SiliconFlow: migrate existing models to base_model

[rekram1-node]

Summary

• migrate 39 duplicated SiliconFlow global/China model TOMLs to canonical base_model metadata
• retain provider overrides and necessary base_model_omit paths while inheriting canonical core facts
• leave 62 full TOMLs without canonical metadata unchanged
• leave 55 regional/catalog symlinks and 2 pre-existing base_model TOMLs unchanged

This is a prerequisite for #2136 and contains none of that PR's catalog or reasoning audit changes. The authored TOMLs are compressed by removing duplicated fields, while resolved generated models continue to contain their evidenced core metadata. The only intentional generated-data change is adding canonical knowledge to 25 migrated models whose original provider TOMLs lacked it.

Audit

• retained model paths audited: 158
• authored regular TOMLs: 103
• migrated: 39
• ineligible because canonical metadata is absent: 62
• already using base_model: 2
• symlinks retained unchanged: 55
• test paths changed: 0
• provider.toml files changed/added: 0
• resolved family coverage: 39/39, no gaps
• resolved release_date coverage: 39/39, no gaps
• resolved last_updated coverage: 39/39, no gaps
• resolved knowledge coverage: 30/39 (5 retained provider values and 25 inherited canonical values)
• forbidden core facts in base_model_omit: 0

The 9 resolved knowledge gaps have no value in either the original provider TOML or canonical metadata:

siliconflow-cn/Pro/MiniMaxAI/MiniMax-M2.1
siliconflow-cn/Pro/MiniMaxAI/MiniMax-M2.5
siliconflow-cn/Pro/zai-org/GLM-5.1
siliconflow-cn/Pro/zai-org/GLM-5
siliconflow/MiniMaxAI/MiniMax-M2.1
siliconflow/MiniMaxAI/MiniMax-M2.5
siliconflow/zai-org/GLM-5.1
siliconflow/zai-org/GLM-5
siliconflow/zai-org/GLM-5V-Turbo

25 models receiving inherited canonical knowledge

siliconflow-cn/Pro/deepseek-ai/DeepSeek-R1: 2024-07
siliconflow-cn/Pro/moonshotai/Kimi-K2-Thinking: 2024-08
siliconflow-cn/Pro/moonshotai/Kimi-K2.5: 2025-01
siliconflow-cn/Pro/moonshotai/Kimi-K2.6: 2025-01
siliconflow-cn/Pro/zai-org/GLM-4.7: 2025-04
siliconflow-cn/deepseek-ai/DeepSeek-V4-Pro: 2025-05
siliconflow-cn/stepfun-ai/Step-3.5-Flash: 2025-01
siliconflow/Qwen/Qwen2.5-VL-72B-Instruct: 2024-04
siliconflow/Qwen/Qwen3-235B-A22B: 2025-04
siliconflow/Qwen/Qwen3-32B: 2025-04
siliconflow/Qwen/Qwen3-Coder-30B-A3B-Instruct: 2025-04
siliconflow/Qwen/Qwen3-Coder-480B-A35B-Instruct: 2025-04
siliconflow/Qwen/Qwen3-Next-80B-A3B-Instruct: 2025-04
siliconflow/Qwen/Qwen3-Next-80B-A3B-Thinking: 2025-04
siliconflow/deepseek-ai/DeepSeek-R1: 2024-07
siliconflow/moonshotai/Kimi-K2-Thinking: 2024-08
siliconflow/moonshotai/Kimi-K2.5: 2025-01
siliconflow/moonshotai/Kimi-K2.6: 2025-01
siliconflow/stepfun-ai/Step-3.5-Flash: 2025-01
siliconflow/zai-org/GLM-4.5-Air: 2025-04
siliconflow/zai-org/GLM-4.5: 2025-04
siliconflow/zai-org/GLM-4.5V: 2025-04
siliconflow/zai-org/GLM-4.6: 2025-04
siliconflow/zai-org/GLM-4.6V: 2025-04
siliconflow/zai-org/GLM-4.7: 2025-04

