feat(mcp): choose tool exposure automatically / 自动选择 MCP 工具暴露模式

[SivanCola]

Summary / 摘要

• Choose configured MCP exposure automatically at session boot; no new config or environment option.
• Keep small, cache-valid surfaces direct for low-latency concrete calls.
• Route large surfaces (at least 16 cached tools), oversized schema surfaces (16 KiB floor, scaled with the model context window), and missing/stale caches through the existing stable use_capability proxy.
• Freeze the choice for the session. Cache-miss discovery still runs in the background through a private registry, so it warms diagnostics and cache state without adding dynamic mcp__* tools to the provider-visible registry.

在会话启动时自动决定配置 MCP 使用直接工具还是 use_capability 代理，不增加用户设置项；小型且缓存命中的工具面保持直连，大型、schema 过大或缓存未知的工具面自动使用稳定代理，并在整个会话内固定选择。

Required impact declarations

Documentation-impact: updated - documented automatic MCP exposure selection in English and Chinese.

Cache-impact: medium - large or unknown MCP surfaces receive a one-time stable proxy prefix; the decision is fixed for the session.

Cache-guard: go test ./internal/boot -run 'TestChooseMCPExposure|TestBuildAutomaticallyUsesCapabilityForLargeMCPSurface|TestBuildKeepsSmallMCPSurfaceDirect'

System-prompt-review: Codex review completed - no system-prompt text changes; only boot-time provider tool selection changes.

Why

Large MCP catalogs can consume a substantial provider tool prefix. A cache miss can also complete its handshake after boot and mutate the visible tool list. The automatic boundary keeps the ordinary path zero-configuration while preserving direct calls where their schema cost is small.

This adapts the MCP surface-collapse direction proposed by @bfxh in #7743. It intentionally does not include the user-facing meta-tool toggle or the unrelated long-horizon/navigation changes from that PR.

Refs #7743

Compatibility and cache impact

• Small configured MCP surfaces with matching caches keep the existing direct mcp__* behavior.
• Large or unknown configured surfaces receive one stable use_capability schema instead of many MCP schemas.
• The exposure decision is boot-scoped, so reconnects cannot change it mid-session.
• Explicit host-session MCP servers remain direct and session-scoped.
• Economy mode keeps its existing on-demand connector behavior.
• Upgrading can cause a one-time provider cache-prefix change when an existing MCP setup crosses the automatic boundary.

Verification

• go test ./...
• go test -race ./internal/boot
• go vet ./internal/boot ./internal/capability ./internal/plugin ./internal/agent
• Integration coverage verifies large surfaces expose only use_capability, small surfaces retain direct tools, and cold/stale caches choose the stable proxy while preserving background discovery diagnostics.