Add SambaNova provider catalog

[rekram1-node]

Summary

• add SambaNova as an OpenAI-compatible provider
• catalog all 6 models returned by the official live SambaCloud /v1/models endpoint on 2026-06-10
• reconcile the 2026-06-09 removals, Preview lifecycle status, and SambaCloud-specific reasoning controls

Audit

• active models: 6 total (4 Production, 2 Preview)
• Preview models: DeepSeek-V3.2 and gemma-4-31B-it, both recorded with status = "beta"
• DeepSeek-V3.1: reasoning toggle established by SambaNova’s first-party SambaCloud launch guide using chat_template_kwargs.enable_thinking
• DeepSeek-V3.2: reasoning toggle established by a SambaNova Systems staff/admin response demonstrating reasoning: false against https://api.sambanova.ai/v1/chat/completions; this supports the repository’s abstract boolean toggle but does not assert that a single wire spelling is stable or exclusive
• gpt-oss-120b: exact low/medium/high reasoning effort established by SambaNova’s first-party SambaCloud launch guide
• MiniMax-M2.7 and gemma-4-31B-it: provider-agnostic reasoning capability is preserved as reasoning = true; reasoning_options is omitted because no SambaCloud-specific control semantics were positively established
• Meta-Llama-3.3-70B-Instruct: non-reasoning with no reasoning options
• generic Responses API effort support was not treated as model-specific evidence
• official unauthenticated SambaCloud model metadata endpoint was queried successfully for catalog, limits, and pricing

Validation

• bun validate passes
• git diff --check passes
• bun test packages/core/test/sambanova.test.ts passes (1/1, 10 assertions across all six models)
• bun test packages/core/test/generate.test.ts: 9/10 pass; only the pre-existing repository-wide missing weights-link inventory fails, listing 24 unrelated Sarvam/NVIDIA/Cohere models