schema that we'll then convert into schemalet (IR)

type

but--implicitly!--that schema must be a string. To handle this properly
required a new SchemaRef that's inforces the string constraint. What I
did has some kludginess, but doesn't break the overall architecture.

constructions let me to see that we're not really using the
schema-specific code. This is a start on the way there, but doesn't
really get there.

null, or either.

to do pattern matching on the input schema. I think there's also an
opportunity to examine the normalized schema as an additional point of
configuration, but this is pretty simple for now.

sequence into a tuple