[taekop]

When a size error happens in summarise() or reframe() after an across() (or an unnamed data frame result) has been expanded, the error handler indexes dots with a position that counts the expanded columns instead of the original expressions:

tibble(x = 1:3, y = 4:6) |>
  reframe(across(everything(), \(x) rep(letters[x], each = x)))
#> Error in `names(dots)[[i]]`:
#> ! subscript out of bounds

The C side reports index_expression as a position in chunks, which has one entry per expanded column. summarise_cols() now records which expression each chunk came from and maps the index back before building the error context. With this the call above gives the usual recycling error (across(...) must be size 3 or 1, not 12), and the same goes for a bad expression that comes after an across(), e.g. summarise(df, across(everything(), mean), z = 1:2) now names z.

I ran a set of summarise/reframe calls (across, data frame results, NULL results, grouped, errors and successes) before and after. Results are identical; the error messages that changed are the ones that used to hit the subscript error, plus one that used to blame the wrong expression: after an across() that selects no columns, summarise(df, across(c(), mean), z = 1:2) used to name across(c(), mean) and now names z. There's no measurable slowdown on a 2000 column across().

The new tests are snapshot tests through summarise() and reframe(); they fail without the change. The full suite has the same failures on main without this change (all-equal, bind-rows, case-when, recode-values, which look like R 4.6 differences).

Closes #7657