Deferred-tools listing in the system prompt busts prompt cache on every MCP discovery / tool reveal #4777

Summary

The "Deferred Tools" listing (MCP tools — always shouldDefer=true, reachable only via ToolSearch) is baked into the cached system prompt. Whenever the deferred set changes — MCP progressive discovery completing after startChat(), or the model revealing a tool via ToolSearch — setTools() / startChat() call setSystemInstruction() to rewrite the whole system instruction. That mutates the cached prefix, invalidating the prompt cache for the rest of the conversation.

Impact

• Every late MCP discovery and every ToolSearch reveal drops the prompt-cache hit → higher latency + cost (cache_read → full input) for the remainder of the session.
• Affects both Anthropic (explicit cache_control on the system block) and Gemini (implicit prefix caching).
• Non-interactive --prompt runs are hit hardest (that's where progressive MCP discovery fires).

Root cause

getMainSessionSystemInstruction() embeds the deferred-tools section (getCoreSystemPrompt(..., deferredTools)), and setTools() rewrites the system instruction to refresh it on every registry change.

Proposed fix

Move the deferred-tools listing out of the cached prefix into a per-turn <system-reminder> in the message tail, rebuilt from live registry state each UserQuery/Cron turn. The system prefix then stays byte-stable across MCP discovery / tool reveals; deferred tools still surface, and remain visible across turns because the reminder persists in chat history.

Files involved:

• packages/core/src/core/client.ts
• packages/core/src/core/prompts.ts

A fix with tests is ready; happy to open a PR.