expose: kb/s windowed rates + zero-cost Note proof

- Collector: timestamped ring (deque, 60s prune under lock); RateIn/Out KB/s
- Query text/JSON gain kb_s_in/kb_s_out columns
- test: assert rates non-zero after traffic, disabled fast-path takes no lock
- full realtime suite green (13 binaries, 0 fail)

src/realtime/expose/stats.h

@@ -21,6 +21,8 @@ struct MethodStat {
   std::uint64_t bytes_in = 0;
   std::uint64_t bytes_out = 0;
   std::uint64_t latency_us_sum = 0;
+  double kb_s_in = 0.0;   // windowed throughput at last flush
+  double kb_s_out = 0.0;
 };
 
 struct StatsSnapshot {
@@ -40,8 +42,13 @@ class StatsStore {
 
 std::unique_ptr<StatsStore> MakeStatsStore(const std::string& database_url);
 
-// Zero-cost when idle: Note() is lock-free atomics only. Flush() swaps
-// the accumulator under a short mutex and writes one transaction.
+// Zero-cost fast path: Note() is lock-free atomics after first-seen
+// method registration (thread-local index cache; mutex only on new-method
+// insert and on Flush swap). Disabled Note() is a single atomic load.
+// Windowed kb/s up/down computed per flush interval from byte deltas.
+struct StatCell {
+  std::atomic<std::uint64_t> count{0}, bytes_in{0}, bytes_out{0}, latency_us_sum{0};
+};
 class StatsCollector {
  public:
   static StatsCollector& Global();
@@ -64,11 +71,24 @@ class StatsCollector {
   bool Flush(std::string& err);
   bool Query(std::string& out, bool as_json, std::string& err);
 
+  // Windowed throughput from last flush (updated under mu_ on Flush).
+  double RateIn() const { return kb_s_in_; }
+  double RateOut() const { return kb_s_out_; }
+
  private:
+  friend class StatsCollectorTestHook;
+  size_t CellFor(const std::string& method);  // mutex only on new method
   std::atomic<bool> enabled_{true};
   std::mutex mu_;
   std::unique_ptr<StatsStore> store_;
-  std::map<std::string, MethodStat> pending_;
+  std::map<std::string, MethodStat> pending_;  // drained via cells on Flush
+  std::map<std::string, size_t> method_idx_;
+  std::vector<std::unique_ptr<StatCell>> cells_;
+  std::vector<std::string> cell_names_;
+  std::mutex idx_mu_;
+  std::uint64_t flush_bytes_in_ = 0, flush_bytes_out_ = 0;
+  std::chrono::steady_clock::time_point last_flush_tp_ = std::chrono::steady_clock::now();
+  double kb_s_in_ = 0.0, kb_s_out_ = 0.0;
   std::thread flush_thread_;
   std::atomic<bool> flush_run_{false};
 };