[siliconflow] Audit models and reasoning controls

[rekram1-node]

Summary

• audit all 103 pre-change SiliconFlow global and China model entries against SiliconFlow official API docs, model pages, and dated release notes
• remove 37 global models whose official deprecation dates have passed and add 10 active global models missing from the catalog
• add exact provider-exposed reasoning controls to every retained reasoning model and correct provider-specific reasoning flags

Reasoning controls

• thinking_budget: all 34 reasoning models, normalized as budget_tokens with official range 128-32768
• enable_thinking: 11 retained models explicitly named by SiliconFlow, normalized as toggle
• Hy3 preview modes: no_think, think_low, think_high, normalized as effort values none, low, high
• fixed/non-configurable: none
• non-reasoning: 40 models, no reasoning options

Evidence

• https://docs.siliconflow.com/en/api-reference/chat-completions/chat-completions
• https://docs.siliconflow.com/en/userguide/capabilities/reasoning
• https://docs.siliconflow.com/en/release-notes/overview
• official SiliconFlow model pages under https://www.siliconflow.com/models

Validation

• bun validate passes
• git diff --check passes
• focused core tests: 43 pass, 1 unrelated pre-existing failure in the repository-wide open-weight metadata test (24 Sarvam/NVIDIA/Cohere models missing weight links)
• live API testing not run because neither SiliconFlow credential is present

Timing

zai-org/GLM-4.7 is retained because its announced deprecation is June 11, 2026; this audit was performed June 10, 2026.