docs(providers): document prompt cache batch 19

[rekram1-node]

Batch 19 prompt-cache audit (11 providers). Header-only changes to provider.toml.

provider	verdict	mechanism
wandb	UNKNOWN	first-party OpenAI-compatible API; no cache docs or extra-field policy found
watsonx	NATIVE_OTHER	native wx.ai chat API via watsonx-ai-provider; no chat-level cache params documented
xai	NATIVE_OTHER	automatic prefix cache; x-grok-conv-id header steers chat hits, prompt_cache_key body key is Responses-only
xiaomi	NATIVE_OTHER	native cached-input billing (cache_read); no client-facing cache key documented
xiaomi-token-plan-ams	NATIVE_OTHER	same Xiaomi API family; cached-input billing, no client key
xiaomi-token-plan-cn	NATIVE_OTHER	same Xiaomi API family; cached-input billing, no client key
xiaomi-token-plan-sgp	NATIVE_OTHER	same Xiaomi API family; cached-input billing, no client key
xpersona	TOLERATES	OpenAI-compatible chat endpoint; schema lists model/messages/reasoning only, no documented cache control
zai	NATIVE_OTHER	automatic implicit context caching via usage.prompt_tokens_details.cached_tokens; no client key
zai-coding-plan	NATIVE_OTHER	same automatic context caching on coding chat endpoint; plan credits bill cached input
zeldoc	UNKNOWN	private LLM plus router; no cache params or cached-token fields documented, caching depends on downstream

bun validate passes.

[github-actions]

No actionable findings.

[github-actions]

Action items

• [medium] [possible mistake] providers/wandb/provider.toml:3 - Check: Provider cache policy must not contradict this provider’s published model costs. Why: The new header says there is no documented prompt cache and to treat usage as uncached, but many providers/wandb/models/** entries already publish non-zero cost.cache_read (e.g. DeepSeek V4 Flash at 0.07). That leaves the catalog internally inconsistent for clients using these comments. Action: Reconcile the policy line with existing cache_read pricing—either rephrase to automatic/undocumented hit billing without a client key (if prices are real) or verify and fix the model costs if cache is not actually billed.
• [medium] [possible mistake] providers/xpersona/provider.toml:4 - Check: Provider cache policy must not contradict this provider’s published model costs. Why: The new header says no documented cache and to treat usage as uncached, while multiple providers/xpersona/models/** entries define cost.cache_read (e.g. 0.15 on GPT-class models). Action: Reconcile the policy with those cache_read values—acknowledge billed/implicit cache hits if valid, or correct the model costs if cache pricing should not be listed.

[github-actions]

Action items

• [medium] [possible mistake] providers/wandb/provider.toml:1 - Check: Prompt-cache docs must stay consistent with catalog pricing/usage facts. Why: The new header says chat usage has “no cached-token field” and to “treat usage as uncached,” but this provider already publishes non-zero cost.cache_read on many models (e.g. DeepSeek V4 Pro, GPT-OSS, Qwen). That is an internal contradiction: either caching/cached billing exists and the comment overstates UNKNOWN, or the existing cache_read rows are wrong. Action: Reconcile the header with the model costs—acknowledge automatic/cached-input billing (and any usage field) if cache_read is correct, or remove/zero those cache_read values if caching truly is not billed/observable.
• [medium] [possible mistake] providers/xpersona/provider.toml:1 - Check: Prompt-cache docs must stay consistent with catalog pricing facts. Why: The new header says there is “no documented prompt cache” and to “treat usage as uncached,” while every Xpersona model already has a non-zero cost.cache_read (e.g. gpt-5.4, Claude Sonnet 4.6). Consumers get opposite signals from provider docs vs pricing. Action: Soften or correct the header to match real billing (e.g. automatic cache hits / priced cache_read with no client cache key), or drop the cache_read costs if caching is not actually offered.

[github-actions]

Action items

• [medium] [possible mistake] providers/wandb/provider.toml:1 - Check: Material factual claims in hand-authored provider docs must be consistent with the PR’s stated audit and supported by the cited source. Why: The PR body still classifies wandb as UNKNOWN with “no cache docs or extra-field policy found,” but the final header claims automatic prefix caching, a chat body field cache_salt, and https://docs.wandb.ai/inference/response-settings/prefix-caching. Those cannot both be right; if cache_salt is wrong, clients will send a non-existent control. Action: Verify the W&B prefix-caching docs for chat cache_salt and usage.prompt_tokens_details.cached_tokens, keep the TOML only if accurate, and update the PR body verdict/mechanism to match.
• [medium] [possible mistake] providers/xai/provider.toml:3 - Check: Prompt-cache header comments must distinguish request controls from response telemetry for this host’s chat API. Why: Final text puts hits under Use body: hits via usage.prompt_tokens_details.cached_tokens, which reads as a request-body cache field. The prior revision correctly said chat has no body cache key and that prompt_cache_key is Responses-only; dropping that can send callers looking for a chat body parameter that does not exist. Action: Restore the chat vs Responses split (headers / no chat body key / response cached_tokens), and only list real request body fields under Use body.
• [low] [possible mistake] providers/xiaomi/provider.toml:2 (same pattern on xiaomi-token-plan-*, zai, zai-coding-plan) - Check: Use body should document request-body cache controls, not only response usage. Why: These hosts are described as automatic prefix cache with no client cache key, yet Use body only names usage.prompt_tokens_details.cached_tokens (response). That blurs “how to request a hit” vs “how to observe a hit.” Action: Rephrase so request body is empty/none (or omitted under the positive-only rule) and hit detection is clearly response/usage-side.