[rmorozov]

Requested by Ruslan · project thread

Parity thread P5 of the 2026-10-24 batch: four small gaps, each with corpus cases, all checked against Jinja2 3.1.6 on Python 3.12.

Problem and resulting behaviour

• 0045: ordering over None and undefined. CompareForOrder let None and undefined operands through. Now min, max, dictsort (key and value) and groupby raise TypeError for None and UndefinedError for an undefined key under every Undefined class, as Python's < does. A single item still needs no comparison ([none]|min is None).
• 0295: select/reject argument binding. Tester now binds as select_or_reject: the attribute and the test name only by position, every other argument (keywords included) to the test, and none at all when there is no test. A missing attribute fails lazily at the first item. User-visible: selectattr(attribute='x') was a Jinja2C++-only spelling and is now an error, as in Python. The two unit rows that used it pass the attribute by position. Also, an undefined attribute name now finds nothing, as in Python, while None still means the item itself.
• 0291: fold by choice. Jinja2 folds a if test else b and and/or by the side a foldable test picks. The new Expression::PrepareFold does one Load walk and records, per operand, whether it folds always or by choice. IsFoldableAt evaluates those tests at render time; foldable tests are constants, so they have the same value Jinja2 folds. FinalizedExpressionRenderer (finalize ordering, 0288) and StaticOutputRenderer (volatile code, 0216) use it. A choice nested in a ~ chain or another composite is filed as 0307, pinned by autoescape.volatile_folded_choice_nested.
• 0041: \N{name} escapes. These are now decoded from a generated Unicode 15.0 name and alias table: src/unicode_names.h, from scripts/gen_unicode_names.py and scripts/unicode_name_aliases.txt, checked in parity-expectations.yml. Lookup matches CPython's: stored names and aliases in any ASCII case, computed CJK unified and Hangul names only as spelled. The table is about 450 KB of read-only data, with a sorted index so each lookup takes about 1 µs. New CMake option JINJA2CPP_WITH_UNICODE_NAMES (default ON, Ruslan's choice in the thread); with OFF, a \N{} escape is a parse error and the tests skip those expectations. detail::LookupUnicodeName is JINJA2CPP_EXPORT so that tests can decode escapes in shared builds; an added symbol only.
• Filed 0308, found by the verifier: 'a' is divisibleby 2 answers False where Python raises. Pinned by tests.divisibleby_string.

Validation

• CI on f039a0b (base 72ad3e6): all 42 checks green. That includes MSVC (3 configurations), AppleClang (4), both sanitizer jobs, the thread job, fuzz, Conan, the parity expectations check and the instruction counts against base. f039a0b only corrects the CMake option's help text (350 → 450 KB) on top of 5182a21, which was also 42/42 green.
• Local Debug GCC: ctest 5/5, parity 2092 cases with the new lines removed from divergences/{autoescape,literals}.txt. generate.py --check and gen_unicode_names.py --check pass under Python 3.12.
• All 430k spellings the generator knows (every name in upper, lower and title case, every alias, CJK variants) were compared with Python's unicode-escape codec: 0 mismatches. The verifier separately compared 163k + 59k names, including mutated ones: 0 mismatches.
• Perf verdict: scripts/preflight.sh --perf passes the instruction gate, and the perf track approved it (job 120893827560). The largest change is Load/strings at +0.24%; every Render benchmark is unchanged.
• The one extra allocation on Load/chat_mistral and Load/strings (+8/+4 B) is the Tester constructor's leadParams.posParams vector. It holds the attribute and test-name arguments split off for ParseParams, and is built once per select/selectattr call site at Load (selectattr("role", "equalto", "user") and select('ne', 'the') in those templates). It is not on the render path.
• A JINJA2CPP_WITH_UNICODE_NAMES=OFF Release build has ctest 5/5.
• A local merge of master a1e4165 (with Call protocol: *args/**kwargs in calls, Python argument order, callables as values (0220, 0294, 0214) #645) was clean and passed ctest 5/5. It was not pushed; the steward stacks it in the train.

Process

• Roles: main session (explore and implement), verifier (Sonnet) 1 round. It found two blocking items, a missing #include <exception> and test-line formatting; preflight had already caught both before any push. Its non-blocking findings were also fixed: basic_string<uint8_t> would break on libc++ 19, the OFF build failed tests, and selectattr(undefined) was wrong. It also checked ASan+UBSan with clang (5/5, clean), a shared build (links, 5/5), linear PrepareFold on deep chains, and that IsFoldableAt's std::exception catch loses no limit error, since the normal evaluation rethrows.
• Red CI pushes: 0 of 2.
• Codex: no result. The review and security review were requested at 15:01; the summary has said "running" ever since, with no findings and no 👍.
• Train note: Statement parity: self after a store, output after extends, iterating loop, extending includes (0211, 0203, 0219, 0298) #643 also touches src/filters.cpp and src/renderer.h.

🤖 Generated with Claude Code

https://claude.ai/code/session_015FP4qroXpXbNJdUJrvfJVQ

[rmorozov]

@codex review

@codex security review

---

Generated by Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	🔄 Running since 2026-10-24T15:01:37.021177Z	5182a21	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.