feat: cache tool definitions and add prompt_caching config toggle

[emal-avala]

Summary

Completes prompt caching support by adding cache_control to tool definitions and exposing a user-facing config toggle.

What was already implemented (this PR does NOT change):

• System prompt cached with cache_control: { type: "ephemeral" }
• Conversation history breakpoints via messages_to_api_params_cached()
• Usage struct tracks cache_read_input_tokens and cache_creation_input_tokens
• /cost command shows cache hit percentage per model
• CacheTracker detects cache breaks via fingerprinting
• anthropic-beta: prompt-caching-2024-07-31 header

What this PR adds:

• Tool definition caching: cache_control: { type: "ephemeral" } on the last tool in the tools array, so the API caches the entire prefix (system + tools) as one block. With 32 tools (~15K tokens), this saves ~$0.003/turn on hits.
• Config toggle: features.prompt_caching (default: true) in config.toml. Users on providers without caching support can disable it to avoid unknown fields in requests.
• Wire config into request: query/mod.rs now reads the feature flag instead of hardcoding enable_caching: true.

Files changed

File	Change
crates/lib/src/llm/anthropic.rs	Add cache_control to last tool definition
crates/lib/src/llm/client.rs	Same in legacy client path
crates/lib/src/config/schema.rs	Add prompt_caching feature flag
crates/lib/src/query/mod.rs	Wire feature flag into ProviderRequest

Implements roadmap item 7.13.

Test plan

• cargo fmt --all -- --check — clean
• cargo clippy -- -D warnings — zero warnings
• cargo test — all tests pass
• Manual: verify tool definitions include cache_control in API request (debug log)
• Manual: verify prompt_caching = false in config.toml disables all cache_control markers
• Manual: verify /cost shows cache hit rate improvement after tools caching

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.