span still wraps inside a list item and a quote, also with a double run
+    assert semicolon_lines("Run `a;\nb` now.\n") == []
+    assert semicolon_lines("- Run `a;\n  b` now.\n") == []
+    assert semicolon_lines("> Run `a;\n> b` now.\n") == []
+    assert semicolon_lines("Run ``a`;\nb;`` now.\n- Next `x;` y.") == []
+    # a line ends a span only where it ends the paragraph: "#" with no space is text,
+    # a lazy line with fewer ">" continues the quote, "2." cannot start a list in a
+    # paragraph, and inline HTML is text
+    for text in ("A `foo;\n#bar`; done.", "> Run `code;\ncontinued` text; here.",
+                 "Run `code;\n2. continued` text; here.",
+                 "Run `code;\n#hashtag continued` text; here.",
+                 "Run `code;\n<span>continued` text; here.",
+                 "> > Run `code;\n> lazy\n> > continued` text; here."):
+        assert semicolon_lines(text) == [text.count("\n") + 1], text
+    # a next item of the open list, a setext underline, and an HTML block end the span
+    assert semicolon_lines("1. Run `a;\n2. b` now; x") == [1, 2]
+    assert semicolon_lines("Run `a;\n===\nb` now; x") == [1, 3]
+    assert semicolon_lines("Run `a;\n<div>\n\nb` now; x") == [1, 4]
+    assert semicolon_lines("Run `a;\n<div>\nb` now; x") == [1, 3]
+    assert semicolon_lines("# Run `a;\nb` x;") == [1, 2]
+    # a lazy underline, a fence or an item that is indented code, and a tab after ">"
+    assert semicolon_lines("> Run `a;\n===\nb` x;") == [3]
+    assert semicolon_lines("- Run `a;\n===\nb` now; x") == [3]
+    assert semicolon_lines("1. Press the ` key; it opens.\n===\nUse `x` here.") == []
+    assert semicolon_lines("- item; `\n===\ntext; `") == [1]
+    # a thematic break is no lazy line: it ends the item, so deeper text after it is code
+    assert semicolon_lines("- Item; here.\n* * *\n    x = 1; y = 2") == [1]
+    assert semicolon_lines("Run `a;\n    ```\nb` x;") == []
+    assert semicolon_lines("Run `a;\n    - b` x;") == [2]
+    assert semicolon_lines("> Run `a;\n>\t- b` x;") == [1, 2]
+    # any list item after a quote ends the quote: it is no lazy line
+    assert semicolon_lines("> Press the ` key; it opens.\n2. Run `x` now; then stop.") == [1, 2]
+    assert semicolon_lines("> Run `a;\n*\nb` x;") == [1, 3]
+    assert semicolon_lines("> Quote text.\n- Press the ` key; it opens.\n> Run `x` now.") == [2]
+    assert semicolon_lines("> > t w1; `\n> - q w2; `\n> > t w3; `") == [1, 2, 3]
+    # each paragraph gets its own run index; a span in a quote uses the quote depth
+    assert semicolon_lines("Run `a;\nb` x\n\nRun `c;\nd` y") == []
+    assert semicolon_lines("> Run `a\n> b; c` now") == []
+    # one scan of the paragraph finds the last line of each run length
+    assert _paragraph_runs(["x ` y", "a `` b ` c", "", "e ` f"], 0, 0, 0) == (2, {1: 1, 2: 1})
+    assert _mask_code_spans("a ` b;", 0, lambda run: False) == ("a ` b;", 0)
+    assert _mask_code_spans("a ` b;", 0, lambda run: run == 1) == ("a     ", 1)
+    # per-file labels
+    findings, _ = lint("a; b", filename="x.md")
+    assert findings[0]["file"] == "x.md"
+    # inline code is masked, not deleted, so columns match the source
+    findings, _ = lint("Run `make all` then spin up the node.")
