ColumnarDataFrame: all-None Option columns

🔗The old failure

ColumnarDataFrame::from_rows discovers column types by probing runtime values. When every row has None for an Option<T> field the probe never sees a Some, the column stays ColumnBuffer::Generic, and R received list(NULL, NULL, …) instead of an atomic vector with NA. Tibble and dplyr treat list(NULL, …) as a list-column — it cannot be compared to scalars, does not coerce cleanly, and appears as <list> rather than <lgl>/<int>/<dbl>/<chr> in str().

🔗The new behaviour

At assembly time, if a ColumnBuffer::Generic column has every entry as None, the column is emitted as an LGLSXP of length nrow filled with NA_logical_ rather than a VECSXP of NULL elements. This is the assembly-time downgrade. No user hint, schema annotation, or derive macro is involved.

The discriminator is in the buffer: Vec<Option<SEXP>> where push_na (pad for missing rows) stores None, and push_value(&None::<T>) serializes through RSerializer::serialize_none → returns SEXP::nil() → stores Some(SEXP::nil()). Both represent “no value” in the generic-list context. The downgrade checks v.iter().all(|e| e.is_none() || e.map_or(false, |s| s.is_nil())) — all entries are either missing or NULL. Only this condition fires the downgrade.

🔗The R coercion guarantee

R’s coercion rules make logical NA invisible downstream:

c(NA, 1L)      # integer NA + integer → integer vector
c(NA, "x")     # logical NA + character → character vector
c(NA, 3.14)    # logical NA + double → double vector

dplyr::bind_rows(), tibble::as_tibble(), mutate(), and coalesce() all coerce on contact. An all-NA logical column is indistinguishable from an all-NA typed column for everything users do downstream.

🔗When this is not what you want

In the rare case where you need a specific typed NA column (for example, R metadata systems that inspect the column type before any values arrive), use with_column to inject a typed NA vector explicitly after assembly:

use miniextendr_api::IntoR;

let na_integer = vec![Option::<i32>::None; nrow].into_sexp(); // INTSXP of NA_integer_
df.with_column("stored_size", na_integer)

This pattern is already described in the issue body for stored_size: Option<u64>.