fix(extract): keep dotted numbers intact when segmenting a source

segment_source split on every `.`, so a version or decimal — 6.8.3, 3.14 —
was fractured across segment boundaries and the number-valued answer atom fell
out of every quotable span. on a synthetic lookup corpus that dropped ~11% of
the ground-truth facts before any budget was applied: the answer simply was
not present in any claim to retrieve.

a `.` flanked by digits is a decimal/version dot, never a sentence boundary,
so the segment regex now keeps it inside the span (sentence-ending periods are
unaffected). measured recall ceiling on that corpus rises from 89% to 100% of
facts — a cap that bounded every downstream compiler, not just this one.

Author: plind-junior

CHANGELOG.md

@@ -78,6 +78,15 @@ All notable changes to vouch are documented here. Format follows
   this is the selection step the compiler thesis needs: fewer, denser claims
   are what move accuracy-per-token against the grep baseline.
 
+### Fixed
+- **extraction no longer fractures dotted numbers.** `segment_source` split
+  on every `.`, so a version or decimal (`6.8.3`, `3.14`) was broken across
+  segment boundaries and its answer atom fell out of every span — measured at
+  ~11% of the ground-truth facts lost on a synthetic lookup corpus *before any
+  budget was applied*. a `.` flanked by digits is now kept inside the span
+  (sentence-ending periods are unaffected), lifting the recall ceiling of the
+  whole ingest pipeline from 89% to 100% of facts on that corpus.
+
 ## [1.5.0] — 2026-07-20
 
 ### Added

src/vouch/extract.py

@@ -28,8 +28,11 @@
 
 # Split on newlines and sentence-ending punctuation. Each match is kept as an
 # exact substring of the source so its receipt verifies; only surrounding
-# whitespace is stripped (the inner run stays contiguous in the source).
-_SEGMENT_RE = re.compile(r"[^\n.!?]+[.!?]?")
+# whitespace is stripped (the inner run stays contiguous in the source). A `.`
+# flanked by digits is a decimal/version dot ("6.8.3", "3.14"), never a
+# sentence boundary — keep it inside the span so the number-valued fact isn't
+# fractured out of every claim.
+_SEGMENT_RE = re.compile(r"(?:[^\n.!?]|(?<=\d)\.(?=\d))+[.!?]?")
 
 DEFAULT_MIN_CHARS = 16
 DEFAULT_MAX_CHARS = 320

tests/test_extract.py

@@ -40,6 +40,18 @@ def test_segment_source_splits_into_verbatim_spans() -> None:
         assert s in text
 
 
+def test_segment_source_keeps_dotted_numbers_intact() -> None:
+    # a period between two digits is a decimal/version dot, never a sentence
+    # end — splitting on it fractures the answer atom out of every span.
+    text = (
+        "The deployed build of Harbor Digest is version 6.8.3 today. "
+        "The measured value of pi is about 3.14 in this document."
+    )
+    segs = extract.segment_source(text)
+    assert any("version 6.8.3 today" in s for s in segs)
+    assert any("3.14" in s for s in segs)
+
+
 def test_segment_source_drops_short_noise_and_dupes() -> None:
     text = (
         "ok. The very same sentence appears here twice in a row now. "