bmad-module-skill-forge 2.1.0 → 3.0.0

package/src/shared/scripts/skf-comention-pairs.py

@@ -1,40 +1,101 @@
 #!/usr/bin/env python3
 # /// script
-# requires-python = ">=3.10"
+# requires-python = ">=3.11"
 # dependencies = []
 # ///
-"""SKF Co-mention Pairs — compute compose-mode co-mention pairs from an
-architecture document with the same precision guards the prompt used to run
-by hand.
-
-`detect-integrations.md` §2 (compose branch) asks the LLM to scan an
-architecture markdown document for skill-name co-mentions using three
-deterministic guards:
-
-  1. Word-boundary matching (`\\b{skill_name}\\b`, case-insensitive) — no
-     substring false positives (`react` must not match inside `reactive`).
-  2. Section filtering — paragraphs governed by an H1/H2 header that
-     normalises to one of {introduction, overview, glossary, table of
-     contents, references, appendix, index} are excluded, and heading text
-     itself is never a co-mention source.
-  3. Two-paragraph minimum — a pair (A, B) qualifies only when at least two
-     distinct body paragraphs each contain word-boundary matches for both A
-     and B.
-
-Given a fixed architecture doc and skill-name set the qualifying-pair set is
-fully determined, so this belongs in a deterministic helper rather than
-in-prompt regex/counting. It mirrors the co-import subprocess the code-mode
-branch already delegates and the `skf-pair-intersect.py` pre-pass. The LLM
-still does the semantic composition (integration-section authoring, tier
-inheritance, VS-verdict parsing) — the script only emits the qualifying pairs
-and their evidence paragraphs.
-
-Subcommand:
+"""SKF Co-mention Pairs: which loaded skills a document names, and which
+pairs of them it names together.
+
+The script reports evidence and the calling step judges it: every pair it
+emits is a candidate, never a confirmed integration. Its three subcommands
+share one skills input, one matcher and one Markdown reader:
+
+  comention  compose-mode co-mention pairs from an architecture document
+             (create-stack-skill detect-integrations.md §2, compose branch)
+  mentions   the skills a document names and the pairs it names together
+             (refine-architecture gap-analysis.md scope sets and documented
+             pairs)
+  infer      candidate pairs when there is no architecture document: one
+             skill's own docs naming the other, or shared domain keywords,
+             never a shared language alone
+
+Skills input (--skills):
+  A JSON array. Each entry is a skill name (`"react"`) or an object:
+    {"name": "<skill>",            required, non-empty
+     "aliases": ["<term>", ...],   other names the skill goes by
+     "keywords": ["<kw>", ...],    domain keywords (read by infer only)
+     "language": "<language>",     the skill's language, an array of them
+                                   (a stack lists its libraries'
+                                   languages) or null (infer only)
+     "docs": ["<path>", ...]}      the skill's own docs (infer only)
+  Other keys are ignored, so an inventory entry, such as a `skills[]` entry
+  of skf-enumerate-stack-skills.py, can be passed as it is. Entries with
+  the same name merge, languages included. Output names a skill by its
+  `name`, never an alias.
+
+Matching:
+  A skill is named where one of its terms (its name and each alias)
+  appears, case-insensitively, with no word character right before or
+  after it: `react` never matches inside `reactive`. For a term that starts
+  and ends with a word character, as every skill name does, this is the
+  `\\b{term}\\b` rule; a term such as `C++` or `.NET` matches too. The
+  words of a multi-word alias may be split by any whitespace, a line break
+  included. comention matches each skill on its own, as it always has, so
+  `react` also matches the `react` in `react-dom`. mentions and infer read
+  each occurrence once, as the longest term that matches there: when the
+  skills include both `react` and `react-dom`, the text `react-dom` names
+  `react-dom` only. Two skills that share a term are both named by it.
+
+Paragraphs, units and evidence kinds:
+  A body paragraph is a run of consecutive non-blank lines that are not
+  ATX headers (1-6 `#` then a space or tab); headers of any level end it,
+  and the nearest H1/H2 above it governs it. A paragraph splits into
+  units: a run of prose, one list item with its continuation lines, one
+  table row, or one line of fenced code. A list item or table row inside a
+  blockquote (`> - item`) is a unit the same way. Each evidence entry is
+  one body paragraph that names both skills of a pair, and its `kind` says
+  how:
+    co-mention  one unit of the paragraph names both skills
+    lead-in     no unit does, but a run of prose names one skill and a list
+                item, table row or code line under it (before the next run
+                of prose) names the other: "The express server mounts:"
+                then "- passport"
+    list-only   neither: the two are named only in separate units, such as
+                separate list items, table rows or code lines
+  A "Tech Stack" table and a "Components" list that each name every
+  library give every pair of them two list-only paragraphs. A lead-in
+  counts only inside its paragraph: with a blank line before the list, the
+  lead-in and the items are two paragraphs. The units are a heuristic, not
+  a Markdown parser: consecutive "Label: value" lines with no list marker
+  are one run of prose, so they give co-mentions, and a line that starts
+  with a number and a period (`2024.`) right after a bullet item is a new
+  item, where CommonMark reads a continuation.
+
+  Each entry also shows where the paragraph names the pair: `unit_excerpt`
+  is the collapsed text of the first unit that names both (a co-mention),
+  or of the lead-in and the item (a lead-in, with `...` for any items
+  between them), cut to 240 characters around the two names; when the
+  words between the names do not fit, the middle is cut instead, so both
+  names stay. `...` marks every cut. `unit_line` is the line (1-based) of
+  the first of the two names. Both are null for a list-only entry, while
+  `excerpt` and `line` are always the paragraph's start and first line.
+
+Subcommands:
   comention --doc <md-file-or-'-'> --skills <json-file-or-'-'>
-      --doc     path to the architecture markdown document, or '-' for stdin.
-      --skills  path to a JSON array of loaded skill names
-                (e.g. `["react", "express"]`), or '-' for stdin.
-                (--doc and --skills cannot both read from stdin.)
+      Emits every pair that one body paragraph or more names in one unit
+      or through a lead-in, by the matching above; heading text is never a
+      source. A pair only ever listed together is left out and counted, as
+      mentions does: that is a structural kind of evidence, and nothing is
+      filtered on meaning. Whether a passage describes an integration is the
+      calling step's judgment, from its excerpt, and two fields carry what
+      it weighs. `excluded_section` is true on an evidence entry whose
+      paragraph an H1/H2 header governs that normalises to one of
+      {introduction, overview, glossary, table of contents, references,
+      appendix, index}, sections that often only list the libraries;
+      `paragraph_count` counts the body paragraphs that name the pair,
+      one being enough. Fenced code is read as body text here, as this
+      mode always read it, so a diagram line names a pair like any other
+      line, and a `#` line inside a fence still counts as a header.
 
       Emit JSON:
         {
@@ -42,45 +103,198 @@ Subcommand:
             {
               "a": "<skill>", "b": "<skill>",
               "paragraph_count": N,
+              "comention_count": C,
+              "lead_in_count": I,
               "evidence": [{"header": "<governing-header-or-null>",
-                            "excerpt": "<collapsed paragraph text>"}, ...]
+                            "excerpt": "<collapsed paragraph text>",
+                            "line": L,
+                            "kind": "co-mention|lead-in|list-only",
+                            "unit_line": U-or-null,
+                            "unit_excerpt": "<text-naming-both>-or-null",
+                            "excluded_section": true|false},
+                           ...]
             },
             ...
           ],
+          "list_only_pair_count": P,
           "excluded_section_count": M
         }
 
