fix: the imperative detector only ever saw sentence starts

'When delay is suspected, obtain a detailed history' is as much an
instruction to the reader as 'Obtain a detailed history', and the first
pass could not see it — so the 819 I reported cleared was the count of
one kind. 71 more were buried mid-sentence, after a comma or a
conjunction. The detector now finds both, and export can be limited to
particular variants so a pass does not collide with one already running
over another view.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TqXevQJhxFrM7jJg82cgZN

backend/scripts/prose_pass.py

@@ -38,24 +38,38 @@ from app.models.article import Article
 
 MARKER = re.compile(r"\[\[(\d+)\|([^\]]+)\]\]")
 NUMBER = re.compile(r"\d+(?:\.\d+)?")
-#: Verbs that open a sentence or a bullet as an instruction to the reader.
+_VERBS = (r"Obtain|Order|Check|Give|Start|Administer|Consider|Assess|Evaluate|Perform|"
-IMPERATIVE = re.compile(
+          r"Measure|Repeat|Refer|Avoid|Ensure|Monitor|Treat|Begin|Initiate|Stop|Use|Send|"
-    r"(?:(?<=^)|(?<=[.!?] )|(?<=^- )|(?<=^\* )|(?<=\n- )|(?<=\n\* ))"
+          r"Screen|Look|Ask|Confirm|Rule out|Do not|Don't|Reserve|Counsel|Admit|Discharge|"
-    r"(Obtain|Order|Check|Give|Start|Administer|Consider|Assess|Evaluate|Perform|"
+          r"Suspect|Exclude|Document|Reassess|Titrate|Correct|Repair|Remove|Apply|Prescribe")
-    r"Measure|Repeat|Refer|Avoid|Ensure|Monitor|Treat|Begin|Initiate|Stop|Use|Send|"
+
-    r"Screen|Look|Ask|Confirm|Rule out|Do not|Don't|Reserve|Counsel|Admit|Discharge|"
+#: An instruction to the reader opening a sentence or a bullet.
-    r"Suspect|Exclude|Document|Reassess|Titrate|Correct|Repair|Remove|Apply|Prescribe)\b")
+IMPERATIVE_START = re.compile(
+    r"(?:(?<=^)|(?<=[.!?] )|(?<=^- )|(?<=^\* )|(?<=\n- )|(?<=\n\* ))(" + _VERBS + r")\b")
+
+#: And one buried mid-sentence, after a clause break — "…is suspected, obtain a
+#: detailed history". These are the ones the first pass missed entirely,
+#: because it only ever looked at what a sentence started with.
+IMPERATIVE_MID = re.compile(
+    r"(?:,|;|\band\b|\bthen\b|\bor\b)\s+(" + _VERBS + r")\b\s+"
+    r"(?:a|an|the|for|all|both|age|serum|blood|urine|IV|oral)\b", re.I)
+
+
+def imperatives(text: str) -> list[str]:
+    return IMPERATIVE_START.findall(text) + IMPERATIVE_MID.findall(text)
 
 MIN_RATIO, MAX_RATIO = 0.7, 1.4
 
 
-def affected(db):
+def affected(db, variants=None):
     """Every section carrying the instructional voice, with its article."""
     out = []
     for article in db.query(Article).order_by(Article.id).all():
         for section in article.sections or []:
+            if variants and (section.get("variant") or "long") not in variants:
+                continue
             content = section.get("content") or ""
-            hits = IMPERATIVE.findall(content)
+            hits = imperatives(content)
             if hits:
                 out.append({
                     "article_id": article.id,
@@ -100,7 +114,7 @@ def cmd_export(args) -> int:
     out.mkdir(parents=True, exist_ok=True)
     db = SessionLocal()
     try:
-        rows = affected(db)
+        rows = affected(db, args.variants.split(",") if args.variants else None)
     finally:
         db.close()
     if not rows:
@@ -191,6 +205,8 @@ def main() -> int:
     export = sub.add_parser("export")
     export.add_argument("--out", required=True)
     export.add_argument("--batches", type=int, default=8)
+    export.add_argument("--variants", default=None,
+                        help="Comma-separated variants to export, e.g. short,long")
     export.set_defaults(func=cmd_export)
 
     imp = sub.add_parser("import")