feat(implement): rewrite reasoning-parser recipe with token discovery guide

[slin1237]

Summary

Rewrite the reasoning-parser recipe from a vague description into a parameterized template with a concrete token discovery guide referencing vLLM and SGLang.

Problem

The old recipe told you to "check model card or API docs for reasoning delimiters" — but reasoning tokens live in different places depending on the model. Contributors had to rediscover this every time.

Changes

• skills/implement/reasoning-parser.md (321 insertions, 44 deletions):
  • Token discovery guide — 6-step priority order:
    1. Check vLLM (17 models) and SGLang first — tokens already validated
    2. HuggingFace tokenizer_config.json → added_tokens_decoder
    3. HuggingFace tokenizer_config.json → chat_template
    4. HuggingFace model card
    5. Provider API docs
    6. Test request observation
  • Reference links to actual GitHub source:
    • vLLM: https://github.com/vllm-project/vllm/tree/main/vllm/reasoning
    • SGLang: https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/parser/reasoning_parser.py
  • Known token tables: 9 in SMG + 8 in vLLM/SGLang not yet in SMG
  • Parameterized template: Complete Rust parser with 7 tests
  • initial_in_reasoning guide: Check add_generation_prompt or observe first token

Test Plan

• All GitHub links resolve to correct files
• Token table covers all existing SMG parsers
• Template compiles when substituted with real values