-      Only unordered pairs (a < b lexicographically) with paragraph_count >= 2
-      are emitted. `pairs[]` is sorted by paragraph_count DESC, then (a, b)
-      ASC — stable, mirroring skf-pair-intersect.py. `evidence[]` is in
-      document order. `excluded_section_count` is the number of DISTINCT
-      excluded H1/H2 sections (by normalised header) that governed at least
-      one body paragraph — a diagnostic for the caller, not a gate.
+      Only unordered pairs (a < b lexicographically) are emitted, each a
+      candidate with comention_count + lead_in_count >= 1 (its co-mention
+      and lead-in entries; a list-only entry from another paragraph stays
+      in its evidence). `list_only_pair_count` counts the pairs only ever
+      listed together, which are not emitted. `pairs[]` is sorted by
+      paragraph_count DESC, then (a, b) ASC, mirroring
+      skf-pair-intersect.py. `evidence[]` is in document order.
+      `excluded_section_count` is the number of DISTINCT such introductory
+      H1/H2 sections (by normalised header) that governed at least one
+      body paragraph: a diagnostic for the caller.
+
+  mentions --doc <md-file-or-'-'> --skills <json-file-or-'-'>
+           [--technologies <json-file-or-'-'>]
+      Fenced code blocks are skipped: their lines are neither paragraphs
+      nor headings. No section is excluded, and one paragraph is enough.
+
+      Emit JSON:
+        {
+          "skills": [
+            {"name": "<skill>",
+             "paragraph_count": N, "heading_count": H, "fenced_count": F,
+             "terms": [{"term": "<the skill's term>",
+                        "kind": "name|alias",
+                        "paragraph_count": N, "heading_count": H,
+                        "fenced_count": F}, ...],
+             "paragraphs": [{"header": "<governing-header-or-null>",
+                             "excerpt": "<collapsed paragraph text>",
+                             "line": L}, ...]},
+            ...
+          ],
+          "mentioned": ["<skill>", ...],
+          "fenced_only": ["<skill>", ...],
+          "unmentioned": ["<skill>", ...],
+          "candidates": [<a pair shaped as a comention pair, its
+                          evidence without `excluded_section`>, ...],
+          "list_only_pair_count": P,
+          "fenced_block_count": B,
+          "fenced_blocks": [{"line": L, "info": "<info string>"}, ...]
+        }
+
+      `skills[]` has one entry per skill, sorted by name: how many body
+      paragraphs, headings (any level) and fenced blocks name it, and each
+      such paragraph in document order. `terms[]` lists each of the
+      skill's terms that named it, spelled as the skills input gives it,
+      its name first and then its aliases in input order, with `kind`
+      saying which, and how many paragraphs, headings and fenced blocks
+      that term named the skill in: the term a scope gate shows for each
+      skill, so a skill named only through an alias reads as such.
+      `mentioned` lists the skills a paragraph or a heading names,
+      `fenced_only` the skills named only inside fenced code (a diagram,
+      for example) and `unmentioned` the rest; the three are sorted and
+      hold every skill once.
+
+      `candidates[]` holds each pair that a body paragraph names in one
+      unit or through a lead-in (comention_count + lead_in_count >= 1),
+      with every paragraph that names both as evidence, sorted by
+      comention_count DESC, lead_in_count DESC, paragraph_count DESC, then
+      (a, b) ASC. `list_only_pair_count` counts the pairs only ever listed
+      together, which are not emitted. `fenced_blocks[]` lists the fenced
+      blocks skipped, in document order: the line of each opening fence and
+      its info string, trimmed (`mermaid` for a Mermaid diagram, "" for
+      none).
+
+      --technologies takes a JSON array of technology names as a document
+      writes them (the caller judges which strings are technologies) and
+      adds two keys:
+        "technologies": [{"name": "<as given>",
+                          "covered_by": [{"skill": "<skill>",
+                                          "term": "<the skill's term>",
+                                          "kind": "name|alias|name-contained|alias-contained"},
+                                         ...]}, ...],
+        "unverified_technologies": ["<name>", ...]
+      in the order given, a repeat (ignoring case and runs of whitespace)
+      dropped. A term covers only by squashed equality: a skill covers a
+      technology when one of its terms, its name or an alias, equals the
+      name once case, whitespace, hyphens, dots and underscores are
+      dropped (a `react-router` skill covers `React Router`, a
+      `tailwindcss` skill `Tailwind CSS`, a skill with the alias `cognee`
+      `Cognee`). A name or alias found only inside a longer name, by the
+      matching above with each occurrence read as its longest term, is
+      listed as name-contained or alias-contained and covers nothing:
+      whether `next` is `Next.js`, `react` is `React Native`, or the
+      common-word alias `ai` is `Azure AI Search`, is a question of meaning,
+      so the name stays unverified for the caller to judge. The match runs
+      one way: a `storybook-react-vite` skill does not cover `React`.
+      `covered_by` is sorted by skill, each with the first of its terms
+      that covers the name (or, for a contained entry, the term found) and
+      its kind, and `unverified_technologies` lists the names no skill
+      covers.
+
+  infer --skills <json-file-or-'-'> [--top-k N]
+      A pair is a candidate on either kind of evidence:
+        docs-mention     one skill's docs name the other. Every line of
+                         each `docs` file is read, front matter and code
+                         included; each occurrence names the skill of the
+                         longest term there (see Matching), and an
+                         occurrence of the documenting skill's own name or
+                         alias names no other skill. A mention is where a
+                         documented contract would show, not proof of one:
+                         the calling step reads the excerpt and counts a
+                         documented contract only when the line says how
+                         the two work together, never on a "see also" or
+                         "unlike" line.
+        shared-keywords  the two skills share a domain keyword. Keywords
+                         compare trimmed and lowercased, with runs of
+                         whitespace collapsed; a keyword equal to any
+                         skill's language does not count.
+      A shared language alone never makes a pair: `language` is read only
+      to set such keywords aside and to count language_only_pair_count,
+      the pairs of skills that share a language and have no evidence.
+
+      Emit JSON:
+        {
+          "pairs": [
+            {
+              "a": "<skill>", "b": "<skill>",
+              "docs_mention_count": 0-2,
+              "shared_keyword_count": K,
+              "evidence": [
+                {"kind": "docs-mention",
+                 "documented_by": "<skill>", "doc": "<path>",
+                 "line": L, "line_count": N,
+                 "excerpt": "<collapsed line text>"},
+                {"kind": "shared-keywords", "keywords": ["<kw>", ...]}
+              ]
+            },
+            ...
+          ],
+          "truncated": <bool>,
+          "total_pairs": T,
+          "language_only_pair_count": P
+        }
+
+      A docs-mention entry is the first line (docs in the order given)
+      where `documented_by` names the other skill, and line_count counts
+      every such line. docs_mention_count is how many of the two skills'
+      docs name the other. `pairs[]` is sorted by docs_mention_count DESC,
+      shared_keyword_count DESC, then (a, b) ASC, and capped at --top-k
+      (default 20, the Top-K of detect-integrations.md §1); `truncated` and
+      `total_pairs` report the cap as skf-pair-intersect.py does.
 
 CLI examples:
   uv run skf-comention-pairs.py comention --doc arch.md --skills skills.json
   echo '["react","express"]' | uv run skf-comention-pairs.py comention \\
       --doc arch.md --skills -
+  uv run skf-comention-pairs.py mentions --doc arch.md --skills skills.json
+  uv run skf-comention-pairs.py mentions --doc arch.md --skills skills.json \\
+      --technologies technologies.json
+  uv run skf-comention-pairs.py infer --skills skills.json --top-k 20
 
 Exit codes:
-  0  — operation succeeded (including: no qualifying pairs → empty list)
-  1  — user error (bad JSON, missing/unreadable file, both inputs on stdin,
-       malformed skills array)
-  2  — unexpected internal error
+  0  operation succeeded (including: nothing qualifies, so empty lists)
+  1  user error (bad JSON, a missing, unreadable or non-UTF-8 file, two
+     inputs on stdin, malformed skills or technologies array, negative
+     --top-k)
+  2  unexpected internal error
 """
 
 from __future__ import annotations
 
 import argparse
+import bisect
+import functools
+import itertools
 import json
 import re
 import sys
 from pathlib import Path
+from typing import NamedTuple
 
 
-# Closed exclusion set — governing H1/H2 headers that normalise to one of
-# these drop their paragraphs from co-mention analysis (they typically
-# enumerate all libraries without describing an integration).
+# The introductory sections: comention marks `excluded_section` on the
+# evidence of a paragraph whose governing H1/H2 header normalises to one of
+# these, since such sections often list every library without describing an
+# integration. The calling step weighs the mark; nothing is dropped.
 EXCLUDED_HEADERS = frozenset({
     "introduction",
     "overview",
@@ -96,21 +310,112 @@ EXCLUDED_HEADERS = frozenset({
 # as body text. Capture group 1 = hashes (level), group 2 = header text.
 _HEADER_RE = re.compile(r"^(#{1,6})[ \t]+(.*)$")
 
