feat(server): add Poolside V1 parsers

[pedro-moser]

Important

This is a draft stacked on #102. Until #102 lands, the GitHub diff against main includes its Laguna GGUF commits. The parser change to review is the final three commits (f275318, c6f0067, and 5f7ad9d, seven files). I will rebase onto main after #102 merges.

Summary

• add poolside_v1 reasoning support for the implicit-open <think> ... </think> protocol
• add poolside_v1 tool-call support using Poolside/GLM-style <tool_call>, <arg_key>, and <arg_value> tags
• auto-select both parsers for Laguna/Poolside architecture and path markers
• expose poolside_v1 as an explicit CLI reasoning/tool parser choice
• mirror the embedded template's default-on reasoning mode while honoring explicit enable_thinking=false
• preserve GLM's schema-aware coercion for explicitly typed non-string arguments while retaining exact string whitespace

Motivation

#102 adds Laguna GGUF serving, but parser auto-selection currently falls through to unrelated generic formats. The model can load and generate while reasoning and tool calls are silently classified incorrectly.

The Poolside tool format is close to GLM 4.7, so this reuses that state machine rather than introducing a second parser implementation. The only Poolside-specific semantic difference is preserving leading/trailing whitespace for arguments whose JSON Schema declares exactly type: "string"; enum-only, untyped, numeric, boolean, array, and object arguments retain existing typed conversion behavior.

Closes #195 after #102 lands.

Validation

On current main plus #102:

PYTHONPATH=$PWD:$PWD/python pytest -q \
  tests/server/test_parser_auto_selection.py \
  tests/server/test_poolside_v1_parsers.py \
  tests/server/test_function_call_parser.py \
  tests/server/test_minimax_m3_parsers.py \
  tests/server/test_muse_glimmer_parsers.py \
  tests/server/test_reasoning_parser_all_models.py \
  tests/server/test_reasoning_parser_dsv4.py

254 passed in 1.55s

Coverage includes:

• buffered and arbitrary-chunk streaming
• escaped string content and chunk boundaries inside tags
• text after a tool call
• schema-declared string whitespace
• schema-aware GLM coercion for uppercase booleans, exponent numbers, Python-literal objects, and invalid integer representations
• enum-only and untyped schemas
• explicit CLI choices and Laguna auto-selection
• regression coverage for existing GLM behavior

A real unsloth/Laguna-S-2.1-GGUF revision 750f92f90cf54159c4d7a610cb7b3e74498e75c6 run auto-selected both parsers with no explicit flags. Buffered and streaming requests produced valid API tool-call transport whenever the checkpoint emitted a closed tool block.

Risk and exclusions

• Existing parser names and GLM behavior remain unchanged.
• The IQ1 checkpoint often emits tool blocks inside unclosed reasoning. Those samples correctly remain reasoning content; this PR does not try to repair model output.
• This PR does not change the Laguna loader, MoE cache, model template, or sampling defaults.