62 intentionally retained full TOMLs

siliconflow/openai/gpt-oss-120b.toml
siliconflow/openai/gpt-oss-20b.toml
siliconflow/meta-llama/Meta-Llama-3.1-8B-Instruct.toml
siliconflow/nex-agi/DeepSeek-V3.1-Nex-N1.toml
siliconflow/tencent/Hunyuan-MT-7B.toml
siliconflow/tencent/Hunyuan-A13B-Instruct.toml
siliconflow/inclusionAI/Ling-flash-2.0.toml
siliconflow/inclusionAI/Ring-flash-2.0.toml
siliconflow/inclusionAI/Ling-mini-2.0.toml
siliconflow/Qwen/Qwen2.5-72B-Instruct-128K.toml
siliconflow/Qwen/Qwen2.5-14B-Instruct.toml
siliconflow/Qwen/Qwen2.5-Coder-32B-Instruct.toml
siliconflow/Qwen/Qwen3-235B-A22B-Instruct-2507.toml
siliconflow/Qwen/Qwen3-VL-235B-A22B-Instruct.toml
siliconflow/Qwen/Qwen3-Omni-30B-A3B-Captioner.toml
siliconflow/Qwen/Qwen3-VL-235B-A22B-Thinking.toml
siliconflow/Qwen/Qwen3-30B-A3B-Instruct-2507.toml
siliconflow/Qwen/Qwen3-14B.toml
siliconflow/Qwen/Qwen2.5-72B-Instruct.toml
siliconflow/Qwen/Qwen2.5-7B-Instruct.toml
siliconflow/Qwen/Qwen2.5-VL-7B-Instruct.toml
siliconflow/Qwen/Qwen3-VL-30B-A3B-Thinking.toml
siliconflow/Qwen/Qwen3-VL-30B-A3B-Instruct.toml
siliconflow/Qwen/Qwen2.5-VL-32B-Instruct.toml
siliconflow/Qwen/Qwen3-8B.toml
siliconflow/Qwen/Qwen3-30B-A3B-Thinking-2507.toml
siliconflow/Qwen/QwQ-32B.toml
siliconflow/Qwen/Qwen3-235B-A22B-Thinking-2507.toml
siliconflow/Qwen/Qwen3-VL-32B-Thinking.toml
siliconflow/Qwen/Qwen3-VL-8B-Thinking.toml
siliconflow/Qwen/Qwen3-Omni-30B-A3B-Instruct.toml
siliconflow/Qwen/Qwen3-VL-32B-Instruct.toml
siliconflow/Qwen/Qwen2.5-32B-Instruct.toml
siliconflow/Qwen/Qwen3-VL-8B-Instruct.toml
siliconflow/Qwen/Qwen3-Omni-30B-A3B-Thinking.toml
siliconflow/deepseek-ai/DeepSeek-V3.2-Exp.toml
siliconflow/deepseek-ai/DeepSeek-V3.1.toml
siliconflow/deepseek-ai/DeepSeek-V3.toml
siliconflow/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B.toml
siliconflow/deepseek-ai/DeepSeek-V3.2.toml
siliconflow/deepseek-ai/deepseek-vl2.toml
siliconflow/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B.toml
siliconflow/deepseek-ai/DeepSeek-V3.1-Terminus.toml
siliconflow/THUDM/GLM-4-32B-0414.toml
siliconflow/THUDM/GLM-4-9B-0414.toml
siliconflow/THUDM/GLM-Z1-32B-0414.toml
siliconflow/THUDM/GLM-Z1-9B-0414.toml
siliconflow/baidu/ERNIE-4.5-300B-A47B.toml
siliconflow/moonshotai/Kimi-K2-Instruct-0905.toml
siliconflow/moonshotai/Kimi-K2-Instruct.toml
siliconflow/ByteDance-Seed/Seed-OSS-36B-Instruct.toml
siliconflow-cn/Qwen/Qwen3.5-9B.toml
siliconflow-cn/Qwen/Qwen3.5-4B.toml
siliconflow-cn/deepseek-ai/DeepSeek-OCR.toml
siliconflow-cn/ascend-tribe/pangu-pro-moe.toml
siliconflow-cn/Pro/deepseek-ai/DeepSeek-V3.toml
siliconflow-cn/Pro/deepseek-ai/DeepSeek-V3.2.toml
siliconflow-cn/Pro/deepseek-ai/DeepSeek-V3.1-Terminus.toml
siliconflow-cn/Pro/moonshotai/Kimi-K2-Instruct-0905.toml
siliconflow-cn/Kwaipilot/KAT-Dev.toml
siliconflow-cn/PaddlePaddle/PaddleOCR-VL-1.5.toml
siliconflow-cn/PaddlePaddle/PaddleOCR-VL.toml

Validation

• bun compare:migrations (reports only the 25 approved canonical knowledge additions)
• bun validate
• git diff --check
• local before/after semantic audit across all 39 resolved models: no removed fields, no changed pre-existing values, and no additions except the 25 canonical knowledge fields listed above
• explicit core-fact audit against original provider and canonical metadata: all evidenced family, release_date, last_updated, and knowledge values resolve correctly