+# A fence line: three or more backticks or tildes after any indentation
+# (the rule the other SKF Markdown scanners use), then the info string.
+_FENCE_RE = re.compile(r"^\s*(`{3,}|~{3,})(.*)$")
+
+# The indentation and blockquote markers (`>`) that can precede a list
+# marker or a table row.
+_QUOTE_RE = re.compile(r"\s*(?:>\s*)*")
+
+# List item markers, at any indentation so that a nested item is a unit of
+# its own, and inside a blockquote: a bullet, or a number with `.` or `)`,
+# then whitespace and text.
+_BULLET_RE = re.compile(r"^\s*(?:>\s*)*[-*+]\s+\S")
+_ORDERED_RE = re.compile(r"^\s*(?:>\s*)*(\d{1,9})[.)]\s+\S")
+
+# A GFM table delimiter row, such as `| --- | :-: |` or `--- | ---`, matched
+# whole against the row with its blockquote markers and outer whitespace
+# removed: leading and trailing `\s*` here would backtrack quadratically on
+# a long run of spaces.
+_TABLE_DELIM_RE = re.compile(r"\|?\s*:?-+:?\s*(?:\|\s*:?-+:?\s*)*\|?")
+
+_WORD_RE = re.compile(r"\S+")
+
 _EXCERPT_LIMIT = 240
 
+# infer's Top-K cap: the cap detect-integrations.md §1 applies to pairs.
+DEFAULT_TOP_K = 20
+
+CO_MENTION = "co-mention"
+LEAD_IN = "lead-in"
+LIST_ONLY = "list-only"
+DOCS_MENTION = "docs-mention"
+SHARED_KEYWORDS = "shared-keywords"
+# The covered_by kinds of a skill name or an alias found only inside a longer
+# technology name: listed, but they cover nothing.
+NAME_CONTAINED = "name-contained"
+ALIAS_CONTAINED = "alias-contained"
+
 
 class UserError(Exception):
     """A caller-facing input error → exit code 1."""
 
 
+class Skill(NamedTuple):
+    """One loaded skill: the terms that name it (its name first, then its
+    aliases) and the inputs only infer reads."""
+
+    name: str
+    terms: tuple[str, ...]
+    keywords: tuple[str, ...] = ()
+    languages: tuple[str, ...] = ()
+    docs: tuple[str, ...] = ()
+
+
+class Unit(NamedTuple):
+    """A part of a paragraph that a pair can share: its first line
+    (1-based), its text, and whether it is a run of prose (the others are
+    list items, table rows and code lines)."""
+
+    line: int
+    text: str
+    prose: bool
+
+
+class Paragraph(NamedTuple):
+    """A body paragraph: its governing H1/H2 header (or None), its first
+    line (1-based), its text and its units."""
+
+    header: str | None
+    line: int
+    text: str
+    units: tuple[Unit, ...]
+
+
+class Fence(NamedTuple):
+    """A fenced code block: the line of its opening fence (1-based), its
+    info string, trimmed, and its content without the fence lines."""
+
+    line: int
+    info: str
+    text: str
+
+
+class MarkdownDoc(NamedTuple):
+    """What read_markdown finds: body paragraphs, the text of every header
+    line, and each fenced block it set apart."""
+
+    paragraphs: list[Paragraph]
+    headings: list[str]
+    fenced: list[Fence]
+
+
 # --------------------------------------------------------------------------
 # Markdown parsing
 # --------------------------------------------------------------------------
 
 
 def _strip_closing_hashes(text: str) -> str:
-    """Remove a closed-ATX trailing run of hashes (e.g. `Overview ##`)."""
-    return re.sub(r"[ \t]*#+[ \t]*$", "", text)
+    """Remove a closed-ATX trailing run of hashes (e.g. `Overview ##`).
+
+    String methods rather than a `[ \\t]*#+[ \\t]*$` search, which
+    backtracks quadratically on a long run of spaces.
+    """
+    body = text.rstrip(" \t")
+    if not body.endswith("#"):
+        return text
+    return body.rstrip("#").rstrip(" \t")
 
 
 def normalize_header(text: str) -> str:
@@ -122,27 +427,163 @@ def normalize_header(text: str) -> str:
     return t
 
 
-def parse_body_paragraphs(doc_text: str) -> list[tuple[str | None, str]]:
-    """Split the document into body paragraphs, each tagged with the raw text
-    of its governing (nearest-preceding) H1/H2 header (or None).
+def _fence_roles(lines: list[str]) -> list[str | None]:
+    """Mark each line of a fenced code block `open`, `body` or `close`, and
+    every other line None. A fence closes on a line of the same character,
+    at least as long, with nothing after it; an unclosed fence runs to the
+    end of the document."""
+    roles: list[str | None] = []
+    fence: str | None = None
+    for line in lines:
+        m = _FENCE_RE.match(line)
+        if fence is None:
+            # A backtick fence's info string cannot hold a backtick.
+            if m is not None and not (m.group(1)[0] == "`" and "`" in m.group(2)):
+                fence = m.group(1)
+                roles.append("open")
+            else:
+                roles.append(None)
+        elif (
+            m is not None
+            and m.group(1)[0] == fence[0]
+            and len(m.group(1)) >= len(fence)
+            and not m.group(2).strip()
+        ):
+            fence = None
+            roles.append("close")
+        else:
+            roles.append("body")
+    return roles
+
+
+def _fenced_blocks(lines: list[str], roles: list[str | None]) -> list[Fence]:
+    """Each fenced code block, with its opening line and info string."""
+    blocks: list[tuple[int, str, list[str]]] = []
+    for idx, (line, role) in enumerate(zip(lines, roles)):
+        if role == "open":
+            info = _FENCE_RE.match(line).group(2).strip()
+            blocks.append((idx + 1, info, []))
+        elif role == "body":
+            blocks[-1][2].append(line)
+    return [Fence(line, info, "\n".join(body)) for line, info, body in blocks]
+
+
+def _unquote(line: str) -> str:
+    """The line without its indentation and blockquote markers."""
+    return line[_QUOTE_RE.match(line).end():]
+
+
+def _is_list_item(line: str) -> bool:
+    return bool(_BULLET_RE.match(line) or _ORDERED_RE.match(line))
+
+
+def _table_rows(lines: list[str], code: list[bool]) -> set[int]:
+    """Indexes of a paragraph's table rows: every line that starts with `|`,
+    and each GFM table (a header row, its delimiter row, then each line up
+    to a list item or code) even when its rows have no leading pipe."""
+    rows = {
+        i for i, line in enumerate(lines)
+        if not code[i] and _unquote(line).startswith("|")
+    }
+    for i in range(1, len(lines)):
+        if (
+            code[i] or code[i - 1]
+            or "|" not in lines[i] or "|" not in lines[i - 1]
+            or not _TABLE_DELIM_RE.fullmatch(_unquote(lines[i]).rstrip())
+        ):
+            continue
+        rows.update((i - 1, i))
+        j = i + 1
+        while j < len(lines) and not code[j] and not _is_list_item(lines[j]):
+            rows.add(j)
+            j += 1
+    return rows
+
+
+def _units(lines: list[str], code: list[bool], first: int) -> tuple[Unit, ...]:
+    """Split a paragraph's lines, the first of which is line `first` of the
+    document, into the units a pair can share: a run of prose, one list item
+    with its continuation lines, one table row, or one line of fenced
+    code."""
+    rows = _table_rows(lines, code)
+    units: list[Unit] = []
+    start = 0  # index of the open unit's first line
+    # What the open unit holds: "prose", "item" (a list item) or None.
+    current: str | None = None
+
+    def close(end: int) -> None:
+        if current is not None:
+            units.append(Unit(
+                first + start, "\n".join(lines[start:end]), current == "prose"
+            ))
+
+    for i, line in enumerate(lines):
+        if code[i] or i in rows:
+            close(i)
+            current = None
+            units.append(Unit(first + i, line, False))
+            continue
+        ordered = _ORDERED_RE.match(line)
+        # A bullet can interrupt running prose; an ordered marker only when
+        # it is 1 (CommonMark), so `2024. Then ...` inside prose stays text.
+        if _BULLET_RE.match(line) or (
+            ordered is not None
+            and (current != "prose" or int(ordered.group(1)) == 1)
+        ):
+            close(i)
+            start, current = i, "item"
+        elif current is None:
+            # Any other line opens a run of prose, or continues the open
+            # run or list item.
+            start, current = i, "prose"
+    close(len(lines))
+    return tuple(units)
+
+
+def read_markdown(doc_text: str, *, skip_fenced: bool) -> MarkdownDoc:
+    """Split the document into body paragraphs, headings and fenced blocks.
 
