fix(serve): drop question/filler stopwords from query terms

`graphify query "<question>"` tokenised the whole question and seeded BFS on
every word, so natural-language scaffolding dominated retrieval. "how does the
frontier cache work" seeded on "how"/"the"/"work" — which prefix-match prose
labels like "Working Principles" at the 100x prefix tier — instead of on
"frontier"/"cache", landing in the wrong part of the graph.

Filter a set of English question/filler words from the query terms so content
words drive seeding, with a fallback to the unfiltered terms when a query is all
stopwords ("how does it work"). Applied to query terms only — node text is never
filtered, so a symbol literally named `work` stays findable via explain/path.

Updates the one test that pinned "what" as a kept term and adds coverage for the
new drop + all-stopword fallback. Full suite green (2745 passed, 28 skipped).

graphify/serve.py

@@ -107,8 +107,29 @@ def _is_searchable(term: str) -> bool:
     return True
 
 
+# English question/filler words dropped from query terms so content words drive
+# BFS seeding. Without this, "how does the frontier cache work" seeds on "how"/
+# "the"/"work" (which prefix-match prose labels like "Working Principles" at 100x)
+# instead of "frontier"/"cache", and lands in the wrong part of the graph. Applied
+# to query terms only — node text is never filtered, so a symbol literally named
+# `work` stays findable via explain/path. `work`/`works`/`working` are included
+# because "how does X work" / "how X works" is the most common question phrasing.
+_QUERY_STOPWORDS = frozenset({
+    "how", "what", "why", "when", "where", "which", "who", "whom", "whose",
+    "does", "did", "is", "are", "was", "were", "be", "been", "being",
+    "can", "could", "should", "would", "will", "shall", "may", "might", "must",
+    "has", "have", "had", "the", "and", "but", "not", "for", "from", "with",
+    "without", "into", "onto", "off", "that", "this", "these", "those", "there",
+    "here", "its", "their", "them", "they", "about", "any", "all", "some",
+    "work", "works", "working",
+})
+
+
 def _query_terms(question: str) -> list[str]:
-    """Split a query into searchable terms, segmenting Chinese text."""
+    """Split a query into searchable terms, segmenting Chinese text, then drop
+    English question/filler words (`_QUERY_STOPWORDS`) so content words drive
+    seeding. Falls back to the unfiltered terms if the query is all stopwords, so
+    a question like "how does it work" still seeds on something."""
     terms: list[str] = []
     for raw in question.split():
         if _has_chinese(raw):
@@ -121,7 +142,8 @@ def _query_terms(question: str) -> list[str]:
             for tok in re.findall(r"\w+", raw.lower()):
                 if _is_searchable(tok):
                     terms.append(tok)
-    return terms
+    content = [t for t in terms if t not in _QUERY_STOPWORDS]
+    return content or terms
 
 
 _EXACT_MATCH_BONUS = 1000.0

tests/test_serve.py

@@ -262,7 +262,20 @@ def test_trigram_index_cached_and_rebuilt_per_graph():
 
 
 def test_query_terms_strips_search_punctuation():
-    assert _query_terms("what calls extract?") == ["what", "calls", "extract"]
+    # "what" is a question stopword (dropped); punctuation is still stripped from "extract?".
+    assert _query_terms("what calls extract?") == ["calls", "extract"]
+
+
+def test_query_terms_drops_question_stopwords():
+    # Natural-language question words are dropped so content words drive seeding:
+    # "how does the frontier cache work" must reduce to the content terms, or it
+    # seeds on "how"/"the"/"work" (which prefix-match prose labels) instead.
+    assert _query_terms("how does the frontier cache work") == ["frontier", "cache"]
+
+
+def test_query_terms_all_stopwords_falls_back_to_unfiltered():
+    # An all-stopword query keeps its terms rather than seeding on nothing.
+    assert _query_terms("how does it work") == ["how", "does", "work"]
 
 
 def test_query_terms_filters_only_short_english_terms(monkeypatch):