[kclaka]

A primary hit with a shadow miss never populated the shadow cache; primary misses also replaced existing shadow hits. This biased the advertised same-traffic policy comparison. Populate shadow misses from primary hits and inspect shadow entries without counting another request or changing eviction state before insertion.

Export real Prometheus cache counters and size/capacity gauges with policy and role labels. Calculate throughput from elapsed time, reset baselines when cache instances change, and preserve monotonic exported counters. Document that throughput counts cache lookups and remove claims for unimplemented latency histograms.

PR #7 (cache capacity/safety) has already been merged into the branch for #14. This PR retains cache-capacity-safety as its base. Its diff contains only comparison and observability changes.

Validation: 62 Rust tests (including prerequisites), with five comparison/metrics regressions; cargo fmt; strict Clippy across all targets. Tests cover both directions of cache divergence, bench-mode isolation, elapsed-time throughput, and rendered Prometheus output. The combined runtime smoke check verified the live metrics endpoint.

Replaces #13 after removing the codex/ branch prefix. Commit contents are preserved; the original PR retains prior discussion and review history.