fix(golden): never write Infinity or NaN into a snapshot file

Symptom: after a pull request replaced DIVIDE() with the / operator and the
denominator collapsed to zero, 'semantic-diff snapshot' wrote a golden file containing
'"value": -Infinity'. Python read it back without complaint, so every local check
passed, but the file is not valid JSON: node, jq and any other consumer reject it. A
snapshot the CI job cannot parse is worse than no snapshot, because the gate reports a
tooling error instead of the regression that caused it.

Root cause: json.dumps emits the non standard literals Infinity, -Infinity and NaN by
default, and compute_snapshot passed the raw float through from DuckDB.

Fix: non finite results are recorded as blank with an explicit reason, which the
comparator already treats as a value movement, so the regression is still caught.

Author: sandeep0430

src/semantic_diff/golden.py

@@ -125,7 +125,25 @@ def compute_snapshot(
                     name, False, reason=f"SQL execution failed: {exc}", sql=translation.sql
                 )
                 continue
-            value = None if raw is None else float(raw)
+            if raw is None:
+                snapshot.values[name] = MeasureValue(name, True, value=None, sql=translation.sql)
+                continue
+            value = float(raw)
+            if not math.isfinite(value):
+                # Infinity and NaN are not valid JSON. Writing them produced a snapshot
+                # file that Python could read back and every other parser rejected.
+                snapshot.values[name] = MeasureValue(
+                    name,
+                    True,
+                    value=None,
+                    reason=(
+                        f"evaluated to the non finite value {value!r}, recorded as blank so the "
+                        "snapshot stays valid JSON; the usual cause is a / operator dividing by zero"
+                    ),
+                    sql=translation.sql,
+                )
+                log.warning("non finite measure value", extra={"measure": name, "value": str(value)})
+                continue
             snapshot.values[name] = MeasureValue(name, True, value=value, sql=translation.sql)
         eval_seconds = time.perf_counter() - eval_start