feat(providers): log prompt cache tokens in LLM response debug output

[vmuliadi-astro]

Problem

The gateway logs only prompt_tokens / completion_tokens / total_tokens on its "LLM response" debug line. Providers like DeepSeek (via Cloudflare AI Gateway) report cache metadata in the same usage object:

"usage": {
  "prompt_tokens": 85,
  "completion_tokens": 10,
  "total_tokens": 95,
  "prompt_tokens_details": {"cached_tokens": 0},
  "prompt_cache_hit_tokens": 0,
  "prompt_cache_miss_tokens": 85
}

This metadata is parsed but then discarded: UsageInfo has no fields for it, and the debug log line never emits it. Any token-usage collector consuming the log therefore cannot report the real cache hit/miss split — it can only estimate (or treat cache as unknown).

Fix

1. pkg/providers/protocoltypes/types.go — extend UsageInfo with:

   • PromptCacheHitTokens *int (prompt_cache_hit_tokens, DeepSeek-style)
   • PromptCacheMissTokens *int (prompt_cache_miss_tokens)
   • PromptTokensDetails *PromptTokensDetails (prompt_tokens_details.cached_tokens, OpenAI-compatible style)

   Pointers distinguish "provider did not report" (nil) from a real zero (cold cache).

2. pkg/agent/pipeline_llm.go — surface those fields on the LLM response debug log line when present.

No provider behavior changes; purely additive logging. Verified against a live DeepSeek-via-AIG call and a full rebuild on linux/arm64.

[github-actions]

This PR has had no activity for 7 days and has been marked as stale. If you are still working on it, please push an update or leave a comment; otherwise it will be closed automatically in 7 days.