docs(providers): prompt cache batch 03

[rekram1-node]

Documents prompt-cache (chat) strategy for batch 03 providers. Header-only changes (leading comment block above first key); no schema fields touched. bun validate passes.

Provider	Verdict	Mechanism	Docs
arcee	TOLERATES	Request schema omits prompt_cache_key with no additionalProperties:false, so unknown key is accepted but ignored.	https://docs.arcee.ai/api-reference/chat-completion
atomic-chat	UNKNOWN	Local OpenAI-compatible server documents no prompt_cache_key or caching behavior; backend-dependent.	https://github.com/AtomicBot-ai/Atomic-Chat
auriko	SUPPORTS	Explicit prompt_cache_key field with OpenAI-compatible passthrough.	https://docs.auriko.ai/api-reference/chat-completions
azure	SUPPORTS	Automatic prefix caching plus prompt_cache_key routing hint (retention/breakpoints on newer models).	https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/prompt-caching
azure-cognitive-services	SUPPORTS	Same Azure OpenAI prompt-caching surface: automatic prefix caching plus prompt_cache_key.	https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/prompt-caching
bailing	UNKNOWN	No prompt_cache_key or prompt-caching behavior documented.	https://alipaytbox.yuque.com/sxs0ba/ling/intro
baseten	REJECTS	Strict ChatCompletionRequest (additionalProperties:false) with no prompt_cache_key; unknown key fails validation.	https://docs.baseten.co/reference/inference-api/chat-completions
berget	TOLERATES	Schema omits prompt_cache_key but permits extras (additionalProperties:{}), so unknown key is accepted without effect.	https://api.berget.ai/openapi.json
blueclaw	UNKNOWN	OpenAI-compatible gateway docs list chat/streaming/tools with no cache param or schema.	https://blueclaw.network/how-it-works.html
bothub	UNKNOWN	No prompt_cache_key or prompt-caching behavior documented for the OpenAI-compatible endpoint.	https://bothub.ru/models
cerebras	SUPPORTS	Automatic prefix caching plus optional prompt_cache_key routing hint (max 1024 chars, longer values rejected with 400).	https://inference-docs.cerebras.ai/capabilities/prompt-caching

Verification notes:

• Cerebras prompt_cache_key limit (1024 chars, 400 if longer) confirmed in docs.
• Baseten REJECTS confirmed via OpenAPI (ChatCompletionRequest additionalProperties:false, no prompt_cache_key).
• Berget TOLERATES confirmed via openapi.json (no prompt_cache_key anywhere; ChatCompletionRequest additionalProperties:{}).
• Auriko SUPPORTS confirmed via OpenAPI (explicit prompt_cache_key: "Prompt caching identifier. Supported by OpenAI.").
• Azure SUPPORTS confirmed via MS prompt-caching guide (automatic caching + prompt_cache_key on GPT-5.6+).

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.