+    phrasal = [f for f in findings if f["rule"] == "phrasal-verb"]
+    assert len(phrasal) == 1 and phrasal[0]["col"] == 21, phrasal
+    # a sentence ends after closing quotes and brackets
+    twenty = " ".join(["word"] * 20)
+    findings, _ = lint(f'He asked "{twenty}?" Then {twenty} ended. (It {twenty}.) Next.')
+    assert not any(f["rule"] == "long-sentence" for f in findings), findings
+    # abbreviations do not end a sentence
+    findings, _ = lint(f"Use a tool, e.g. the {twenty} linter here now.")
+    assert any(f["rule"] == "long-sentence" for f in findings), findings
+    findings, _ = lint(f"Compare A vs. B in the {twenty} test.")
+    assert any(f["rule"] == "long-sentence" for f in findings), findings
+    # a sentence wrapped across lines is counted once, at its start
+    findings, _ = lint(f"Intro text here.\nThe {twenty}\nwrapped tail words go here.")
+    long_sentences = [f for f in findings if f["rule"] == "long-sentence"]
+    assert len(long_sentences) == 1, long_sentences
+    assert (long_sentences[0]["line"], long_sentences[0]["col"]) == (2, 1), long_sentences
+    findings, _ = lint(f"- Item {twenty}\n  continued in the same item.")
+    assert any(f["rule"] == "long-sentence" for f in findings), findings
+    # leading front matter is skipped except its description; line numbers after it stay true
+    findings, words_total = lint(f"---\nname: x\ndescription: {twenty} {twenty}.\n"
+                                 "tags:\n  - a and\n---\nThe job is robust.")
+    assert [(f["line"], f["rule"]) for f in findings] == [
+        (3, "long-sentence"), (7, "marketing-adjective")], findings
+    assert words_total == 44, words_total
+    # the description is one paragraph, in a quoted, plain or block scalar
+    for text, expected in (
+            ("---\nname: x\ndescription: Applies to `thor/` only; the rules are in a file.\n---\n\nText.\n",
+             [(3, 37, "semicolon")]),
+            ('---\ndescription: "Use it; then stop."\n---\n', [(2, 21, "semicolon")]),
+            ("---\ndescription: >-\n    Use it\n    robustly; then stop.\nname: x\n---\n",
+             [(4, 5, "marketing-adjective"), (4, 13, "semicolon")]),
+            ("---\ndescription: Use `a\n  b; c` now; stop.\n---\n", [(3, 12, "semicolon")]),
+            ("---\ndescription: ```\n---\n\nProse; here.\n", [(5, 6, "semicolon")]),
+            ("---\n# note\nname: x\n---\n", []),
+            ("---\nname: x\n# note; here\nsummary: a; b\n---\n", []),
+            # a comment before the first key, and a comment after the value
+            ("---\n# metadata\nname: robust-tool\ndescription: Read the file.\n---\n", []),
+            ("---\ndescription: Read the file. # robust; note\n---\n", []),
+            ("---\ndescription: \"Read the file.\" # robust; note\n---\n", []),
+            ("---\ndescription: Use it\n  # robust; note\n---\n", []),
+            ("---\ndescription: \"Use it\n  across; lines\" # robust\n---\n", [(3, 9, "semicolon")]),
+            # a link title that goes on in the next lines of the description
+            ("---\ndescription: See [a](b \"x\n  y`z\") `w; v` q.\n---\n", []),
+            ("---\ndescription: See [a](b \"x\n  y\n  z`w\") `p; q` r.\n---\n", []),
+            ("---\ndescription: See [a](b \"x\n  y\") c; d.\n---\n", [(3, 8, "semicolon")]),
+            ("---\ndescription: See [a](b \"x\n  y`z\") ``c ` d`` text; last `.\n---\n", [(3, 23, "semicolon")]),
+            ("---\ndescription: ]( ;)\n---\n", [(2, 17, "semicolon")]),
+            ("---\ndescription: See [a\n  ](b \";\") c; d.\n---\n", [(3, 13, "semicolon")]),
+            ("---\ndescription: \\```;``\n---\n", []),
