refactor(pfs,llm,bib): one definition of the chunk-metadata array predicates (GIN-index expression), normalize_key for family keys everywhere, error-isolation docstring on code_cited_sources (refs #705)

notebooks/code_families.py

@@ -615,6 +615,7 @@ def _(code, mo, not_built, pl, store):
 
             from llm.config import load as _load_llm_cfg
             from llm.config import pg_url as _pg_url
+            from pfs.anchors import families_array_sql as _families_array_sql
 
             _eng = _sa_engine(_pg_url(_load_llm_cfg()))
             with _eng.begin() as _conn:
@@ -623,8 +624,7 @@ def _(code, mo, not_built, pl, store):
                         "SELECT c.name, count(*) AS n "
                         "FROM langchain_pg_embedding e "
                         "JOIN langchain_pg_collection c ON c.uuid = e.collection_id "
-                        "WHERE string_to_array(COALESCE(e.cmetadata->>'families', ''), "
-                        "' ') && ARRAY[:key] "
+                        f"WHERE {_families_array_sql('e')} && ARRAY[:key] "
                         "GROUP BY c.name ORDER BY c.name"
                     ),
                     {"key": _key},

src/bib/tag.py

@@ -271,16 +271,19 @@ class Tag(BaseModel):
         Parameters
         ----------
         key : str
-            The family key exactly as registered (e.g. ``CCM``,
-            ``APCM``, or a derived ``pfs.code_family`` key) — never
-            lower-cased, since derived keys aren't necessarily
-            uppercase-safe round trips.
+            The family key as registered (e.g. ``CCM``, ``APCM``, or a
+            derived ``pfs.code_family`` key); normalised through
+            ``pfs.families.normalize_key`` (stripped, upper-case — every
+            registry key is stored that way) so one spelling never
+            becomes two tags.
 
         Examples::
 
             Tag.family("CCM")   # family:CCM
         """
-        return cls(namespace="family", value=key)
+        from pfs.families import normalize_key
+
+        return cls(namespace="family", value=normalize_key(key))
 
     @classmethod
     def fn(cls, qualified_name: str) -> Tag:

src/cli/pfs.py

@@ -42,7 +42,13 @@ from pfs.codetables import (
 )
 from pfs.cpt_load import ingest as cpt_ingest
 from pfs.extract import extract_code
-from pfs.families import FAMILIES, HAND_FAMILIES, derive_families, refresh_from
+from pfs.families import (
+    FAMILIES,
+    HAND_FAMILIES,
+    derive_families,
+    normalize_key,
+    refresh_from,
+)
 from pfs.guidance import build as build_guidance
 from pfs.lineage import lineage, lineage_all
 from pfs.reaction import series as reaction_series
@@ -136,7 +142,7 @@ def _codes_for(
     if families:
         refresh_from(con)
         for family in families:
-            fam = FAMILIES.get(family.upper())
+            fam = FAMILIES.get(normalize_key(family))
             if fam is None:
                 # List only the hand families here (M3) — FAMILIES can hold
                 # ~15k derived keys once `refresh_from` has run, and that
@@ -475,7 +481,7 @@ def guidance(
     family's FR paragraphs and the CPT manual cite, with FR provenance."""
     if not family:
         raise typer.BadParameter("pass --family (repeatable)")
-    keys = [f.upper() for f in family]
+    keys = [normalize_key(f) for f in family]
     store = _store()
     if write:
         # Ruling A13 (mirrors `stack pfs reaction`): resolve every family
@@ -559,7 +565,7 @@ def _reaction_targets(
     refresh_from(con)
     targets: list[tuple[str, tuple[str, ...], str | None]] = []
     for f in family:
-        key = f.upper()
+        key = normalize_key(f)
         fam = FAMILIES.get(key)
         if fam is None:
             raise typer.BadParameter(

