feat(metrics): per-call-site cache hit rate tracking

[Siddhant-K-code]

Closes #47

What

Adds CallSiteTracker to pkg/metrics. Records Anthropic API usage per call site and surfaces the worst performers first, making it easy to identify which agents or code paths are getting poor cache hit rates.

Changes

pkg/metrics/callsite.go (new)

• CallSiteTracker: thread-safe tracker backed by sync.RWMutex
• Record(callSite, UsageRecord): accumulates CacheCreationTokens, CacheReadTokens, UncachedInputTokens, OutputTokens; increments CacheHitRequests when cache_read_input_tokens > 0
• HitRate(): cache_read / (cache_read + cache_creation + uncached_input)
• WriteEfficiency(): cache_read / cache_creation — values < 1.0 mean writes that expire before being read
• RequestHitRate(): fraction of requests with at least one cache read token
• AllStats(): all call sites sorted by hit rate ascending (worst first)
• Summary(): human-readable table of call site, hit%, efficiency, request count
• Reset / ResetAll for test isolation

pkg/metrics/callsite_test.go (new) — 7 tests

Usage

tracker := metrics.NewCallSiteTracker()

tracker.Record("agent/planner.go:84", metrics.UsageRecord{
    CacheCreationInputTokens: resp.Usage.CacheCreationInputTokens,
    CacheReadInputTokens:     resp.Usage.CacheReadInputTokens,
    InputTokens:              resp.Usage.InputTokens,
})

// Worst performers first
for _, s := range tracker.AllStats() {
    fmt.Printf("%-40s hit=%.0f%%  eff=%.1fx  reqs=%d\n",
        s.CallSite, s.HitRate()*100, s.WriteEfficiency(), s.TotalRequests)
}