feat: Add MiniMax M3 tool-calling, reasoning, and conformance coverage

[indrajit96]

Description

This adds MiniMax M3 support to the frontend parser crates and conformance matrix.

MiniMax M3 uses a namespace-token XML-ish tool-call format:

]<]minimax[>[<tool_call>
]<]minimax[>[<invoke name="get_weather">
]<]minimax[>[<location>NYC]<]minimax[>[</location>
]<]minimax[>[</invoke>
]<]minimax[>[</tool_call>

It also uses <mm:think>...</mm:think> reasoning markers, where the opener can
be prompt-prefilled by the chat template.

Changes

• Added minimax_m3 / minimax-m3 tool-call parser registration.
• Added MiniMaxM3ParserConfig and ParserConfig::MiniMaxM3.
• Added a MiniMax M3 parser for:
  • namespace-prefixed tool-call blocks
  • <invoke name="..."> calls
  • parameter tags where tag names are argument names
  • duplicate parameter tags as arrays
  • schema-based scalar coercion
  • nested MiniMax XML object/array payloads
  • EOF recovery for complete invokes with a missing outer close
• Added minimax_m3 / minimax-m3 reasoning parser registration.
• Added opt-in streaming dangling-end recovery for MiniMax M3 reasoning, so
  reason</mm:think>answer is parsed as reasoning plus answer when the opener
  was prompt-prefilled.
• Added MiniMax M3 conformance coverage:
  • tool-calling batch fixtures
  • tool-calling stream-v2 inventory fixtures
  • reasoning batch fixtures
  • reasoning stream fixtures
  • Top-N table/registry metadata
• Wired MiniMax M3 into the vLLM/SGLang conformance wrappers, with pinned peer
  engines marked unavailable where they do not yet expose M3 parsers.

Notes

• TC stream-v2 coverage for MiniMax M3 is table/inventory coverage only. It marks
  dynamo_rust, vllm_rust, vllm_python, and sglang_python unavailable for
  now.
• This PR does not add a real parsers_v2 MiniMax M3 streaming parser.

Validation

Rust parser coverage

Targeted MiniMax M3 parser tests:

cargo test --locked -p dynamo-parsers minimax_m3 -- --nocapture

This covers:

• MiniMax M3 reasoning parser registration for both minimax_m3 and
  minimax-m3.
• Batch reasoning extraction for visible <mm:think>...</mm:think> spans.
• Streaming reasoning extraction across split chunks.
• Prompt-prefilled reasoning opener behavior:
  reason</mm:think>answer.
• Dangling end-marker recovery in batch and streaming paths.
• Close-marker-only streaming case where the close marker is stripped and does
  not leak into normal text.
• MiniMax M3 tool-call parsing for the namespace-token XML format.
• EOF recovery for a complete <invoke> when the outer tool_call close is
  missing.
• Streaming jail start detection regression:
  normal text ending in ] or ]< does not trigger MiniMax M3 tool-call mode,
  while ]<] and the full opener still do.

Live E2E validation

In addition to the repo-local parser/conformance tests above, these
frontend-crates changes were built into the downstream Dynamo vLLM runtime image
and exercised in a live K8s deployment. Used vllm/vllm-openai:minimax-m3 vllm base container.

[github-actions]

📊 Conformance matrix rendered — view in CI summary

[keivenchang]

Reviewed the M3 work. Locally at 56447bd: 628 parser tests pass (12 new), all 3 parity suites pass, clippy clean.

]<]minimax[>[ is verified real -- MiniMax M3 namespace token 200058; checked the upstream chat_template.jinja (params encoded as tag names), so the config.rs M2-difference comment is accurate. The shared BasicReasoningParser opt-in recover_dangling_end is the right pattern -- extends the parent, no fork, no regression.

Pushed 14eb3bc: the HTML highlighter was rendering ]<]minimax[>[ as a red orphan tag before every structural tag; now muted (.tt-ns) in static + streaming.

Inline notes below. Nothing blocking.

[rmccorm4]

LGTM after open comments addressed

[indrajit96]

Merging, as all opens addressed.