fix(llm,pfs): I5/M1/M2 — probe the model pool once per classifier, strip FR citations before lineage pattern matching, drop unused extract_text param

I5: `closed_vocab_classifier` called `pool.check(cfg.instruct_model)` on
every `classify()` call — a host-liveness probe, not per-request work.
A code with hundreds of element lines re-probed the pool hundreds of
times. Move the check to the closure's construction.

M1: `fr_events` matched event patterns (crosswalk, replaces, …) against
the raw FR paragraph text, so a citation like "91 FR 99490" inside a
"crosswalk ... 91 FR 99490" sentence read as a real crosswalk event for
99490 — the citation's page number is not a mention of the code. Strip
FR citations with the now-public `pfs.families.FR_CITE_RE` before both
the presence check and every event pattern.

M2: `extract_text`'s `classify` parameter was unused; no call site
passed it. Drop it.

Claude-Session: https://claude.ai/code/session_01Aum3pEMAM3yQVdFSdVe6Gc

src/llm/classify.py

@@ -56,11 +56,15 @@ def _default_post(url: str, json: dict) -> dict:
 def closed_vocab_classifier(
     cfg: Any, pool: Any, *, post: Callable[..., dict] | None = None
 ) -> Classifier:
-    """Return ``classify(text, choices) -> slug | None`` bound to *pool*."""
+    """Return ``classify(text, choices) -> slug | None`` bound to *pool*.
+
+    ``pool.check`` runs once here, not per call (I5) — it's a host-liveness
+    probe, not per-request work, and a code with hundreds of element lines
+    used to re-probe the pool hundreds of times."""
     send = post or _default_post
+    pool.check(cfg.instruct_model)
 
     def classify(text: str, choices: Sequence[str]) -> str | None:
-        pool.check(cfg.instruct_model)
         with pool.acquire_generation() as host:
             model = pick_model(cfg, pool, host)
             data = send(

src/pfs/extract.py

@@ -80,9 +80,7 @@ def extract_run(
     return Extraction(run.code, tuple(ordered), tuple(reviews))
 
 
-def extract_text(
-    code: str, text: str, *, year: int, source: str, classify: Classifier | None = None
-) -> Extraction:
+def extract_text(code: str, text: str, *, year: int, source: str) -> Extraction:
     rows = [
         ElementRow(code, year, e.type.value, e.value, e.detail, text, "", 0, 0, source)
         for e in parse_descriptor(text)

src/pfs/lineage.py

@@ -19,7 +19,7 @@ from typing import Any
 
 from pfs.codetables import EventRow
 from pfs.descriptors import rule_year_of
-from pfs.families import find_codes
+from pfs.families import FR_CITE_RE, find_codes
 
 RVU_KINDS = (
     "appeared",
@@ -199,6 +199,11 @@ def fr_events(store: Any, code: str) -> list[EventRow]:
     out: list[EventRow] = []
     for r in rows:
         text = _expand_ranges(r["text"])
+        # An FR page citation's number ("91 FR 99490") is not a mention of
+        # the code — strip it before both the presence check and every
+        # event pattern below (M1: a "crosswalk ... 91 FR 99490" sentence
+        # used to read as a crosswalk event for 99490).
+        text = FR_CITE_RE.sub(" ", text)
         if code not in text.upper():
             continue
         year = rule_year_of(r["title"], r["date_published"])

tests/llm/test_classify.py

@@ -93,3 +93,14 @@ class TestClassifier:
             post=lambda url, json: {"message": {"content": "maybe consent?"}},
         )
         assert classify("x", ["consent"]) is None
+
+    def test_pool_check_runs_once_not_per_call(self):
+        # I5: pool.check is a host-liveness probe, done once when the
+        # classifier closure is built — not per classify() call.
+        pool = _Pool()
+        classify = closed_vocab_classifier(
+            _Cfg(), pool, post=lambda url, json: {"message": {"content": "consent"}}
+        )
+        classify("a", ["consent"])
+        classify("b", ["consent"])
+        assert pool.checked == ["qwen2.5:14b"]

tests/pfs/test_lineage.py

@@ -189,6 +189,23 @@ class TestFrEvents:
         kinds = {e.kind for e in fr_events(store, "99441")}
         assert {"deleted", "bundled", "telehealth_list"} <= kinds
 
+    def test_fr_citation_page_number_is_not_a_crosswalk_mention(self, store):
+        # M1: an FR page citation ("91 FR 99490") is not a real mention of
+        # 99490 — it must be stripped before pattern matching, or a
+        # "crosswalk ... 91 FR 99490" sentence reads as a crosswalk event.
+        store.con.execute(
+            "INSERT INTO fr_anchors VALUES (?,?,?,?,?)",
+            (
+                "DE2VH9PD",
+                1500,
+                67999,
+                1500,
+                "We note the crosswalk methodology discussed at 91 FR 99490 for related codes.",
+            ),
+        )
+        ev = fr_events(store, "99490")
+        assert not any(e.kind == "crosswalk" for e in ev)
+
     def test_deleted_via_range_prefilter_middle_member(self, store):
         # 90868 is a middle member of a "90867-90869" range and appears
         # nowhere else in the paragraph literally — the SQL prefilter must