feat(llm): make Anthropic prompt caching configurable

[OUAO-FRANK]

Summary

• add an llm.prompt_caching setting for Anthropic-compatible endpoints
• allow a provider-level prompt_caching override
• omit all cache_control breakpoints when prompt caching is disabled
• add resolver, configuration, and request-builder regression tests

This keeps prompt caching enabled by default for backward compatibility while allowing gateways that reject Anthropic's cache_control field to opt out.

Validation

• git diff --check passed
• Go tests could not be run locally because the Go toolchain is unavailable in this environment

The implementation and tests were reviewed manually. This PR was developed with GPT-5.6 through Codex, using Git and GitHub CLI; I reviewed the complete diff and validation results before submission.

Fixes #700

[github-actions]

🔍 OpenCodeReview found 1 issue(s) in this PR.

• ✅ Successfully posted inline: 1 comment(s)

[github-actions]

Critical bug: When promptCachingEnabled() returns false, params.System and params.Tools are never assigned, which means system prompts and tool definitions are silently dropped from the API request. Previously (before this change), these were always set when non-empty regardless of caching.

The fix should separate the cache-control mutation from the assignment. For example:

if len(systemBlocks) > 0 {
    if c.promptCachingEnabled() {
        systemBlocks[len(systemBlocks)-1].CacheControl = anthropic.NewCacheControlEphemeralParam()
    }
    params.System = systemBlocks
}
if len(tools) > 0 {
    if c.promptCachingEnabled() {
        tools[len(tools)-1].OfTool.CacheControl = anthropic.NewCacheControlEphemeralParam()
    }
    params.Tools = tools
    // ... tool choice logic
}

The same pattern applies to the messages block below — while messages are always set via the struct literal, the cache_control breakpoint on the last message should also be guarded without preventing any essential assignment.

[NanaseInori]

The resolver correctly distinguishes unset from false, and the client gates all three automatically added Anthropic cache breakpoints.

One behavior to adjust or clarify: applyCreateCustomProvider explicitly writes PromptCaching=true for every new Anthropic custom provider. With llm.prompt_caching=false, creating a provider through the TUI therefore re-enables caching, although the user never chose a provider-specific override. Leaving the pointer nil would inherit the global setting and still retain the historical enabled default when no setting exists.