chore: observability audit fixes and documentation updates - #41

summary

• fixes connected_clients in INFO to report active connections instead of total accepted (was a monotonically increasing counter, now tracks live connections)
• adds connections_active atomic counter to ServerContext, incremented on accept and decremented on handler exit
• moves metrics_enabled into ServerContext to reduce parameter threading through handle() and process()
• gates Instant::now() behind a metrics_enabled || slowlog.is_enabled() check so there's zero timing overhead when both are disabled
• replaces .expect("slowlog lock poisoned") on all 4 mutex acquisitions in SlowLog with graceful poison recovery (clears entries and continues rather than crashing the server)
• updates README, Dockerfile, crate READMEs, and CLAUDE.md to reflect observability features

tested

• cargo clippy --workspace -- -D warnings — clean
• cargo test --workspace — all 579 tests pass
• reviewed all 4 mutex sites in slowlog.rs for consistent poison handling
• verified INFO CLIENTS section now reads connections_active instead of connections_accepted

design considerations

• poison recovery in SlowLog clears entries on the write path (maybe_record) but preserves them on read paths (get, len). the slow log is best-effort observability — losing entries after a panic is acceptable, crashing the server is not.
• is_enabled() on SlowLog reads config.enabled which is set at construction time and never changes, so no synchronization needed.
• timing gate uses Option<Instant> instead of a separate code path to keep the logic simple and avoid duplication.