docs(#2870): scope Kimi / OpenRouter / vLLM / SGLang as fourth-party …

[Skobeltsyn]

…— caching docs match ModelProvider enum

ChatGPT audit (#2865) of v0.6.5 flagged that docs/caching.md mentioned Kimi, OpenRouter, vLLM, and SGLang in the shipped-providers list while the ModelProvider enum has only OLLAMA / ANTHROPIC / OPENAI / DEEPSEEK. Trust products need unusually clean docs.

Investigation:

• ModelProvider.entries is { OLLAMA, ANTHROPIC, OPENAI, DEEPSEEK } (verified — single source of truth in ModelConfig.kt:15).
• Kimi / OpenRouter / vLLM / SGLang are real backends that consumers CAN reach by pointing the OpenAI adapter's openAiBaseUrl at them, but they're fourth-party deployments, not first-party adapters with their own wiring, caching tests, or live integration coverage.
• The May 2026 upstream-evaluation batch (per local memory) brought these names into the conversation but never landed as enum entries.

Fixed:

• docs/caching.md:5 (the lead paragraph) — removed Kimi / OpenRouter / vLLM / SGLang from the shipped-providers list. Added a forward reference to the new ## Under evaluation section.
• docs/caching.md per-provider behaviour table — collapsed from 5 rows back to 4 (the ModelProvider entries). Added a one-liner under the table calling out the single source of truth (ModelProvider. entries).
• docs/caching.md cache-reads bullet — dropped Kimi from the Anthropic / OpenAI / DeepSeek list.
• New ## Under evaluation section at the bottom documents Kimi, OpenRouter, vLLM, and SGLang as OpenAI-compatible endpoints reachable via openAiBaseUrl — with the explicit qualifier that they're NOT first-party adapters and that promoting any of them to ModelProvider is a separate ticket.

docs/premortem-0.5.0-streaming.md still mentions OpenRouter — that's a historical roadmap doc listing future provider work, not a shipped claim. Left alone.

No code changes. Full ./gradlew test detekt green.