chat : add MiniMax M2 specialized tool-call handler

[doctorjei]

Fixes autoparser response (failure/abort) to Minimax-M2.7 system prompt.

Overview

Adds specialized tool-call handler for MiniMax M2.7 template (and probably later versions, I'm guessing). Right now, without it, M2.7 output with tools crashes llama-server (GGML_ABORT) at src/llama-grammar.cpp:1435 (EOG with non-empty stack) when <invoke> is emitted.

Why? (Reproducing the Issue)

Reproducible in-tree via tests/test-chat.cpp on current master (b8840). Parallel <invoke> elements inside a <minimax:tool_call> wrapper confuse the peg_tester.

The autoparser (peg-native) infers grammar structure from the template via differential rendering. MiniMax's template uses XML with repeatable invoke elements for parallel calls. The parser correctly infers per-invoke structure but mis-specifies the repetition rule, so any second invoke is lost.

This is a regression; an earlier working version was in the mainline (#16932, 1920345) (via a generalized XML tool-call parser), but the autoparser refactoring (#18675) replaced it. This PR restores specialized handling for MiniMax M2.7 (and likely other M2 versions) without reverting the broader refactor.

Implementation

This implementation follows the Kimi K2 / DeepSeek V3.2 pattern for templates the autoparser cannot handle.

• common_chat_params_init_minimax prepares PEG for wrapper/invoke/param grammar (parallel calls).
• Reasoning is extracted (<think>…</think> blocks) ahead of tool calls.
• String-typed parameters are captured verbatim (tool_arg_string_value) to preserve embedded XML-style content; non-strings are reconstructed through JSON.
• Dispatch in common_chat_try_specialized_template requires three MiniMax-specific literals in template source (<minimax:tool_call>, <invoke name=, <parameter name=)

Testing

Extends the existing MiniMax block in tests/test-chat.cpp with five test cases.

Test Case	Purpose	Master (no fix)
Parallel <invoke> elements; two different tools	Reproduces crash pattern	Crashes (Invalid diff: now finding less tool calls!)
Parallel <invoke> elements; same tool twice	Additional variant of crash	Crashes (same)
String parameter with embedded <div><script>…</script></div>	Verifies tool_arg_string_value is verbatim	Passes
Multi-line string parameter (Python code with \n)	Verifies until("</parameter>") boundary on multi-line content	Passes
Two integer parameters in one <invoke>	Verifies zero_or_more over parameter list + non-string JSON reconstruction	Passes

The passing test cases are also focused on repetition (vs content shape) to provide additional regression coverage.

Additional information

• Precedent: chat: dedicated DeepSeek v3.2 parser + "official" template #21785 (DeepSeek V3.2 dedicated parser)
• Earlier working implementation: common: Generalized XML-style tool-call parsing with streaming support (GLM 4.5/4.6 + MiniMax M2 + SeedOSS + Kimi-K2 + Qwen3-Coder + Apriel-1.5 + Xiaomi-MiMo) #16932 (superseded by Autoparser - complete refactoring of parser architecture #18675)
• Crash assertion: src/llama-grammar.cpp:1435 (GGML_ABORT("fatal error") when EOG token is accepted with non-empty grammar stacks).

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES

AI was used to identify the appropriate strategy, draft a harness, and draft initial code snippets. Every line was reviewed, edited as appropriate, and included manually in commits.

[ggml-gh-bot]

Hi @doctorjei, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• Multiple open PRs from a new contributor: We limit new contributors (those without a previously merged PR) to 1 open PR at a time. You currently have 2 open PRs.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.