iter 7: prompt rules 27-29 driven by real-draft testing

First exercise of the tool against unfamiliar real-world drafts (user's
in-progress academic papers, not derived from CMOS quick guide
examples). Found three systemic patterns affecting ~14% of entries
across 134-entry and 13-entry drafts; this commit addresses all three.

Prompt additions:

 27. GOVERNMENT DOCUMENTS / INSTITUTIONAL REPORTS — when the author is
     a government body (GAO, US Census, WHO, Pew, etc.) and the source
     is a standalone report, italicize the title like a book; do NOT
     wrap it in quotes. CMOS 14.272.

 28. PRESERVE DELIBERATE LOWERCASING of proper nouns. Examples cited in
     the prompt: the journal "portal: Libraries and the Academy"
     (deliberately lowercase "p"), authors bell hooks / danah boyd /
     e e cummings, brand names iPhone / eBay. Headline-case
     normalization must NOT touch these.

Rule 17 also extended (was newspaper/magazine only):

 17. PERIODICAL NAMES — drop leading "The" from ANY periodical:
     newspapers, magazines, AND scholarly journals. Write "Journal of
     Academic Librarianship", not "The Journal of Academic
     Librarianship"; "Library Quarterly", not "The Library Quarterly";
     "American Archivist", not "The American Archivist". CMOS 14.191.
     Rule explicitly excepts BOOK titles and report titles, which keep
     their leading "The" (e.g., *The Library's Guide to Sexual and
     Reproductive Health Information*).

Empirical impact, Reading_Disrepair (13 entries):
  before: 3 errors (2 leading-The, 1 GAO format, 1 portal capitalized)
  after:  0 errors

Empirical impact, HML Submission (134 entries):
  before: ~21 errors (18 leading-The, 0 GAO in this draft, 3 portal)
  after:  ~1 error (a multi-author 2nd-author inversion bug newly
          surfaced — separate fix)

Also adds scripts/analyze_draft_output.py — a triage helper that
diff-walks input and output, surfacing common failure patterns
(leading-The, doi: prefix, bare-hyphen ranges, gov-report quoted, and
intentional lowercasing stripped). False positives are tolerated since
it's a human-review tool, not a validator.

Exemplar corpus is unchanged. Re-baseline against 14 canary exemplars
under the new rules: scalar 1.000, canary 1.000 (no regressions).

User's real drafts under rough_drafts/ remain untracked — they are
in-progress academic work and don't belong in git history.