     Paragraphs are maximal runs of consecutive non-blank, non-header lines,
-    separated by blank lines. Header lines of ANY level are never emitted as
-    paragraphs (headings are not a co-mention source). Only H1/H2 headers
-    update the governing header; H3-H6 leave it unchanged.
+    separated by blank lines. Header lines of ANY level are never emitted
+    as paragraphs (headings are not a co-mention source). Only H1/H2
+    headers update the governing header; H3-H6 leave it unchanged.
+
+    With skip_fenced (mentions), fenced code blocks are set apart first:
+    their lines end a paragraph, are never headers, and each block lands in
+    `fenced`. Without it (comention), fence lines are body text and a `#`
+    line inside a fence is a header, as that mode has always read them; the
+    units still know which lines are code.
     """
-    paragraphs: list[tuple[str | None, str]] = []
+    lines = doc_text.splitlines()
+    roles = _fence_roles(lines)
+    doc = MarkdownDoc(
+        paragraphs=[],
+        headings=[],
+        fenced=_fenced_blocks(lines, roles) if skip_fenced else [],
+    )
     governing: str | None = None
-    buf: list[str] = []
+    buf: list[int] = []  # line indexes of the open paragraph
 
     def _flush() -> None:
         if buf:
-            text = "\n".join(buf).strip()
+            text = "\n".join(lines[i] for i in buf).strip()
             if text:
-                paragraphs.append((governing, text))
+                doc.paragraphs.append(Paragraph(
+                    header=governing,
+                    line=buf[0] + 1,
+                    text=text,
+                    units=_units(
+                        [lines[i] for i in buf],
+                        [roles[i] is not None for i in buf],
+                        buf[0] + 1,
+                    ),
+                ))
             buf.clear()
 
-    for raw_line in doc_text.splitlines():
+    for idx, raw_line in enumerate(lines):
+        if skip_fenced and roles[idx] is not None:
+            _flush()
+            continue
         m = _HEADER_RE.match(raw_line)
         if m is not None:
             # A header terminates the current paragraph; the header line
@@ -150,19 +591,202 @@ def parse_body_paragraphs(doc_text: str) -> list[tuple[str | None, str]]:
             _flush()
             level = len(m.group(1))
             header_text = _strip_closing_hashes(m.group(2)).strip()
+            doc.headings.append(header_text)
             if level <= 2:
                 governing = header_text if header_text else None
             continue
         if raw_line.strip() == "":
             _flush()
             continue
-        buf.append(raw_line)
+        buf.append(idx)
     _flush()
-    return paragraphs
+    return doc
+
+
+def parse_body_paragraphs(doc_text: str) -> list[tuple[str | None, str]]:
+    """Body paragraphs as (governing header text or None, paragraph text),
+    read the way comention reads them (fenced code is body text)."""
+    return [
+        (p.header, p.text)
+        for p in read_markdown(doc_text, skip_fenced=False).paragraphs
+    ]
 
 
 # --------------------------------------------------------------------------
-# Co-mention analysis
+# Skills and matching
+# --------------------------------------------------------------------------
+
+
+def _norm(text: str) -> str:
+    """Lowercase, trimmed, with each run of whitespace collapsed."""
+    return " ".join(text.lower().split())
+
+
+def _dedupe(items: tuple[str, ...], key=lambda s: s) -> tuple[str, ...]:
+    """Drop repeats (by key), keeping the first of each in order."""
+    seen: set[str] = set()
+    out: list[str] = []
+    for item in items:
+        k = key(item)
+        if k not in seen:
+            seen.add(k)
+            out.append(item)
+    return tuple(out)
+
+
+def _term_body(term: str) -> str:
+    """Regex source for a term's words, with any whitespace between them."""
+    return r"\s+".join(re.escape(word) for word in term.split())
+
+
+def _skill_pattern(skill: Skill) -> re.Pattern[str]:
+    """One case-insensitive pattern that matches any of the skill's terms
+    with no word character on either side."""
+    terms = sorted(skill.terms, key=lambda t: (-len(t), t))
+    return re.compile(
+        "|".join(rf"(?:(?<!\w){_term_body(t)}(?!\w))" for t in terms),
+        re.IGNORECASE,
+    )
+
+
+class _Occurrence(NamedTuple):
+    """A match of a term in a text, and the skills it names there. `term`
+    is the matched term, normalised (mentions and infer only)."""
+
+    start: int
+    end: int
+    names: frozenset[str]
+    term: str = ""
+
+
+class _EachSkill:
+    """comention's matcher: each skill's terms on their own, so `react`
+    also matches the `react` in `react-dom`."""
+
+    def __init__(self, specs: dict[str, Skill]) -> None:
+        self.patterns = {n: _skill_pattern(specs[n]) for n in sorted(specs)}
+
+    def names(self, text: str) -> set[str]:
+        return {n for n, pattern in self.patterns.items() if pattern.search(text)}
+
+    def find(self, text: str, among: frozenset[str]) -> list[_Occurrence]:
+        """Every match in the text of a term of a skill in `among`."""
+        return [
+            _Occurrence(m.start(), m.end(), frozenset((n,)))
+            for n in sorted(among)
+            for m in self.patterns[n].finditer(text)
+        ]
+
+
+class _LongestTerm:
+    """The matcher of mentions and infer: one pattern over every skill's
+    terms, longest first, so each occurrence is the longest term that
+    matches there, and names the skills of that term."""
+
+    def __init__(self, specs: dict[str, Skill]) -> None:
+        owners: dict[str, set[str]] = {}
+        for skill in specs.values():
+            for term in skill.terms:
+                owners.setdefault(_norm(term), set()).add(skill.name)
+        self.owners = {t: frozenset(n) for t, n in owners.items()}
+        self.terms = sorted(self.owners, key=lambda t: (-len(t), t))
+        # Plain alternatives (no groups) keep the regex engine's fast path.
+        alternatives = "|".join(_term_body(t) for t in self.terms)
+        self.pattern = re.compile(
+            rf"(?<!\w)(?:{alternatives})(?!\w)", re.IGNORECASE
+        ) if self.terms else None
+
+    def names(self, text: str) -> set[str]:
+        return set().union(*(o.names for o in self.find(text)))
+
+    def find(
+        self, text: str, among: frozenset[str] | None = None
+    ) -> list[_Occurrence]:
+        """Each occurrence in the text, in order, with the skills it names
+        (only those in `among`, when given; an occurrence naming none of
+        them is left out)."""
+        found: list[_Occurrence] = []
+        for m in self.pattern.finditer(text) if self.pattern else ():
+            term = self._term(m.group(0))
+            names = self.owners.get(term, frozenset())
+            if among is not None:
+                names &= among
+            if names:
+                found.append(_Occurrence(m.start(), m.end(), names, term))
+        return found
+
+    def _term(self, text: str) -> str:
+        """The normalised term the pattern matched in `text`."""
+        term = _norm(text)
+        if term in self.owners:
+            return term
+        # A case-insensitive match that lowercasing does not map back to its
+        # term (the long s matches `s`): the first term, longest first, that
+        # matches the text whole is the one the pattern chose.
+        for term in self.terms:
+            if re.fullmatch(_term_body(term), text, re.IGNORECASE):
+                return term
+        return ""
+
+
+def _coerce(skills: list[Skill | str]) -> dict[str, Skill]:
+    """Index skills by name, merging entries that share a name. A plain
+    string is a skill with no alias."""
+    merged: dict[str, Skill] = {}
+    for item in skills:
+        skill = item if isinstance(item, Skill) else Skill(item, (item,))
+        prev = merged.get(skill.name)
+        if prev is not None:
+            skill = Skill(
+                name=skill.name,
+                terms=_dedupe(prev.terms + skill.terms, key=_norm),
+                keywords=tuple(sorted(set(prev.keywords + skill.keywords))),
+                languages=tuple(sorted(set(prev.languages + skill.languages))),
+                docs=_dedupe(prev.docs + skill.docs),
+            )
+        merged[skill.name] = skill
+    return merged
+
+
+def _string_list(entry: dict, key: str, where: str) -> tuple[str, ...]:
+    """The entry's optional `key`: an array of non-empty strings."""
+    value = entry.get(key)
+    if value is None:
+        return ()
+    if not isinstance(value, list):
+        raise UserError(
+            f"{where}.{key} must be an array of non-empty strings; "
+            f"got {type(value).__name__}"
+        )
+    out: list[str] = []
+    for j, item in enumerate(value):
+        if not isinstance(item, str) or not item.strip():
+            raise UserError(
+                f"{where}.{key}[{j}] must be a non-empty string; got {item!r}"
+            )
+        out.append(item.strip())
+    return tuple(out)
+
+
+def _languages(entry: dict, where: str) -> tuple[str, ...]:
+    """The entry's optional `language`: a string, or an array of strings (a
+    stack lists its libraries' languages), normalised and sorted."""
+    value = entry.get("language")
+    if value is None:
+        return ()
+    items = [value] if isinstance(value, str) else value
+    if not isinstance(items, list) or not all(
+        isinstance(item, str) and item.strip() for item in items
+    ):
+        raise UserError(
+            f"{where}.language must be a non-empty string or an array of "
+            f"non-empty strings; got {value!r}"
+        )
+    return tuple(sorted({_norm(item) for item in items}))
+
+
+# --------------------------------------------------------------------------
+# Evidence
 # --------------------------------------------------------------------------
 
