Support Jinja2 implicit adjacent string-literal concatenation in minja

[jiafatom]

Problem

Jinja2 (like Python) implicitly concatenates adjacent string literals, e.g. {{ "foo" "bar" }} renders as foobar. minja's parseConstant only consumed a single string literal, so it threw Expected closing parenthesis in call args whenever a template placed two string literals next to each other.

This breaks real, valid templates. For example google/gemma-4-E2B-it's chat_template.jinja splits a long raise_exception(...) message across adjacent string literals:

{{- raise_exception("chat_template: tool_calls[].function.arguments must be a "
                    "JSON object (mapping), not a string. Deserialize arguments "
                    "before passing to the template.") -}}

minja::Parser::parse() fails at that line, so the entire template is rejected and OrtxApplyChatTemplate (used by onnxruntime-genai) becomes unusable for the model.

Before (parsing the Gemma-4 template):

PARSE FAILED: Expected closing parenthesis in call args at row 259, column 29

After: PARSE OK.

Fix

In parseConstant, after reading a string literal, loop and merge any run of consecutive string literals into a single constant — matching Jinja2/Python semantics. parseString() already skips leading whitespace and returns nullptr when the next token is not a string literal, so the loop terminates naturally and does not affect any other expression form.

Test

Adds OrtxTokenizerTest.AdjacentStringLiteralConcatenation, covering single-line ("Hello, " "world" '!'), cross-newline, and mixed-quote concatenation. Also verified existing behavior is unaffected (explicit ~ concat, numbers, conditionals) via a standalone render harness.

[Copilot]

Pull request overview

This PR updates the minja template parser to match Jinja2/Python behavior by implicitly concatenating adjacent string literals (e.g., {{ "foo" "bar" }} → foobar). This fixes parsing of real-world Jinja templates that split long strings across multiple adjacent literals, enabling OrtxApplyChatTemplate to handle such templates successfully.

Changes:

• Extend minja::Parser::parseConstant() to merge runs of consecutive string literals into a single constant value.
• Add a tokenizer/chat-template regression test that validates same-line, newline-separated, and mixed-quote adjacent literal concatenation.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
test/pp_api_test/test_tokenizer_chat.cc	Adds a regression test covering implicit adjacent string-literal concatenation via OrtxApplyChatTemplate.
shared/api/minja.hpp	Implements Jinja2-style implicit concatenation by looping parseString() calls and appending consecutive literals.

---

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