scripts/analyze_draft_output.py

@@ -0,0 +1,183 @@
+"""Analyze a cmos-formatted draft output against its input draft.
+
+Usage:
+    uv run python scripts/analyze_draft_output.py INPUT.md OUTPUT.md
+
+Extracts the bibliography sections from both files, pairs them line-by-
+line, and surfaces common failure patterns:
+
+- "The " leading article preserved in periodical names
+- "doi:" prefix not converted
+- Bare ASCII hyphen in page ranges (2+ digits - 2+ digits)
+- Quoted titles that look like they should be italicized (book/report)
+- Entries that are byte-identical (no change applied)
+- Entries with large length deltas (potential truncation or runaway)
+
+Output is a summary plus a per-entry table. Not a substitute for human
+review but cheap triage on long bibliographies.
+"""
+
+from __future__ import annotations
+
+import argparse
+import re
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+
+from cmos.parser import NoBibliographyError, split_bibliography  # noqa: E402
+
+
+def entries_from(path: Path) -> list[str]:
+    text = path.read_text(encoding="utf-8")
+    try:
+        return split_bibliography(text).entries
+    except NoBibliographyError:
+        print(f"no ## Bibliography section in {path}", file=sys.stderr)
+        return []
+
+
+# --- Failure-mode detectors ---------------------------------------------------
+
+_LEADING_THE_PERIODICAL = re.compile(r"\*(The [A-Z][^*]+)\*")
+_DOI_PREFIX = re.compile(r"\bdoi:\s*10\.", re.IGNORECASE)
+_BARE_HYPHEN_RANGE = re.compile(r"(?<!\w)(\d{2,4})-(\d{2,4})(?!\w)")
+_GOV_AUTHOR_MARKERS = (
+    "Accountability Office",
+    "Department of ",
+    "Bureau of ",
+    "Government of ",
+    "U.S. ",
+    "United States ",
+    "Census Bureau",
+    "National Institute",
+    "Congressional Research Service",
+)
+
+
+def check_leading_the(entry: str) -> str | None:
+    m = _LEADING_THE_PERIODICAL.search(entry)
+    if m:
+        return f"leading 'The' in italicized periodical: '*{m.group(1)}*'"
+    return None
+
+
+def check_doi_prefix(entry: str) -> str | None:
+    if _DOI_PREFIX.search(entry):
+        return "has 'doi:' prefix instead of https://doi.org/"
+    return None
+
+
+def check_bare_hyphen_range(entry: str) -> str | None:
+    # Allow en-dash (which we want) and hyphenated year ranges inside titles
+    # are sometimes legitimate, but inside a colon-prefixed page range they
+    # should be en-dash. Just flag all and let a human decide.
+    m = _BARE_HYPHEN_RANGE.search(entry)
+    if m:
+        return f"bare hyphen in number range: '{m.group(0)}'"
+    return None
+
+
+def check_gov_report_quoted(entry: str) -> str | None:
+    # Only flag entries whose author field is clearly a government body AND
+    # the title is wrapped in quotes (rather than italicized).
+    has_gov_author = any(marker in entry for marker in _GOV_AUTHOR_MARKERS)
+    if not has_gov_author:
+        return None
+    # Title in quotes (not italicized) on a gov-authored entry = likely wrong.
+    if re.search(r'"[^"]{20,}"', entry):
+        return "government-body author with quoted title; should be italicized like a book"
+    return None
+
+
+def check_intentional_lowercase_stripped(entry: str, source_entry: str | None) -> str | None:
+    # If the source entry had an all-lowercase word-token that appears
+    # title-cased in the output, flag it as potentially wrongful normalization.
+    if source_entry is None:
+        return None
+    # Find all-lowercase alphabetic tokens in the source that would normally
+    # be capitalized in a proper noun. Skip common short words.
+    skip = {
+        "a", "an", "and", "as", "at", "but", "by", "for", "if", "in", "nor",
+        "of", "on", "or", "so", "the", "to", "up", "via", "vs", "yet", "no",
+        "not", "with", "from", "into", "over", "per", "than",
+    }
+    lowered_tokens = [
+        t for t in re.findall(r"\b[a-z]{4,}\b", source_entry) if t not in skip
+    ]
+    for tok in lowered_tokens:
+        # Is the same token title-cased in the output?
+        if tok[0].lower() + tok[1:] not in entry and tok.capitalize() in entry:
+            return f"source had '{tok}' (lowercase); output may have normalized to '{tok.capitalize()}'"
+    return None
+
+
+DETECTORS = [
+    ("leading_the", check_leading_the),
+    ("doi_prefix", check_doi_prefix),
+    ("bare_hyphen_range", check_bare_hyphen_range),
+    ("gov_report_quoted", check_gov_report_quoted),
+]
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("input", type=Path)
+    ap.add_argument("output", type=Path)
+    ap.add_argument("--show", type=int, default=5, help="entries to show per bucket")
+    args = ap.parse_args()
+
+    inputs = entries_from(args.input)
+    outputs = entries_from(args.output)
+    if not inputs or not outputs:
+        return 2
+
+    print(f"input entries:  {len(inputs)}")
+    print(f"output entries: {len(outputs)}")
+    if len(inputs) != len(outputs):
+        print(f"!! count mismatch — pairing by index will misalign")
+    print()
+
+    # Count issues across all output entries.
+    issue_counts: dict[str, list[tuple[int, str, str]]] = {name: [] for name, _ in DETECTORS}
+    issue_counts["intentional_lowercase_stripped"] = []
+    unchanged = []
+    for i, out in enumerate(outputs):
+        for name, detector in DETECTORS:
+            msg = detector(out)
+            if msg:
+                issue_counts[name].append((i, msg, out))
+        src = inputs[i] if i < len(inputs) else None
+        msg = check_intentional_lowercase_stripped(out, src)
+        if msg:
+            issue_counts["intentional_lowercase_stripped"].append((i, msg, out))
+        if src is not None and out.strip() == src.strip():
+            unchanged.append((i, out))
+
+    print("=== Failure-mode summary ===")
+    for name, hits in issue_counts.items():
+        print(f"  {name:<30} {len(hits)}")
+    print(f"  {'byte_identical':<30} {len(unchanged)}")
+    print()
+
+    for name, hits in issue_counts.items():
+        if not hits:
+            continue
+        print(f"--- {name} (first {min(args.show, len(hits))} of {len(hits)}) ---")
+        for idx, msg, entry in hits[: args.show]:
+            print(f"  [{idx}] {msg}")
+            print(f"       {entry[:140]}{'...' if len(entry) > 140 else ''}")
+        print()
+
+    if unchanged:
+        print(f"--- byte_identical (first {min(args.show, len(unchanged))} of {len(unchanged)}) ---")
+        for idx, entry in unchanged[: args.show]:
+            print(f"  [{idx}] {entry[:140]}")
+        print()
+
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

src/cmos/formatter.py

@@ -109,16 +109,31 @@ Rules you must follow (CMOS 18th edition specifically):
     provides them (not initials), e.g., "Snyder, Carl D., Manuel
     Bedrossian, Casey Barr, et al." — NOT "Snyder, C. D., ..."
 