+            # a "#" in quotes, in a block scalar, or inside a word is text
+            ("---\ndescription: \"Read # the; file.\"\n---\n", [(2, 25, "semicolon")]),
+            ("---\ndescription: 'It''s robust' # c\n---\n", [(2, 21, "marketing-adjective")]),
+            ("---\ndescription: >\n  a # robust; b\n---\n", [(3, 7, "marketing-adjective"), (3, 13, "semicolon")]),
+            ("---\ndescription: a#b; c\n---\n", [(2, 17, "semicolon")]),
+            # the key "description" in quotes is the same key
+            ("---\n\"description\": robust;\n---\n", [(2, 16, "marketing-adjective"), (2, 22, "semicolon")]),
+            ("---\n'description' : a; b\n---\n", [(2, 18, "semicolon")]),
+            ("---\n\"description\": |\n  a; b\n---\n", [(3, 4, "semicolon")]),
+            ("---\n\"descriptions\": a; b\n---\n", []),
+            # a later line that starts with the key can be in a quoted value of another key
+            ("---\ndescription: robust;\nexample: \"foo\ndescription: plain\"\n---\n",
+             [(2, 14, "marketing-adjective"), (2, 20, "semicolon")]),
+            # a comment after a block scalar has less indent than its text
+            ("---\ndescription: |\n    Use it.\n  # robust; note\n---\n", []),
+            # quoted keys, keys with spaces and keys with other characters are YAML
+            ("---\n\"robust-kit\": patch\n---\n", []),
+            ("---\nallowed-tools: Bash; Read\n---\n", []),
+            ("---\n$schema: robust; y\n---\n", []),
+            ("---\nrésumé: a; b\n---\n", [])):
+        findings, _ = lint(text)
+        assert [(f["line"], f["col"], f["rule"]) for f in findings] == expected, (text, findings)
+    # a block that is not YAML is Markdown: thematic breaks around prose
+    findings, _ = lint("---\n# Instructions\n\nStart; stop.\n\n- First.\n\n---\n")
+    assert [(f["line"], f["rule"]) for f in findings] == [(4, "semicolon")], findings
+    findings, _ = lint("---\nStart; stop.\n---\n")
+    assert [(f["line"], f["rule"]) for f in findings] == [(2, "semicolon")], findings
+    # a plain key with a space reads as prose, as in "Step 1: run it"
+    findings, _ = lint("---\nStep 1: run it; then: stop.\n---\n")
+    assert [(f["line"], f["rule"]) for f in findings] == [(2, "semicolon")], findings
+    # a block of comments only is Markdown: a heading between thematic breaks
+    findings, _ = lint("---\n# Note\n---\n\nText; here.\n")
+    assert [(f["line"], f["rule"]) for f in findings] == [(5, "semicolon")], findings
+    # a later --- pair is a thematic break, not front matter
+    findings, _ = lint(f"Intro.\n\n---\n{twenty} {twenty}.\n---")
+    assert any(f["rule"] == "long-sentence" for f in findings), findings
+    # blank lines, headings, rules, list items, and quotes end a prose block
+    for sep in ("\n\n", "\n# Heading\n", "\n---\n", "\n- ", "\n> "):
+        findings, _ = lint(f"First {twenty} line{sep}Second {twenty} line.")
+        assert not any(f["rule"] == "long-sentence" for f in findings), (sep, findings)
+    # a GitHub alert marker on the first line of a quote is not a word of the body
+    body = " ".join(["word"] * 25) + "."
+    for alert in ("> [!NOTE]", "> [!tip]", ">[!Warning]", ">  [!CAUTION]  "):
+        findings, _ = lint(f"{alert}\n> {body}")
+        assert not any(f["rule"] == "long-sentence" for f in findings), (alert, findings)
+    for text in (f"Text.\n\n> [!NOTE]\n> {body}", f"> Text.\n~~~\ncode\n~~~\n> [!NOTE]\n> {body}"):
+        findings, _ = lint(text)
+        assert not any(f["rule"] == "long-sentence" for f in findings), (text, findings)
+    # Inside HTML containers the marker is text: marker plus body is 26 words.
