JSON null and missing are not the same fact.

The difficult example: same empty cell, different source state

Record two has no a key. Record one has a: null. Record three has a concrete value and an explicit null in b. Under the default both-empty policy, the first two facts intentionally render alike, and preview reports that loss.

[
  {"a":null,"b":1},
  {"b":2},
  {"a":3,"b":null}
]

a,b
,1
,2
3,

Policy decision: switch Null values to “Literal null” when downstream consumers must distinguish explicit null from a missing key. The union schema still keeps both a and b as columns.

Why this matters downstream

• Analytics often treats explicit null as known absence and missing as a producer/schema problem.
• Empty CSV cells are convenient but lossy; the preview warning records that choice before download.
• The policy is conversion-wide and deterministic, so the same input does not change meaning between runs.