 
@@ -174,53 +798,488 @@ def _make_excerpt(text: str, limit: int = _EXCERPT_LIMIT) -> str:
     return collapsed
 
 
-def analyze(doc_text: str, skill_names: list[str]) -> dict:
-    """Compute qualifying co-mention pairs and their evidence.
+@functools.lru_cache(maxsize=256)
+def _words(text: str) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
+    """The text's words, found once per text: where each starts and ends,
+    and the running total of their lengths."""
+    spans = [m.span() for m in _WORD_RE.finditer(text)]
+    return (
+        tuple(s for s, _ in spans),
+        tuple(e for _, e in spans),
+        tuple(itertools.accumulate((e - s for s, e in spans), initial=0)),
+    )
 
-    Deterministic: same (doc_text, skill_names) → identical output.
-    """
-    names = sorted(set(skill_names))
-    patterns = {
-        n: re.compile(r"\b" + re.escape(n) + r"\b", re.IGNORECASE)
-        for n in names
+
+def _excerpt_around(
+    text: str,
+    first: tuple[int, int],
+    second: tuple[int, int],
+    limit: int = _EXCERPT_LIMIT,
+) -> str:
+    """Collapse `text` like an excerpt and cut it to `limit` characters
+    around two names, at the spans `first` and `second` (first starting
+    first): the words around both, or, when the words between them do not
+    fit, the words around each with the middle cut. `...` marks each cut."""
+    starts, ends, pre = _words(text)
+    last = len(starts) - 1
+
+    def join(i: int, j: int) -> str:
+        return " ".join(text[starts[k]:ends[k]] for k in range(i, j + 1))
+
+    if pre[-1] + last <= limit:  # the whole text fits
+        return join(0, last)
+
+    def word(pos: int) -> int:
+        """The index of the word that holds text[pos]."""
+        return bisect.bisect_right(ends, pos)
+
+    def width(i: int, j: int) -> int:
+        """Words i..j joined, with a `...` for each side cut."""
+        return pre[j + 1] - pre[i] + (j - i) + 3 * (i > 0) + 3 * (j < last)
+
+    def grow(i: int, j: int, room: int) -> tuple[int, int]:
+        """Add words on both sides of i..j, in turn, while they fit."""
+        grown = True
+        while grown:
+            grown = False
+            if i > 0 and width(i - 1, j) <= room:
+                i, grown = i - 1, True
+            if j < last and width(i, j + 1) <= room:
+                j, grown = j + 1, True
+        return i, j
+
+    i, j = word(first[0]), word(max(first[1], second[1]) - 1)
+    first_end, second_start = word(first[1] - 1), word(second[0])
+    if width(i, j) <= limit or second_start <= first_end:
+        windows = [grow(i, j, limit)]
+    else:
+        # The words between the names do not fit: keep each name with the
+        # words around it, and cut the middle.
+        windows = [
+            grow(i, first_end, limit // 2), grow(second_start, j, limit // 2),
+        ]
+    quote = " ... ".join(join(a, b) for a, b in windows)
+    if windows[0][0] > 0:
+        quote = "..." + quote
+    if windows[-1][1] < last:
+        quote += "..."
+    # Only a name inside a word too long for its window overflows the limit.
+    return _make_excerpt(quote, limit)
+
+
+def _pair_quote(
+    text: str,
+    line: int,
+    found_a: list[_Occurrence],
+    found_b: list[_Occurrence],
+) -> tuple[int, str]:
+    """Where `text`, whose first line is `line`, names two skills nearest
+    each other, given where it names each: the line of the first of the
+    two names, and the excerpt around them."""
+    # One pass in text order: each occurrence pairs with the latest one
+    # before it of the other skill (an occurrence of both, with itself).
+    marked = sorted(
+        [(o.start, o.end, 0, o) for o in found_a]
+        + [(o.start, o.end, 1, o) for o in found_b],
+        key=lambda m: m[:3],
+    )
+    latest: list[_Occurrence | None] = [None, None]
+    best: tuple[tuple[int, int], _Occurrence, _Occurrence] | None = None
+    for _, _, side, o in marked:
+        p = latest[1 - side]
+        if p is not None:
+            key = (max(p.end, o.end) - p.start, p.start)
+            if best is None or key < best[0]:
+                best = (key, p, o)
+        latest[side] = o
+    _, first, second = best  # the text names both skills
+    return (
+        line + text.count("\n", 0, first.start),
+        _excerpt_around(
+            text, (first.start, first.end), (second.start, second.end)
+        ),
+    )
+
+
+def _paragraph_evidence(
+    para: Paragraph, present: list[str], matcher: _EachSkill | _LongestTerm
+) -> list[tuple[tuple[str, str], dict]]:
+    """One evidence entry per pair of the skills a paragraph names, with
+    its kind and where the paragraph names the pair. `present` is sorted,
+    so every pair is (a, b) with a < b."""
+    if len(present) < 2:
+        return []
+    among = frozenset(present)
+    # Each unit, with where it names each skill.
+    units: list[tuple[Unit, dict[str, list[_Occurrence]]]] = []
+    for unit in para.units:
+        named: dict[str, list[_Occurrence]] = {}
+        for o in matcher.find(unit.text, among):
+            for n in o.names:
+                named.setdefault(n, []).append(o)
+        units.append((unit, named))
+    # pair -> (kind, unit_line, unit_excerpt); a co-mention wins over a
+    # lead-in, and the first unit (or lead-in and item) over later ones.
+    where: dict[tuple[str, str], tuple[str, int, str]] = {}
+    for unit, named in units:
+        for a, b in itertools.combinations(sorted(named), 2):
+            if (a, b) not in where:
+                where[a, b] = (CO_MENTION, *_pair_quote(
+                    unit.text, unit.line, named[a], named[b]
+                ))
+    lead: int | None = None  # the run of prose the next units sit under
+    for k, (unit, named) in enumerate(units):
+        if unit.prose:
+            lead = k
+            continue
+        if lead is None:
+            continue
+        lead_unit, lead_named = units[lead]
+        # Each skill of such a pair is named in one of the two units only:
+        # a unit that named both would have made it a co-mention.
+        pairs = [
+            (min(x, y), max(x, y))
+            for x, y in itertools.product(sorted(lead_named), sorted(named))
+            if (min(x, y), max(x, y)) not in where and x != y
+        ]
+        if not pairs:
+            continue
+        # The lead-in, then the item, with `...` for any items between.
+        gap = "\n" if k == lead + 1 else "\n...\n"
+        text = lead_unit.text + gap + unit.text
+        shift = len(lead_unit.text) + len(gap)
+        for a, b in pairs:
+            found = [
+                lead_named[n] if n in lead_named else [
+                    o._replace(start=o.start + shift, end=o.end + shift)
+                    for o in named[n]
+                ]
+                for n in (a, b)
+            ]
+            where[a, b] = (LEAD_IN, *_pair_quote(text, lead_unit.line, *found))
+    excerpt = _make_excerpt(para.text)
+    evidence = []
+    for pair in itertools.combinations(present, 2):
+        kind, unit_line, unit_excerpt = where.get(pair, (LIST_ONLY, None, None))
+        evidence.append((pair, {
+            "header": para.header,
+            "excerpt": excerpt,
+            "line": para.line,
+            "kind": kind,
+            "unit_line": unit_line,
+            "unit_excerpt": unit_excerpt,
+        }))
+    return evidence
+
+
+def _pair_entry(a: str, b: str, evidence: list[dict]) -> dict:
+    return {
+        "a": a,
+        "b": b,
+        "paragraph_count": len(evidence),
+        "comention_count": sum(1 for e in evidence if e["kind"] == CO_MENTION),
+        "lead_in_count": sum(1 for e in evidence if e["kind"] == LEAD_IN),
+        "evidence": evidence,
     }
-    paragraphs = parse_body_paragraphs(doc_text)
+
+
+# --------------------------------------------------------------------------
+# Subcommand logic
+# --------------------------------------------------------------------------
+
+
+def analyze(doc_text: str, skills: list[Skill | str]) -> dict:
+    """comention: every candidate pair a body paragraph names in one unit
+    or through a lead-in, with its evidence, each entry marking whether an
+    introductory section holds it, and the count of pairs only listed
+    together.
+
+    Deterministic: same (doc_text, skills) → identical output.
+    """
+    matcher = _EachSkill(_coerce(skills))
 
