fix(identity): treat comma/period as OCR noise in squash

packages/flow/openadapt_flow/runtime/identity.py

@@ -364,8 +364,31 @@
 
 
 def squash(text: str) -> str:
-    """Lowercase and remove ALL whitespace (OCR-tolerant comparison form)."""
-    return "".join(text.lower().split())
+    """Lowercase and drop whitespace and OCR-volatile name punctuation.
+
+    Comma and period on letter tokens are stripped because thin identity-band
+    crops routinely make RapidOCR read ``Smith,`` as ``Smith.`` (and the
+    reverse). Leaving them in the comparison form turned a glyph-collapse
+    ABSTAIN into a false affirmative MISMATCH on native Windows while Mac
+    stayed consistent.
+
+    Periods that separate digits (``01.01.1900``) are kept: collapsing them
+    turns a date into a digit run that the glyph-vulnerable-identifier gate
+    then mistreats as an MRN-like token and spuriously abstains.
+    """
+    lowered = text.lower()
+    out: list[str] = []
+    for i, ch in enumerate(lowered):
+        if ch.isspace():
+            continue
+        if ch in ",.":
+            prev = lowered[i - 1] if i else ""
+            nxt = lowered[i + 1] if i + 1 < len(lowered) else ""
+            if prev.isdigit() and nxt.isdigit():
+                out.append(ch)  # date / decimal separator
+            continue
+        out.append(ch)
+    return "".join(out)
 
 
 def band_region(point: Point, band_height: int, viewport: tuple[int, int]) -> Region:

packages/flow/tests/test_identity.py

@@ -347,6 +347,23 @@ def test_tokenize(self):
         assert tokenize("  Jane   Li \n panel ") == ["jane", "li", "panel"]
         assert tokenize("") == []
 
+    def test_comma_period_name_punctuation_is_ocr_noise(self):
+        # Recorded thin-band context often stores "Smith." while the live
+        # band (or another engine) returns "Smith," — same patient row.
+        match = band_match(
+            "100512 Smith. John 01/15/1980 Active",
+            "100512 Smith, John 01/15/1980 Active",
+        )
+        assert match.coverage == 1.0
+        assert match.contradicted_chars == 0
+        assert match.unexplained_name_tokens == 0
+        # A real near-name sibling must still mismatch.
+        bad = band_match(
+            "100512 Smith. John 01/15/1980 Active",
+            "100512 Smith, Joan 01/15/1980 Active",
+        )
+        assert bad.contradicted_chars > 0 or bad.unexplained_name_tokens > 0
+
 
 class TestOperatingPoint:
     """Pin the ROC-chosen decision parameters and their boundaries
@@ -456,6 +473,19 @@ def test_coverage_ignores_scattered_short_blocks(self):
         assert coverage("abcdef", "abcdef") == 1.0
         assert coverage("", "x") == 0.0
 
+    def test_squash_drops_comma_period(self):
+        # Thin identity-band crops make RapidOCR read "Smith," as "Smith."
+        # (Windows native); both must compare equal so glyph abstain is not
+        # overwritten by a false punctuation mismatch.
+        assert squash("Smith,") == squash("Smith.")
+        assert squash("Smith, John") == "smithjohn"
+        assert squash("Jr.") == "jr"
+        # Date / digit separators must survive: collapsing "01.01.1900" into
+        # "01011900" makes the glyph-vulnerable-identifier gate spuriously
+        # abstain on a clean MRN + period-separated DOB pair (CI fuzz).
+        assert squash("01.01.1900") == "01.01.1900"
+        assert squash("01/01/1900") == "01/01/1900"
+
     def test_band_region_clamps_to_viewport(self):
         assert band_region((500, 10), 64, (1280, 800)) == (0, 0, 1280, 64)
         assert band_region((500, 795), 64, (1280, 800)) == (0, 736, 1280, 64)