<regex>: Emit complete _N_str nodes only during NFA construction

[muellerj3]

This is a side quest to #5962: If the parser no longer generates _N_group nodes, the current node in the NFA is the node just before the group when parsing of the group contents begins. This might be an _N_str node and if the groups starts with some character sequence, the current logic will attach these characters to the _N_str node just before the group. This would effectively move these characters outside of the non-capturing group without the _N_group node, thus miscompiling the regex.

For this reason, this PR changes this character sequence pooling logic a bit: Characters now get added to a vector _Chars internal to the _Builder3 class, until the following subexpression in the regex results in the attachment of some other kind of node to the NFA or the parser must remember the current position in the NFA. At that point, the builder will emit the string node and then append the other kind of node to the string node or return a pointer to the string node. This means that string nodes are no longer modified after they have been generated.

Another nice bonus for the future: The parser does not really need the geometric growth logic inside the _Buf class anymore for _N_str nodes (although it is still used for now). I plan to rework the parsing logic for character classes as well and intend to remove all of that resize logic inside the _Buf class at that time.

The _Builder2 and _Parser2 classes are renamed because the layout of the builder class changed, so these classes must be renamed to avoid an ABI break. The version numbers of _Parser3::_Alternative() and _Builder3::_Else_if() are removed because the class renaming renders them unnecessary.

[muellerj3]

@microsoft-github-policy-service agree

[Copilot]

Pull request overview

Updates the <regex> NFA builder to defer creation of _N_str nodes until the builder must commit pending literal characters, preventing incorrect pooling of characters across group boundaries when _N_group nodes are no longer generated (per #5962 / GH-6289).

Changes:

• Reworked the regex NFA builder to buffer literal characters in _Builder3::_Chars and only emit _N_str nodes at commit points during construction.
• Renamed internal parser/builder types (_Parser2/_Builder2 → _Parser3/_Builder3) to avoid ABI breaks after layout changes, and simplified related method naming.
• Added targeted regression tests covering anchors, word boundaries, groups, lookahead, alternation, and captures.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
tests/std/tests/VSO_0000000_regex_use/test.cpp	Adds GH-6289 regression tests for string-node emission behavior across various regex constructs.
stl/inc/regex	Introduces _Builder3 buffering/emission logic for _N_str nodes and renames _Parser2/_Builder2 to _Parser3/_Builder3.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[muellerj3]

Good catch.

[StephanTLavavej]

/azp run STL-CI

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[muellerj2]

In hindsight, a string is the more appropriate container here.