Thanks for putting this together, @thgschwind, and apologies for the slow turnaround.

Revisiting this in light of the substrait v0.97.0 spec change (July 12): the premise that interval_day may be used without a precision no longer holds, so I don't think we can take this approach.

• Proto: v0.97.0 changed IntervalDay.precision's contract to "Implementations should reject an unset precision" (previously an unset precision was treated as 6/microseconds). This PR does the opposite — it produces a Type.IntervalDay with unset precision, i.e. exactly the message the spec now says to reject.
• Grammar: interval_day has no bare form in SubstraitType.g4 — the only alternative is interval_day<precision>. evaluate("interval_day") is actually a syntax error (mismatched input '<EOF>' expecting {'<','?'}); it only reaches the precision is None branch because ANTLR error recovery hands back a partial parse tree.

Revisiting my earlier question about whether this is IntervalDay-specific — it isn't. Every precision-parameterized type (decimal, interval_day, interval_compound, precision_time, precision_timestamp, precision_timestamp_tz) requires explicit parameters, so bare forms are all invalid in the same way.

Closing without merging, since defaulting to unset precision is now spec-noncompliant. The one real remaining problem — the confusing Unknown token type NoneType on invalid syntax — is worth fixing on its own and I've split it out into #216.