[nebius] Audit reasoning controls

[rekram1-node]

Summary

• audit the 29 remaining Nebius Token Factory model entries without broad catalog refresh
• preserve the prior lifecycle fixes: remove 2 models disabled April 13, 2026 and mark the exact 11 June 22 shutdown models deprecated
• correct Qwen/Qwen3-235B-A22B-Instruct-2507 as non-reasoning
• add only positively supported reasoning controls

Resolved reasoning matrix

• effort low|medium|high: openai/gpt-oss-120b, openai/gpt-oss-120b-fast
• toggle: Qwen/Qwen3.5-397B-A17B, Qwen/Qwen3.5-397B-A17B-fast, zai-org/GLM-5
• fixed: none
• token budget: none

Nebius's GPT-OSS guide establishes adjustable effort and its API schema supplies the exact values. Nebius model guides establish hybrid modes for Qwen3.5 and GLM-5. Nebius flavor documentation establishes Base/Fast output parity.

Qwen/Qwen3-235B-A22B-Thinking-2507-fast intentionally omits options: separate Thinking/Instruct variants disprove a hybrid toggle but do not prove fixed reasoning or exclude other controls.

Credential-blocked unknown controls

• MiniMaxAI/MiniMax-M2.5-fast
• MiniMaxAI/MiniMax-M2.5
• NousResearch/Hermes-4-405B
• NousResearch/Hermes-4-70B
• nvidia/Nemotron-3-Nano-Omni
• nvidia/nemotron-3-super-120b-a12b
• Qwen/Qwen3-Next-80B-A3B-Thinking-fast
• Qwen/Qwen3-Next-80B-A3B-Thinking
• deepseek-ai/DeepSeek-V3.2
• deepseek-ai/DeepSeek-V3.2-fast
• deepseek-ai/DeepSeek-V4-Pro
• moonshotai/Kimi-K2.5-fast
• moonshotai/Kimi-K2.5

No Nebius credentials were available, so these 13 models retain reasoning = true with omitted reasoning_options.

Validation

• bun validate
• focused reasoning matrix assertion: 2 effort, 3 toggle, 0 fixed, 13 credential-blocked, 1 separately unresolved Thinking variant
• focused lifecycle assertion: exactly 11 deprecated models
• git diff --check
• Nebius-only scope check