-17. NEWSPAPER AND MAGAZINE NAMES:
-    - Expand common abbreviations. "NYT" → "New York Times". "WSJ" →
-      "Wall Street Journal". "WaPo" → "Washington Post". "LAT" →
-      "Los Angeles Times".
-    - Drop a leading "The" from the publication name. Write "New Yorker",
-      NOT "The New Yorker". Write "New York Times", NOT "The New York
-      Times". Write "Atlantic", NOT "The Atlantic".
-    - Italicize newspaper and magazine names: *New Yorker*.
-    - Use a comma (not a period) between the italicized name and the date:
-      *New Yorker*, December 18, 2023.
+17. PERIODICAL NAMES (newspapers, magazines, AND scholarly journals):
+    - Expand common newspaper abbreviations. "NYT" → "New York Times".
+      "WSJ" → "Wall Street Journal". "WaPo" → "Washington Post". "LAT"
+      → "Los Angeles Times".
+    - Drop a leading "The" from ANY periodical name — newspapers,
+      magazines, AND scholarly journals. Write "New Yorker", NOT "The
+      New Yorker". Write "New York Times", NOT "The New York Times".
+      Write "Atlantic", NOT "The Atlantic". Write "Journal of Academic
+      Librarianship", NOT "The Journal of Academic Librarianship". Write
+      "Library Quarterly", NOT "The Library Quarterly". Write "American
+      Archivist", NOT "The American Archivist". This is a CMOS 14.191
+      convention for all periodicals.
+    - IMPORTANT EXCEPTION: this "drop leading The" rule applies ONLY to
+      periodical names. Do NOT drop the leading "The" from BOOK titles,
+      REPORT titles, or any other non-periodical title. Book titles and
+      report titles keep their "The" exactly as given. For example,
+      *The Library's Guide to Sexual and Reproductive Health
+      Information* (a book) keeps its "The"; *The Free Press of
+      Glencoe* (a publisher name that begins with "The") keeps its
+      "The".
+    - Italicize newspaper, magazine, and journal names.
+    - Between an italicized newspaper or magazine name and a full date,
+      use a comma: *New Yorker*, December 18, 2023. (Scholarly journals
+      use the volume/issue/year-in-parens form instead, and do NOT get
+      a comma before the parenthetical.)
 
 18. E-BOOK FORMAT INDICATOR: write just the platform name with a period.
     "Kindle." NOT "Kindle edition." "Nook." NOT "Nook e-book."
@@ -180,6 +195,37 @@ Rules you must follow (CMOS 18th edition specifically):
     title and the publisher:
     Lastname, First. *Title*. 2nd ed. Publisher, Year.
 
+27. GOVERNMENT DOCUMENTS AND INSTITUTIONAL REPORTS: when the author is a
+    government body, agency, institution, or NGO (e.g., "Government
+    Accountability Office", "U.S. Department of Education", "Congressional
+    Research Service", "World Health Organization", "Pew Research
+    Center") and the source is a STANDALONE report or document (not an
+    article in a periodical), treat the report like a BOOK: italicize
+    the title with Markdown *...*, do NOT wrap it in quotes. Include any
+    report number as an identifier after the title. Form:
+    Author-Body. *Title of Report*. Report Number. Publisher (often the
+    same body), Date. URL.
+    Example: Government Accountability Office. *Public Libraries: Many
+    Buildings Are Reported to Be in Poor Condition*. GAO-26-107262.
+    December 18, 2025. https://files.gao.gov/reports/GAO-26-107262/index.html.
+    This is CMOS 14.272.
+
+28. PRESERVE DELIBERATE LOWERCASING of proper nouns. Some author names,
+    journal names, organization names, and brand names are intentionally
+    all-lowercase as a stylistic or branding choice and MUST be preserved
+    as-is. Known cases include:
+    - The scholarly journal "portal: Libraries and the Academy" — always
+      written with a lowercase "p". Do NOT title-case it to
+      "Portal: Libraries and the Academy".
+    - Authors: "bell hooks", "danah boyd", "e e cummings", "k.d. lang".
+    - Brand / product names: "iPhone", "eBay" — preserve the exact
+      capitalization shown on their own sites.
+    If the source has a proper noun in ALL LOWERCASE and you might
+    otherwise normalize it to headline case, STOP and leave it
+    lowercase. This is a deliberate authorial/brand choice, not sloppy
+    capitalization. When in doubt about whether a lowercased form is
+    intentional, err on the side of preserving the source's form.
+
 Output ONLY the single reformatted entry. No preamble, no explanation.
 """