     excluded_sections: set[str] = set()
-    # (a, b) -> list of evidence dicts, one per qualifying paragraph
+    # (a, b) -> list of evidence dicts, one per paragraph naming both
     pair_evidence: dict[tuple[str, str], list[dict]] = {}
 
-    for governing, text in paragraphs:
-        if governing is not None:
-            norm = normalize_header(governing)
+    for para in read_markdown(doc_text, skip_fenced=False).paragraphs:
+        excluded = False
+        if para.header is not None:
+            norm = normalize_header(para.header)
             if norm in EXCLUDED_HEADERS:
                 excluded_sections.add(norm)
-                continue
-        # names is sorted, so `present` is sorted → pairs are (a < b)
-        present = [n for n in names if patterns[n].search(text)]
-        if len(present) < 2:
-            continue
-        excerpt = _make_excerpt(text)
-        evidence_entry = {"header": governing, "excerpt": excerpt}
-        for i in range(len(present)):
-            for j in range(i + 1, len(present)):
-                key = (present[i], present[j])
-                pair_evidence.setdefault(key, []).append(evidence_entry)
+                excluded = True
+        # `present` is sorted → pairs are (a < b)
+        present = sorted(matcher.names(para.text))
+        for pair, entry in _paragraph_evidence(para, present, matcher):
+            entry["excluded_section"] = excluded
+            pair_evidence.setdefault(pair, []).append(entry)
 
