docs(cache): audit prompt-cache strategy batch 08

[rekram1-node]

Batch 08 (11 providers): impossibl, inception, inceptron, inference, inferx, infomaniak, io-net, iteracompute, jalapeno, jiekou, kenari.

Leading header comment only (# Prompt cache (chat) + # Docs) per provider.

• SUPPORTS (2): inception (automatic prefix cache, discounted cached-input rates, usage.prompt_tokens_details.cached_tokens), jalapeno (automatic prefix cache + optional prompt_cache_key, dedicated caching doc).
• NATIVE_OTHER (3): impossibl (automatic on most routes; Anthropic needs cache_control breakpoint — prior verdict verified), jiekou (explicit Anthropic cache_control + implicit OpenAI/Gemini — prior verdict verified), io-net (automatic prefix cache, no request fields; per-model supports_prompt_cache + cache read/write rates).
• TOLERATES (1): infomaniak (OpenAI-compatible chat API, no cache params/usage/rates published).
• UNKNOWN (5): inceptron (setup-only docs), inference (closed param table, no cache fields), inferx (no public API reference), iteracompute (no cache params; usage is prompt/completion/total only), kenari (site/docs unreachable).

Note: bun validate fails identically on the clean base (pre-existing AuthoredModelShape.deepPartial environment error); all 11 TOML files verified parseable.

[github-actions]

No actionable findings.

[github-actions]

Action items

• [high] [possible mistake] providers/jiekou/provider.toml:1 - Check: Cache-key strategy must match documented request fields. Why: The header says send prompt_cache_key, but the same line only cites cache_control as a tolerated OpenAI-chat extension and implicit OpenAI/Gemini prefix caching. That does not establish prompt_cache_key as a client control; patch 1 also treated this as Anthropic cache_control + implicit cache, not a cache-key API. Action: Reclassify to the correct key policy (likely strip prompt_cache_key unless docs explicitly define prompt_cache_key), and keep cache_control guidance under cache policy only.
• [medium] [possible mistake] providers/impossibl/provider.toml:1 - Check: send prompt_cache_key requires a documented/supported cache-key field, not mere unknown-field tolerance. Why: The rationale is that the gateway tolerates extra cache fields / ignores cache_control on non-Anthropic routes (no 400). Patch 1 stated there is no prompt_cache_key field. Tolerance of extras is not evidence that clients should send prompt_cache_key. Action: Verify Impossibl docs for an actual prompt_cache_key control; if absent, use strip (or another evidenced policy) and describe automatic/cache_control behavior only in cache policy.
• [medium] [possible mistake] providers/infomaniak/provider.toml:3 - Check: Cache policy claims must not contradict the provider’s own documented surface. Why: Patch 1 reported no published cache params, cached-token usage, or cached-input rates (TOLERATES). Patch 2 still strips prompt_cache_key for a closed schema, but asserts “automatic server-side prefix caching” and verification via usage.prompt_tokens_details.cached_tokens. Catalog models also lack cache_read/cache_write. Action: Confirm Infomaniak documents automatic prefix cache and cached_tokens; if not, rewrite cache policy to “no documented prompt cache / upstream-implicit only” (or cite the specific doc that supports the stronger claim).