src/llm/evidence.py

@@ -24,6 +24,7 @@ import duckdb
 
 from llm.config import LlmConfig
 from llm.links import as_source
+from pfs.anchors import codes_array_sql, families_array_sql
 from pfs.codetables import is_missing_table_error
 from pfs.families import FAMILIES, Detection, detect_codes, refresh_from
 from pfs.valuation import ValuationRow, valuation
@@ -296,7 +297,7 @@ _CITED_SQL = (
     ") AS rn "
     "FROM langchain_pg_embedding e "
     "JOIN unnest(CAST(:codes AS text[])) AS w(code) "
-    "ON w.code = ANY(string_to_array(COALESCE(e.cmetadata->>'codes', ''), ' ')) "
+    f"ON w.code = ANY({codes_array_sql('e')}) "
     "WHERE e.collection_id = (SELECT uuid FROM langchain_pg_collection "
     "WHERE name = :collection) "
     ") t "
@@ -332,7 +333,7 @@ _FAMILY_CITED_SQL = (
     ") AS item_rn "
     "FROM langchain_pg_embedding e "
     "JOIN unnest(CAST(:families AS text[])) AS w(family) "
-    "ON string_to_array(COALESCE(e.cmetadata->>'families', ''), ' ') && ARRAY[w.family] "
+    f"ON {families_array_sql('e')} && ARRAY[w.family] "
     "WHERE e.collection_id = (SELECT uuid FROM langchain_pg_collection "
     "WHERE name = :collection) "
     ") one_per_item "
@@ -358,7 +359,7 @@ _DOCKET_CITED_SQL = (
     ") AS rn "
     "FROM langchain_pg_embedding e "
     "JOIN unnest(CAST(:codes AS text[])) AS w(code) "
-    "ON w.code = ANY(string_to_array(COALESCE(e.cmetadata->>'codes', ''), ' ')) "
+    f"ON w.code = ANY({codes_array_sql('e')}) "
     "WHERE e.collection_id = (SELECT uuid FROM langchain_pg_collection "
     "WHERE name = :collection) "
     ") t "

src/pfs/anchors.py

@@ -16,6 +16,32 @@ from pfs.elements import parse_descriptor
 from pfs.families import FAMILIES, Family, find_codes
 
 
+#: The chunk-metadata arrays every code/family predicate reads. ONE
+#: definition (#705 item 4): the ``&&`` GIN index on pgvector was built on
+#: exactly ``string_to_array(coalesce(cmetadata->>'families',''),' ')`` and a
+#: predicate that spells it differently silently loses the index.
+def codes_array_sql(alias: str = "e") -> str:
+    """SQL for the chunk's ``codes`` metadata as a text[] (space-joined)."""
+    return f"string_to_array(COALESCE({alias}.cmetadata->>'codes', ''), ' ')"
+
+
+def families_array_sql(alias: str = "e") -> str:
+    """SQL for the chunk's ``families`` metadata as a text[] (space-joined)."""
+    return f"string_to_array(COALESCE({alias}.cmetadata->>'families', ''), ' ')"
+
+
+def code_or_family_predicate(
+    alias: str = "e", *, codes_param: str = ":codes", family_param: str = ":family_key"
+) -> str:
+    """``(codes && CAST(:codes AS text[])) OR (families && ARRAY[:family_key])``
+    — the two-way match the chat's cited-source windows and the reaction
+    series share."""
+    return (
+        f"({codes_array_sql(alias)} && CAST({codes_param} AS text[])) "
+        f"OR ({families_array_sql(alias)} && ARRAY[{family_param}])"
+    )
+
+
 def code_family_index(families: Mapping[str, Family]) -> dict[str, tuple[str, ...]]:
     """Code -> sorted tuple of every qualifying family key, built once so
     ``anchor_metadata`` doesn't scan every family for every chunk (a

src/pfs/families.py

@@ -28,6 +28,14 @@ _FR_CITE_RE = re.compile(r"\b\d{1,3}\s+FR\s+\d{3,6}\b", re.IGNORECASE)
 FR_CITE_RE = _FR_CITE_RE
 