     pairs: list[dict] = []
+    list_only_pairs = 0
     for (a, b), evidence in pair_evidence.items():
-        if len(evidence) >= 2:
+        entry = _pair_entry(a, b, evidence)
+        if entry["comention_count"] or entry["lead_in_count"]:
+            pairs.append(entry)
+        else:
+            list_only_pairs += 1
+    pairs.sort(key=lambda p: (-p["paragraph_count"], p["a"], p["b"]))
+
+    return {
+        "pairs": pairs,
+        "list_only_pair_count": list_only_pairs,
+        "excluded_section_count": len(excluded_sections),
+    }
+
+
+def _named(occurrences: list[_Occurrence]) -> set[str]:
+    return set().union(*(o.names for o in occurrences))
+
+
+def _spelling(skill: Skill, term: str) -> tuple[str, str]:
+    """The skill's own spelling of a normalised term, and its kind."""
+    for i, t in enumerate(skill.terms):
+        if _norm(t) == term:
+            return t, "name" if i == 0 else "alias"
+    return term, "alias"
+
+
+def _squash(text: str) -> str:
+    """Lowercased, with whitespace, hyphens, dots and underscores dropped."""
+    return re.sub(r"[\s\-._]+", "", text.lower())
+
+
+def coverage(
+    technologies: list[str], specs: dict[str, Skill], matcher: _LongestTerm
+) -> tuple[list[dict], list[str]]:
+    """Which skills cover each technology name, and the names none covers.
+    Names are trimmed, and a repeat (ignoring case and runs of whitespace)
+    is dropped. A term covers only by squashed equality: a name or an
+    alias found only inside the name is listed as `name-contained` or
+    `alias-contained` and covers nothing."""
+    technologies = list(_dedupe(tuple(t.strip() for t in technologies), key=_norm))
+    squashed = [
+        (n, t, "name" if i == 0 else "alias", _squash(t))
+        for n in sorted(specs) for i, t in enumerate(specs[n].terms)
+    ]
+    entries: list[dict] = []
+    unverified: list[str] = []
+    for tech in technologies:
+        # skill -> (term, kind): the skills that cover the name.
+        found: dict[str, tuple[str, str]] = {}
+        # skill -> (term, kind): the skills one of whose terms sits inside
+        # the name.
+        contained: dict[str, tuple[str, str]] = {}
+        for occurrence in matcher.find(tech):
+            for n in sorted(occurrence.names):
+                term, kind = _spelling(specs[n], occurrence.term)
+                kind = NAME_CONTAINED if kind == "name" else ALIAS_CONTAINED
+                contained.setdefault(n, (term, kind))
+        target = _squash(tech)
+        for n, term, kind, flat in squashed:
+            if flat and flat == target:
+                found.setdefault(n, (term, kind))
+        covered_by = {
+            n: {"skill": n, "term": term, "kind": kind}
+            for n, (term, kind) in found.items()
+        }
+        for n, (term, kind) in contained.items():
+            covered_by.setdefault(n, {"skill": n, "term": term, "kind": kind})
+        entries.append({
+            "name": tech,
+            "covered_by": [covered_by[n] for n in sorted(covered_by)],
+        })
+        if not found:
+            unverified.append(tech)
+    return entries, unverified
+
+
+def mentions(
+    doc_text: str,
+    skills: list[Skill | str],
+    technologies: list[str] | None = None,
+) -> dict:
+    """mentions: which skills the document names outside fenced code, the
+    term that named each, the candidate pairs a body paragraph names in one
+    unit or through a lead-in, and, given technology names, which skills
+    cover each."""
+    specs = _coerce(skills)
+    names = sorted(specs)
+    matcher = _LongestTerm(specs)
+    doc = read_markdown(doc_text, skip_fenced=True)
+
+    # (skill, normalised term) -> [paragraphs, headings, fenced blocks]
+    term_counts: dict[tuple[str, str], list[int]] = {}
+
+    def count_terms(occurrences: list[_Occurrence], slot: int) -> None:
+        for name, term in {(n, o.term) for o in occurrences for n in o.names}:
+            term_counts.setdefault((name, term), [0, 0, 0])[slot] += 1
+
+    hits: dict[str, list[dict]] = {n: [] for n in names}
+    pair_evidence: dict[tuple[str, str], list[dict]] = {}
+    for para in doc.paragraphs:
+        occurrences = matcher.find(para.text)
+        count_terms(occurrences, 0)
+        present = sorted(_named(occurrences))
+        if not present:
+            continue
+        excerpt = _make_excerpt(para.text)
+        for n in present:
+            hits[n].append(
+                {"header": para.header, "excerpt": excerpt, "line": para.line}
+            )
+        for pair, entry in _paragraph_evidence(para, present, matcher):
+            pair_evidence.setdefault(pair, []).append(entry)
+
+    in_headings: list[set[str]] = []
+    for heading in doc.headings:
+        occurrences = matcher.find(heading)
+        count_terms(occurrences, 1)
+        in_headings.append(_named(occurrences))
+    in_fenced: list[set[str]] = []
+    for fence in doc.fenced:
+        occurrences = matcher.find(fence.text)
+        count_terms(occurrences, 2)
+        in_fenced.append(_named(occurrences))
+    skills_out: list[dict] = []
+    mentioned: list[str] = []
+    fenced_only: list[str] = []
+    unmentioned: list[str] = []
+    for n in names:
+        heading_count = sum(1 for found in in_headings if n in found)
+        fenced_count = sum(1 for found in in_fenced if n in found)
+        terms = []
+        for term in specs[n].terms:
+            counts = term_counts.get((n, _norm(term)))
+            if counts:
+                terms.append({
+                    "term": term,
+                    "kind": _spelling(specs[n], _norm(term))[1],
+                    "paragraph_count": counts[0],
+                    "heading_count": counts[1],
+                    "fenced_count": counts[2],
+                })
+        skills_out.append({
+            "name": n,
+            "paragraph_count": len(hits[n]),
+            "heading_count": heading_count,
+            "fenced_count": fenced_count,
+            "terms": terms,
+            "paragraphs": hits[n],
+        })
+        if hits[n] or heading_count:
+            mentioned.append(n)
+        elif fenced_count:
+            fenced_only.append(n)
+        else:
+            unmentioned.append(n)
+
+    candidates: list[dict] = []
+    list_only_pairs = 0
+    for (a, b), evidence in pair_evidence.items():
+        entry = _pair_entry(a, b, evidence)
+        if entry["comention_count"] or entry["lead_in_count"]:
+            candidates.append(entry)
+        else:
+            list_only_pairs += 1
+    candidates.sort(key=lambda p: (
+        -p["comention_count"], -p["lead_in_count"], -p["paragraph_count"],
+        p["a"], p["b"],
+    ))
+
+    result = {
+        "skills": skills_out,
+        "mentioned": mentioned,
+        "fenced_only": fenced_only,
+        "unmentioned": unmentioned,
+        "candidates": candidates,
+        "list_only_pair_count": list_only_pairs,
+        "fenced_block_count": len(doc.fenced),
+        "fenced_blocks": [{"line": f.line, "info": f.info} for f in doc.fenced],
+    }
+    if technologies is not None:
+        entries, unverified = coverage(technologies, specs, matcher)
+        result["technologies"] = entries
+        result["unverified_technologies"] = unverified
+    return result
+
+
+def _read_doc(path_text: str, owner: str) -> str:
+    """Read one of a skill's `docs` files (infer)."""
+    path = Path(path_text)
+    if not path.is_file():
+        raise UserError(f"docs file of skill {owner!r} not found: {path}")
+    try:
+        return path.read_text(encoding="utf-8")
+    except (OSError, UnicodeDecodeError) as exc:
+        raise UserError(
+            f"failed to read docs file of skill {owner!r} ({path}): {exc}"
+        ) from exc
+
+
+def infer(skills: list[Skill | str], top_k: int = DEFAULT_TOP_K) -> dict:
+    """infer: candidate pairs without an architecture document, from one
+    skill's docs naming the other or shared domain keywords, never a shared
+    language alone. Top-K capped."""
+    specs = _coerce(skills)
+    names = sorted(specs)
+    languages = {lang for s in specs.values() for lang in s.languages}
+    matcher = _LongestTerm(specs)
+
+    # (documenting skill, named skill) -> its docs-mention entry
+    mentioned: dict[tuple[str, str], dict] = {}
+    for a in names:
+        for doc in specs[a].docs:
+            for lineno, line in enumerate(_read_doc(doc, a).splitlines(), start=1):
+                named: set[str] = set()
+                for occurrence in matcher.find(line):
+                    # An occurrence of the documenting skill's own term
+                    # names that skill, never another one.
+                    if a not in occurrence.names:
+                        named |= occurrence.names
+                for b in sorted(named):
+                    found = mentioned.get((a, b))
+                    if found is None:
+                        mentioned[(a, b)] = {
+                            "kind": DOCS_MENTION,
+                            "documented_by": a,
+                            "doc": doc.replace("\\", "/"),
+                            "line": lineno,
+                            "line_count": 1,
+                            "excerpt": _make_excerpt(line),
+                        }
+                    else:
+                        found["line_count"] += 1
+
+    pairs: list[dict] = []
+    language_only_pairs = 0
+    for a, b in itertools.combinations(names, 2):
+        evidence = [mentioned[k] for k in ((a, b), (b, a)) if k in mentioned]
+        docs_mention_count = len(evidence)
+        shared = sorted(
+            (set(specs[a].keywords) & set(specs[b].keywords)) - languages
+        )
+        if shared:
+            evidence.append({"kind": SHARED_KEYWORDS, "keywords": shared})
+        if evidence:
             pairs.append({
                 "a": a,
                 "b": b,
-                "paragraph_count": len(evidence),
+                "docs_mention_count": docs_mention_count,
+                "shared_keyword_count": len(shared),
                 "evidence": evidence,
             })
-    pairs.sort(key=lambda p: (-p["paragraph_count"], p["a"], p["b"]))
+        elif set(specs[a].languages) & set(specs[b].languages):
+            language_only_pairs += 1
+    pairs.sort(key=lambda p: (
+        -p["docs_mention_count"], -p["shared_keyword_count"], p["a"], p["b"],
+    ))
 
     return {
-        "pairs": pairs,
-        "excluded_section_count": len(excluded_sections),
+        "pairs": pairs[:top_k],
+        "truncated": len(pairs) > top_k,
+        "total_pairs": len(pairs),
+        "language_only_pair_count": language_only_pairs,
     }
 
 
@@ -234,39 +1293,76 @@ def _read_source(source: str, label: str) -> str:
     if source == "-":
         try:
             return sys.stdin.read()
-        except OSError as exc:
+        except (OSError, UnicodeDecodeError) as exc:
             raise UserError(f"failed to read {label} from stdin: {exc}") from exc
     path = Path(source)
     if not path.is_file():
         raise UserError(f"{label} file not found: {path}")
     try:
         return path.read_text(encoding="utf-8")
-    except OSError as exc:
+    except (OSError, UnicodeDecodeError) as exc:
         raise UserError(f"failed to read {label} file {path}: {exc}") from exc
 
 
-def parse_skills(raw_text: str) -> list[str]:
-    """Parse and validate the --skills JSON array of non-empty strings."""
+def parse_skills(raw_text: str) -> list[Skill]:
+    """Parse and validate the --skills JSON array of skill names or objects."""
     try:
         data = json.loads(raw_text)
     except json.JSONDecodeError as exc:
         raise UserError(f"malformed JSON in --skills input: {exc}") from exc
     if not isinstance(data, list):
         raise UserError(
-            "--skills input must be a JSON array of strings; "
+            "--skills input must be a JSON array of skill names or objects; "
+            f"got {type(data).__name__}"
+        )
+    skills: list[Skill] = []
+    for idx, item in enumerate(data):
+        where = f"--skills[{idx}]"
+        if isinstance(item, str) and item.strip():
+            skills.append(Skill(item, (item,)))
+            continue
+        if not isinstance(item, dict):
+            raise UserError(
+                f"{where} must be a non-empty string or an object with a "
+                f"non-empty `name`; got {item!r}"
+            )
+        name = item.get("name")
+        if not isinstance(name, str) or not name.strip():
+            raise UserError(
+                f"{where}.name must be a non-empty string; got {name!r}"
+            )
+        skills.append(Skill(
+            name=name,
+            terms=_dedupe((name, *_string_list(item, "aliases", where)), key=_norm),
+            keywords=tuple(sorted({
+                _norm(k) for k in _string_list(item, "keywords", where)
+            })),
+            languages=_languages(item, where),
+            docs=_dedupe(_string_list(item, "docs", where)),
+        ))
+    return skills
+
+
+def parse_technologies(raw_text: str) -> list[str]:
+    """Parse and validate the --technologies JSON array of names."""
+    try:
+        data = json.loads(raw_text)
+    except json.JSONDecodeError as exc:
+        raise UserError(f"malformed JSON in --technologies input: {exc}") from exc
+    if not isinstance(data, list):
+        raise UserError(
+            "--technologies input must be a JSON array of names; "
             f"got {type(data).__name__}"
         )
-    names: list[str] = []
     for idx, item in enumerate(data):
-        if not isinstance(item, str) or not item:
+        if not isinstance(item, str) or not item.strip():
             raise UserError(
-                f"--skills[{idx}] must be a non-empty string; got {item!r}"
+                f"--technologies[{idx}] must be a non-empty string; got {item!r}"
             )
-        names.append(item)
-    return names
+    return data
 
