Error informatively when xobs_only() arguments are named or unknown

[joethorley]

Closes #111.

A named argument inside xobs_only() failed with an internal stopifnot() from xcast() and, when it was the only argument, a dplyr cross-join deprecation warning attributed to newdata:

d <- tibble::tibble(a = 1:5 + 0.5, b = factor(letters[1:5]))
xnew_data(d, xobs_only(z = b))
#> Error in xcast(out, .data = .data):
#>   all(names(values) %in% names(.data)) is not TRUE

Approach

Names are rejected rather than applied.
xobs_only() preserves the combinations observed in .data, and a new column has no observed combinations, so a rename has nothing well defined to join on for a bare vector such as z = new_seq(b).
Naming the whole call, xnew_data(d, z = xobs_only(b)), continues to create a new column as #102 documents.

• xobs_only() errors when any argument is named, naming the offending arguments.
• semi_crossing() errors when a resulting column is not in .data, which also covers an unnamed bare vector, instead of falling through to an empty by and a cross join.
• The ... documentation states that arguments must be unnamed variables in .data.

xnew_data(d, xobs_only(z = b))
#> Error: `xobs_only()` arguments must not be named ('z') as observed combinations must refer to columns of `.data`.
xnew_data(d, xobs_only(new_seq(b, .length_out = 2)))
#> Error: `xobs_only()` arguments must refer to columns of `.data` (unrecognised: 'new_seq(b, .length_out = 2)').

Checks

• R CMD check: 0 errors, 0 warnings, 0 notes.
• New test-xobs-only.R covers the observed-combination behaviour and snapshots the four error cases.

🤖 Generated with Claude Code