[shikokuchuo]

Closes #36. See the issue for the full design history and motivation.

crate() no longer requires .fn to be defined inside the call. Existing functions are handled by type:

• Plain closures are re-environed into the crate's evaluation environment, equivalent to the previous crate(rlang::set_env(fn)) workaround.
• Namespace functions and primitives are wrapped in a thin closure that delegates to them, leaving their original environment untouched.
• Existing crates are passed through unchanged.
• Supplying ... data that would be unreachable from the resulting crate (existing crate, namespace function, or primitive) is an error.

The factory-closure pitfall (e.g. stats::ecdf() results losing captured state) is documented with an example rather than guarded heuristically, as discussed in #36.

[hadley]

This approach is fine with me. I assume we should do a revdep pass for purrr just in case any CRAN package relies on the existing behaviour for function factories?

[shikokuchuo]

Revdeps for both carrier and purrr come back clean. I checked purrr using cloud_check(revdep_packages = cran_revdeps("purrr")) from this branch.