 
+def normalize_key(key: str) -> str:
+    """The canonical form of a family key: stripped, upper-case. Every
+    registry key (hand and derived) is stored this way; ``Tag.family``,
+    the ``stack pfs`` commands and the chat all normalise through here
+    (#705 item 5) so one spelling never becomes two tags or two families."""
+    return (key or "").strip().upper()
+
+
 def find_codes(text: str) -> tuple[str, ...]:
     """Sorted unique codes literally present in *text*, upper-cased.
 

src/pfs/reaction.py

@@ -37,6 +37,7 @@ from typing import Any, Callable, Sequence
 
 from sqlalchemy import text
 
+from pfs.anchors import code_or_family_predicate
 from pfs.descriptors import rule_year_of
 from pfs.families import code_pattern, codes_in
 
@@ -56,8 +57,7 @@ _DOCKET_COUNTS_SQL = text(
     "count(DISTINCT item_key) AS n_total "
     "FROM ( "
     "SELECT e.cmetadata->>'docket' AS docket, e.cmetadata->>'item_key' AS item_key, "
-    "(string_to_array(COALESCE(e.cmetadata->>'codes', ''), ' ') && CAST(:codes AS text[])) "
-    "OR (string_to_array(COALESCE(e.cmetadata->>'families', ''), ' ') && ARRAY[:family_key]) "
+    f"{code_or_family_predicate('e')} "
     "AS matched "
     "FROM langchain_pg_embedding e "
     "WHERE e.collection_id = (SELECT uuid FROM langchain_pg_collection WHERE name = :collection) "
@@ -83,8 +83,7 @@ _STANCE_SAMPLE_SQL = text(
     "WHERE e.collection_id = (SELECT uuid FROM langchain_pg_collection WHERE name = :collection) "
     "AND e.cmetadata->>'docket' = :docket "
     "AND ( "
-    "(string_to_array(COALESCE(e.cmetadata->>'codes', ''), ' ') && CAST(:codes AS text[])) "
-    "OR (string_to_array(COALESCE(e.cmetadata->>'families', ''), ' ') && ARRAY[:family_key]) "
+    f"{code_or_family_predicate('e')} "
     ") "
     ") one_per_item "
     "WHERE item_rn = 1 "

tests/bib/test_codetags.py

@@ -28,6 +28,7 @@ class TestTagFactories:
 
     def test_family_label(self) -> None:
         assert Tag.family("CCM").label == "family:CCM"
+        assert Tag.family(" ccm ").label == "family:CCM"  # normalised (#705)
 
 
 class TestItemCodesFromChunks:

tests/pfs/test_anchors.py

@@ -73,3 +73,33 @@ def test_single_code_family_never_stamped():
     assert (
         anchor_metadata("code 54321", families=fams, code_index=index)["families"] == ""
     )
+
+
+class TestSharedPredicateSql:
+    """#705 item 4: one definition of the metadata-array expressions the
+    GIN index was built on; every predicate composes from it."""
+
+    def test_arrays_match_the_index_expression(self):
+        from pfs.anchors import codes_array_sql, families_array_sql
+
+        assert codes_array_sql("e") == (
+            "string_to_array(COALESCE(e.cmetadata->>'codes', ''), ' ')"
+        )
+        assert families_array_sql("x") == (
+            "string_to_array(COALESCE(x.cmetadata->>'families', ''), ' ')"
+        )
+
+    def test_predicate_composes_both_sides(self):
+        from pfs.anchors import code_or_family_predicate
+
+        p = code_or_family_predicate("e")
+        assert "&& CAST(:codes AS text[])" in p and "&& ARRAY[:family_key]" in p
+        assert p.startswith("(string_to_array(COALESCE(e.cmetadata->>'codes'")
+
+    def test_consumers_use_the_shared_text(self):
+        import llm.evidence as ev
+        import pfs.reaction as rx
+        from pfs.anchors import code_or_family_predicate, codes_array_sql
+
+        assert codes_array_sql("e") in ev._CITED_SQL
+        assert code_or_family_predicate("e") in str(rx._DOCKET_COUNTS_SQL)