feat(chat): Laguna reasoning + tool-call parser detector (CoT leaks into content)

[zentradev-rabih]

Problem

Serving Laguna-S-2.1 GGUFs via this fork, chain-of-thought is not separated —
it leaks into message.content instead of message.reasoning_content, even with
--reasoning-format deepseek --jinja.

Root cause: Laguna's laguna_glm_thinking chat template matches none of the
common_chat_params_init_* detectors in common/chat.cpp
(common_chat_try_specialized_template returns std::nullopt), so it falls back
to the generic format, which has no thinking-tag handling. --reasoning-format
sets the extraction mode, but the <think>/</think> tags are assigned per
template by a detector that doesn't exist for Laguna.

A subtlety that rules out reusing an existing detector: the assistant turn is
generated already inside a reasoning block. The template's reasoning directive
is prose ("Put your reasoning inside … tags"), so the model emits its thoughts and
then a literal </think> before the answer, with no emitted opening <think>.
Detectors that gate on tmpl.source().find("<think>") (e.g. LFM2) therefore never
enable extraction for Laguna.

Fix

Add common_chat_params_init_laguna:

• Force-opened reasoning: content up to the first </think> is the thought;
  the remainder is the answer. Gated on reasoning_format != NONE (not on a literal
  <think> in the template).
• <system> / <user> / <assistant> roles, generation prompt <assistant>\n.
• <tool_call>{"name":…,"arguments":…}</tool_call> tool calls via
  standard_json_tools.
• Dispatch keyed on two template-unique strings (render_assistant_messages_raw
  and <available_tools>) so it never shadows another family.

Also includes a one-line build fix: common/speculative.cpp uses std::isfinite
without #include <cmath>, which fails to compile on GCC 13.

Verification (Laguna-S-2.1-Q4_K_M, sm_121 / GB10 CUDA build)

• Reasoning prompt → reasoning_content populated, content is the clean answer
  (no CoT, no </think>), finish_reason: stop.
• Tool prompt → tool_calls: [get_weather({"city":"Paris"})], reasoning still
  separated, finish_reason: tool_calls.

Happy to adjust the dispatch markers or tool-call keys if you have canonical ones —
this is the interim serving path for us and we wanted to surface the gap so it can
be fixed properly upstream.

🤖 This fix was developed with the assistance of Claude Code.