 
-def _cmd_comention(args: argparse.Namespace) -> int:
+def _read_doc_and_skills(args: argparse.Namespace) -> tuple[str, list[Skill]]:
     if args.doc == "-" and args.skills == "-":
         raise UserError(
             "--doc and --skills cannot both read from stdin ('-')"
@@ -278,27 +1374,102 @@ def _cmd_comention(args: argparse.Namespace) -> int:
     else:
         doc_text = _read_source(args.doc, "--doc")
         skills_text = _read_source(args.skills, "--skills")
-    skill_names = parse_skills(skills_text)
+    return doc_text, parse_skills(skills_text)
+
 
-    result = analyze(doc_text, skill_names)
+def _emit(result: dict) -> int:
+    json.dump(result, sys.stdout, indent=2)
+    sys.stdout.write("\n")
+    return 0
+
+
+def _cmd_comention(args: argparse.Namespace) -> int:
+    doc_text, skills = _read_doc_and_skills(args)
+    result = analyze(doc_text, skills)
     if args.verbose:
         print(
-            f"analyzed {len(set(skill_names))} distinct skill names; "
-            f"{len(result['pairs'])} qualifying pairs; "
-            f"{result['excluded_section_count']} excluded sections",
+            f"analyzed {len({s.name for s in skills})} distinct skill names; "
+            f"{len(result['pairs'])} candidate pairs; "
+            f"{result['list_only_pair_count']} list-only pairs; "
+            f"{result['excluded_section_count']} introductory sections",
             file=sys.stderr,
         )
-    json.dump(result, sys.stdout, indent=2)
-    sys.stdout.write("\n")
-    return 0
+    return _emit(result)
+
+
+def _cmd_mentions(args: argparse.Namespace) -> int:
+    if args.technologies == "-" and "-" in (args.doc, args.skills):
+        raise UserError(
+            "only one of --doc, --skills and --technologies can read from "
+            "stdin ('-')"
+        )
+    doc_text, skills = _read_doc_and_skills(args)
+    technologies = None
+    if args.technologies is not None:
+        technologies = parse_technologies(
+            _read_source(args.technologies, "--technologies")
+        )
+    result = mentions(doc_text, skills, technologies)
+    if args.verbose:
+        summary = (
+            f"analyzed {len(result['skills'])} distinct skill names; "
+            f"{len(result['mentioned'])} mentioned, "
+            f"{len(result['fenced_only'])} only in fenced code, "
+            f"{len(result['unmentioned'])} unmentioned; "
+            f"{len(result['candidates'])} candidate pairs; "
+            f"{result['list_only_pair_count']} list-only pairs"
+        )
+        if technologies is not None:
+            summary += (
+                f"; {len(result['technologies'])} technologies, "
+                f"{len(result['unverified_technologies'])} not covered"
+            )
+        print(summary, file=sys.stderr)
+    return _emit(result)
+
+
+def _cmd_infer(args: argparse.Namespace) -> int:
+    if args.top_k < 0:
+        raise UserError(f"--top-k must be >= 0; got {args.top_k}")
+    skills = parse_skills(_read_source(args.skills, "--skills"))
+    result = infer(skills, top_k=args.top_k)
+    if args.verbose:
+        print(
+            f"analyzed {len({s.name for s in skills})} distinct skill names; "
+            f"{result['total_pairs']} candidate pairs, "
+            f"{len(result['pairs'])} emitted; "
+            f"{result['language_only_pair_count']} language-only pairs",
+            file=sys.stderr,
+        )
+    return _emit(result)
+
+
+def _add_doc_and_skills(p: argparse.ArgumentParser, doc_help: str) -> None:
+    p.add_argument("--doc", required=True, help=doc_help)
+    p.add_argument(
+        "--skills",
+        required=True,
+        help=(
+            "path to a JSON array of loaded skills, or '-' for stdin. "
+            "Shape: [\"<skill>\" | {\"name\": \"<skill>\", "
+            "\"aliases\": [\"<term>\", ...]}, ...]"
+        ),
+    )
+    p.add_argument(
+        "--verbose",
+        action="store_true",
+        help="print a one-line summary to stderr",
+    )
 
 
 def _build_parser() -> argparse.ArgumentParser:
     parser = argparse.ArgumentParser(
         prog="skf-comention-pairs",
         description=(
-            "Compute compose-mode co-mention pairs from an architecture "
-            "document (detect-integrations.md §2 compose branch)."
+            "Report which loaded skills a document names and the pairs it "
+            "names together (comention: detect-integrations.md §2 compose "
+            "branch; mentions: refine-architecture scope sets), or candidate "
+            "pairs from the skills' own docs and keywords (infer)."
         ),
     )
     sub = parser.add_subparsers(dest="cmd", required=True)
@@ -306,29 +1477,68 @@ def _build_parser() -> argparse.ArgumentParser:
     p_cm = sub.add_parser(
         "comention",
         help=(
-            "emit qualifying co-mention pairs (>=2 body paragraphs, "
-            "word-boundary, section-filtered) as JSON"
+            "emit every candidate pair a body paragraph names in one unit "
+            "or through a lead-in (word-boundary), with evidence kinds and "
+            "introductory sections marked, and count the pairs only listed "
+            "together, as JSON"
         ),
     )
-    p_cm.add_argument(
-        "--doc",
-        required=True,
-        help="path to the architecture markdown document, or '-' for stdin",
+    _add_doc_and_skills(
+        p_cm, "path to the architecture markdown document, or '-' for stdin"
+    )
+    p_cm.set_defaults(func=_cmd_comention)
+
+    p_mn = sub.add_parser(
+        "mentions",
+        help=(
+            "emit which skills a document names (fenced code skipped), the "
+            "term that named each, the pairs a paragraph names together "
+            "and, with --technologies, which skills cover each name, as JSON"
+        ),
+    )
+    _add_doc_and_skills(
+        p_mn, "path to the markdown document, or '-' for stdin"
+    )
+    p_mn.add_argument(
+        "--technologies",
+        help=(
+            "path to a JSON array of technology names as the document "
+            "writes them, or '-' for stdin: adds which skills cover each "
+            "name and the names none covers"
+        ),
+    )
+    p_mn.set_defaults(func=_cmd_mentions)
+
+    p_in = sub.add_parser(
+        "infer",
+        help=(
+            "emit candidate pairs with no architecture document: one "
+            "skill's docs naming the other or shared domain keywords, never "
+            "a shared language alone, Top-K capped, as JSON"
+        ),
     )
-    p_cm.add_argument(
+    p_in.add_argument(
         "--skills",
         required=True,
         help=(
-            "path to a JSON array of loaded skill names, or '-' for stdin. "
-            "Shape: [\"<skill>\", ...]"
+            "path to a JSON array of loaded skills, or '-' for stdin. "
+            "Shape: [{\"name\": \"<skill>\", \"aliases\": [...], "
+            "\"keywords\": [...], \"language\": \"<language>\" | [...], "
+            "\"docs\": [\"<path>\", ...]}, ...]"
         ),
     )
-    p_cm.add_argument(
+    p_in.add_argument(
+        "--top-k",
+        type=int,
+        default=DEFAULT_TOP_K,
+        help=f"cap the output at the top N pairs (default {DEFAULT_TOP_K})",
+    )
+    p_in.add_argument(
         "--verbose",
         action="store_true",
         help="print a one-line summary to stderr",
     )
-    p_cm.set_defaults(func=_cmd_comention)
+    p_in.set_defaults(func=_cmd_infer)
 
     return parser