+    for opening, closing in (("<details>", "</details>"),
+                             ('<DIV class="note">', "</DIV>"),
+                             ("<section><details>", "</details></section>")):
+        findings, _ = lint(f"{opening}\n\n> [!NOTE]\n> {body}\n\n{closing}")
+        long_sentences = [f for f in findings if f["rule"] == "long-sentence"]
+        assert [f["match"] for f in long_sentences] == ["26 words"], (opening, findings)
+        findings, _ = lint(f"{opening}\n{closing}\n\n> [!NOTE]\n> {body}")
+        assert not any(f["rule"] == "long-sentence" for f in findings), (opening, findings)
+    findings, _ = lint(f"> [!NOTE]\n> {body}")
+    assert not any(f["rule"] == "long-sentence" for f in findings), findings
+    # An inner close or a fenced closing tag must not close the outer container.
+    for content in ("<details>\n<div>\n</div>",
+                    "<details>\n\n~~~\n</details>\n~~~"):
+        findings, _ = lint(f"{content}\n\n> [!NOTE]\n> {body}\n\n</details>")
+        long_sentences = [f for f in findings if f["rule"] == "long-sentence"]
+        assert [f["match"] for f in long_sentences] == ["26 words"], (content, findings)
+    # Directly after <details>, ~~~ is HTML text, not a fence. So the </details> line after
+    # it closes the container, and the alert after it is outside, as GitHub shows it.
+    findings, _ = lint(f"<details>\n~~~\n</details>\n~~~\n\n> [!NOTE]\n> {body}\n\n</details>")
+    assert not any(f["rule"] == "long-sentence" for f in findings), findings
+    # Fenced, commented, inline, void, and self-closing tags do not open containers.
+    for content in ("```html\n<details>\n```", "<!-- <details> -->",
+                    "<span>Text.</span>", "<hr>", "<div />",
+                    "<details></details>", "</details>"):
+        findings, _ = lint(f"{content}\n\n> [!NOTE]\n> {body}")
+        assert not any(f["rule"] == "long-sentence" for f in findings), (content, findings)
+    # GitHub shows these markers as text, so they stay words. GitHub shows the
+    # last two as alerts, but no blank line or fence proves that a quote starts.
+    for text in (f"> [!FOO]\n> {body}", f"> [!NOTE] {body}", f"> Intro.\n> [!NOTE]\n> {body}",
+                 f"> > [!NOTE]\n> > {body}", f"> Intro.\ncontinuation.\n> [!NOTE]\n> {body}",
+                 f"- Item.\n\n  > [!NOTE]\n  > {body}",
+                 f"Text.\n> [!NOTE]\n> {body}", f"   > [!NOTE]\n   > {body}"):
+        findings, _ = lint(text)
+        long_sentences = [f for f in findings if f["rule"] == "long-sentence"]
+        assert [f["match"] for f in long_sentences] == ["26 words"], (text, long_sentences)
+    # a bold or italic lead sentence ends at its closing emphasis marker
+    findings, _ = lint(f"**Lead.** The {twenty} more words here now.")
+    assert not any(f["rule"] == "long-sentence" for f in findings), findings
+    findings, _ = lint(f"_Note._ The {twenty} more words here now.")
+    assert not any(f["rule"] == "long-sentence" for f in findings), findings
+    findings, _ = lint(f"See **e.g.** the {twenty} linter here now.")
+    assert any(f["rule"] == "long-sentence" for f in findings), findings
+    # a dotted abbreviation ends a sentence before a capitalized word
+    findings, _ = lint(f"The team shipped it in the U.S. Then {twenty} today.")
+    assert not any(f["rule"] == "long-sentence" for f in findings), findings
+    # nested list items, quoted list items, and HTML lines end a prose block
+    for text in (f"- First item\n    - First {twenty} line\n      - Second {twenty} line",
+                 f"> - First {twenty} line\n> - Second {twenty} line",
+                 f"<!-- First {twenty} line -->\nSecond {twenty} line.",
+                 f"> First {twenty} line.\n>\n> Second {twenty} line."):
+        findings, _ = lint(text)
+        assert not any(f["rule"] == "long-sentence" for f in findings), (text, findings)
+    # a blockquote prefix is not a word in a wrapped quoted sentence
+    findings, _ = lint("> " + "\n> ".join(["word word word word word"] * 5) + ".")
