Wouldn't this mean that a glob string py3.* would match py310 when it shouldn't have? i.e. we should first escape the existing . to \., and then replace * by .*?

Yes, I am trying to think of other cases convertions to regular expression might have changed (e.g. pattern with ?).

I will add more test cases.

Arf, I support your remark, but this clashes with behavior:

In conda, the "full regex mode" only kicks in if the build string is wrapped in ^...$. Otherwise, it's understood as a normal glob string. See docstring and implementation.

I believe this is the semantics we want to use, yet it clashes with ChimeraString's (see https://github.com/mamba-org/mamba/pull/3735/files#r1908971307).

I would find it clearer to have something like:

RegexSpec::RegexSpec(std::string raw_pattern)
    : m_raw_pattern(format_raw_pattern(std::move(raw_pattern))
    , m_pattern(std::regex(m_raw_pattern))
{
}

And have all the logic in format_raw_pattern (or any better name you can come with) rather than in the constructor, with a return statement.

I thought of this, but does this guarantee that m_raw_pattern is strictly initialized before m_pattern?

I am not sure it is safe as m_pattern is declared before m_raw_pattern:

Ah good catch, you have to invert them in the declaration. The intialization order is guanranteed to be the same as the declaration order.

Could this change the ABI?