[DEMO, do not merge] Evidence: tool-call-parser top-tier + recipe links

[wjhrdy]

Throwaway PR to capture real GitHub Actions evidence for two changes just added to #107:

1. tool-call-parser conflicts now render top-tier ("Worth a look"), same as reasoning-parser.
2. Findings link to the actual recipe page as evidence.

Deliberately reintroduces both reasoning-parser: deepseek_r1 and tool-call-parser: hermes on Qwen/Qwen3.6-35B-A3B/performance/server.yml (both wrong, both should render top-tier with a working recipes.vllm.ai link).

Will push a follow-up commit reverting both once the check's output is confirmed, then close without merging (base branch ci/recipes-check predates the real fix, so this must never land).

[github-actions]

vllm-project/recipes check

No high-confidence conflicting values found on changed models.

Other observations (performance tuning, env vars, non-blocking)

• Qwen/Qwen3.6-35B-A3B -> Qwen/Qwen3.6-35B-A3B (match: exact)
  • enable-chunked-prefill [performance/default (hw=h100)]: ours=True vs recipe=None (extra_local)
  • gpu-memory-utilization [performance/default (hw=h100)]: ours=0.95 vs recipe=None (extra_local)
  • max-num-seqs [performance/default (hw=h100)]: ours=64 vs recipe=None (extra_local)
  • mm-encoder-tp-mode [performance/default (hw=h100)]: ours=None vs recipe='data' (missing_local)
  • enable-chunked-prefill [accuracy/default (hw=h100)]: ours=True vs recipe=None (extra_local)
  • gpu-memory-utilization [accuracy/default (hw=h100)]: ours=0.95 vs recipe=None (extra_local)
  • max-num-seqs [accuracy/default (hw=h100)]: ours=64 vs recipe=None (extra_local)
  • mm-encoder-tp-mode [accuracy/default (hw=h100)]: ours=None vs recipe='data' (missing_local)
  • upstream recipe requires vLLM >= 0.17.0

[wjhrdy]

Closing — evidence captured and linked from #107. Base branch predates the real reasoning-parser/tool-call-parser fix, so this must never merge.