+    assert not any(f["rule"] == "long-sentence" for f in findings), findings
+    # a list marker is not a word either, in every marker style
+    twenty_five = " ".join(f"term{number}" for number in range(1, 26)) + "."
+    for marker in ("-", "*", "+", "1.", "1)", "  -", "> -"):
+        findings, _ = lint(f"{marker} {twenty_five}")
+        assert not any(f["rule"] == "long-sentence" for f in findings), (marker, findings)
+    findings, _ = lint(f"- term0 {twenty_five}")
+    long_sentences = [f for f in findings if f["rule"] == "long-sentence"]
+    assert [(f["match"], f["col"]) for f in long_sentences] == [("26 words", 3)], long_sentences
+    # a code span at the line start does not turn the next token into a list marker
+    findings, _ = lint(f"`label` 1) {twenty_five}")
+    long_sentences = [f for f in findings if f["rule"] == "long-sentence"]
+    assert [f["match"] for f in long_sentences] == ["26 words"], long_sentences
+    # bare punctuation is not a word: code-span separators and spaced dashes
+    for separator in (", ", " ", " / "):
+        names = separator.join(f"`a{number}`" for number in range(30))
+        findings, _ = lint(f"The adapters are {names}.")
+        assert not any(f["rule"] == "long-sentence" for f in findings), (separator, findings)
+    findings, _ = lint(" ".join(f"term{number}" for number in range(1, 25)) + " — term25.")
+    assert not any(f["rule"] == "long-sentence" for f in findings), findings
+    findings, _ = lint(f"Use `make` — {twenty_five}")
+    long_sentences = [f for f in findings if f["rule"] == "long-sentence"]
+    assert [f["match"] for f in long_sentences] == ["26 words"], long_sentences
+    # a long block stays linear: one short sentence per line, no blank lines
+    findings, _ = lint("The agent read the log.\n" * 20000)
+    assert not any(f["rule"] == "long-sentence" for f in findings)
+    # a long line with many code spans, and many links that go on in the next line, stay linear
+    findings, _ = lint("`x` " * 60000 + "c; d.")
+    assert [(f["rule"], f["col"]) for f in findings] == [("semicolon", 240002)], findings
+    findings, _ = lint("[a](x;y\n)\n" * 5000)
+    assert not any(f["rule"] == "semicolon" for f in findings), findings
+    # flags may appear between file paths
+    with tempfile.TemporaryDirectory() as tmp:
+        paths = [os.path.join(tmp, name) for name in ("a.md", "b.md")]
+        for path in paths:
+            with open(path, "w", encoding="utf-8") as fh:
+                fh.write("Short line here.\n")
+        out = io.StringIO()
+        with contextlib.redirect_stdout(out):
+            assert main([paths[0], "--json", paths[1]]) == 0
+        assert json.loads(out.getvalue())["words"] == 6, out.getvalue()
+        # repeated --disable flags add up instead of replacing each other
+        with open(paths[0], "w", encoding="utf-8") as fh:
+            fh.write("a; b. " + " ".join(["word"] * 30) + ".\n")
+        out = io.StringIO()
+        with contextlib.redirect_stdout(out):
+            assert main([paths[0], "--json", "--disable", "semicolon",
+                         "--disable", "long-sentence"]) == 0
+        assert json.loads(out.getvalue())["count"] == 0, out.getvalue()
+    # command-line parsing rejects typos and bad values
+    for bad in (["--basline", "3"], ["--baseline"], ["--baseline", "x"],
+                ["--baseline", "-1"], ["--disable", "passive"], ["--base", "3"]):
+        try:
+            with contextlib.redirect_stderr(io.StringIO()):
+                main(bad)
+        except SystemExit as e:
+            assert e.code == 2, (bad, e.code)
+        else:
+            raise AssertionError(bad)
+    assert _rule_list("passive-voice,present-perfect") == {"passive-voice", "present-perfect"}
+    # an HTML block in a block quote holds text, and it ends with its quote
+    findings, _ = lint("> <div>\n> ```\n> This is prose; keep it.\n> ```\n> </div>")
+    assert [(f["rule"], f["line"]) for f in findings] == [("semicolon", 3)], findings
+    findings, _ = lint("> <div>\n\n```\nx; y\n```\n\nProse; here.")
+    assert [(f["rule"], f["line"]) for f in findings] == [("semicolon", 7)], findings
+    # a ">" line in an open HTML block is its text, not a quote end
+    findings, _ = lint("<div>\n> Note.\n>\n> ```\n> a; b\n> ```\n</div>")
+    assert [(f["rule"], f["line"]) for f in findings] == [("semicolon", 5)], findings
+    findings, _ = lint("<div>\n>\n```\nText; here.\n```")
+    assert [(f["rule"], f["line"]) for f in findings] == [("semicolon", 4)], findings
+    # a quote after a paragraph starts a block quote, so a type 7 tag in it starts a block
+    findings, _ = lint("Paragraph.\n> <widget>\n> ```\n> Text; here.\n> ```")
+    assert [(f["rule"], f["line"]) for f in findings] == [("semicolon", 4)], findings
+    print("selftest OK")
+
+
+RULE_IDS = {rule_id for rule_id, *_ in RULES} | {
+    "long-sentence", "synonym-rotation", "dangling-conjunction"}
+
+
+def _rule_list(value):
+    rules = {r for r in value.split(",") if r}
+    unknown = rules - RULE_IDS
+    if unknown:
+        raise argparse.ArgumentTypeError(
+            f"unknown rule(s): {', '.join(sorted(unknown))}. "
+            f"Known rules: {', '.join(sorted(RULE_IDS))}")
+    return rules
+
+
+def _non_negative_int(value):
+    try:
+        n = int(value)
+    except ValueError:
+        n = -1
+    if n < 0:
+        raise argparse.ArgumentTypeError(f"expected a non-negative integer, got {value!r}")
+    return n
+
+
+def main(argv):
+    parser = argparse.ArgumentParser(
+        prog="ste-lint.py", allow_abbrev=False,
+        description="Deterministic linter for the structural STE rules. "
+                    "Reads stdin when no FILE is given.")
+    parser.add_argument("paths", nargs="*", metavar="FILE")
+    parser.add_argument("--json", action="store_true", help="print findings as JSON")
+    parser.add_argument("--baseline", type=_non_negative_int, default=0, metavar="N",
+                        help="pass unless hard violations exceed N (default 0)")
+    parser.add_argument("--disable", type=_rule_list, action="append", default=[],
+                        metavar="RULES",
+                        help="comma-separated rule IDs to silence; may be repeated")
+    parser.add_argument("--selftest", action="store_true", help="run the built-in tests")
+    # intermixed parsing keeps flags valid between file paths
+    args = parser.parse_intermixed_args(argv)
+    if args.selftest:
+        selftest()
+        return 0
+    as_json, baseline, paths = args.json, args.baseline, args.paths
+    disabled = set().union(*args.disable)
+
+    findings, words_total = [], 0
+    if paths:
+        for p in paths:
+            try:
+                with open(p, encoding="utf-8") as fh:
+                    text = fh.read()
+            except OSError as e:
+                parser.error(f"cannot read {p}: {e.strerror}")
+            f, w = lint(text, filename=p)
+            findings.extend(f)
+            words_total += w
+    else:
+        findings, words_total = lint(sys.stdin.read())
+
+    findings = [f for f in findings if f["rule"] not in disabled]
+    hard_count = sum(1 for f in findings if f["level"] == "advisory-free")
+    report(findings, words_total, as_json, hard_count, baseline)
+    return 1 if hard_count > baseline else 0
+
+
+if __name__ == "__main__":
+    sys.exit(main(sys.argv[1:]))