diff --git a/AGENTS.md b/AGENTS.md
index 088cfce5..77284502 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -28,6 +28,14 @@ Avoid dropping an error without logging it. Use `tracing` for logging.
 let _ = ...
 ```
 
+**ALWAYS rate-limit logging that can fire from a loop or repeated error path.** Any `warn!`/`error!` reachable from a background task loop, retry loop, or per-request hot path MUST be wrapped in `rate_limited!`:
+```rust
+rate_limited!(Duration::from_secs(60), {
+    tracing::warn!("...: {e}");
+});
+```
+Unguarded logging in loops causes log spam that degrades observability and can itself become a performance problem. One-time paths (startup, shutdown, per-thread init) are exempt.
+
 ## Running tests
 
 - Always run `cargo nextest run` to run tests
diff --git a/dial9-tokio-telemetry/src/telemetry/recorder/mod.rs b/dial9-tokio-telemetry/src/telemetry/recorder/mod.rs
index 4da7a824..2ec08683 100644
--- a/dial9-tokio-telemetry/src/telemetry/recorder/mod.rs
+++ b/dial9-tokio-telemetry/src/telemetry/recorder/mod.rs
@@ -1704,7 +1704,9 @@ fn run_flush_loop(
             && !exit
             && let Err(e) = event_writer.drained()
         {
-            tracing::warn!("failed to complete post-drain action: {e}");
+            rate_limited!(Duration::from_secs(60), {
+                tracing::warn!("failed to complete post-drain action: {e}");
+            });
         }
 
         // Create the metrics guard up front; mutate on the exit path,
diff --git a/dial9-tokio-telemetry/src/telemetry/recorder/shared_state.rs b/dial9-tokio-telemetry/src/telemetry/recorder/shared_state.rs
index ff1cb56e..fc0ced59 100644
--- a/dial9-tokio-telemetry/src/telemetry/recorder/shared_state.rs
+++ b/dial9-tokio-telemetry/src/telemetry/recorder/shared_state.rs
@@ -472,6 +472,7 @@ mod shuttle_tests {
     use dial9_trace_format::TraceEvent;
     use shuttle::rand::Rng;
     use std::collections::HashMap;
+    use std::io;
 
     // ── Event definition ────────────────────────────────────────────────
 
@@ -673,4 +674,173 @@ mod shuttle_tests {
     fn pct_real_pipeline() {
         shuttle::check_pct(test_telemetry_core_pipeline, 10000, 3);
     }
+
+    // ── Always-erroring writer ──────────────────────────────────────────
+    //
+    // The companion to `InvariantCheckingWriter`: every fallible method
+    // returns an `io::Error`. This exercises the error paths in the flush
+    // loop and lets us assert that rate limiting bounds log output even
+    // when every operation fails.
+
+    use std::sync::Arc as StdArc;
+    use std::sync::atomic::{AtomicU64 as StdAtomicU64, Ordering as StdOrdering};
+
+    struct AlwaysErroringWriter;
+
+    impl AlwaysErroringWriter {
+        fn err() -> io::Error {
+            io::Error::from(io::ErrorKind::PermissionDenied)
+        }
+    }
+
+    impl TraceWriter for AlwaysErroringWriter {
+        fn write_encoded_batch(&mut self, _batch: &Batch) -> io::Result<()> {
+            Err(Self::err())
+        }
+
+        fn flush(&mut self) -> io::Result<()> {
+            Err(Self::err())
+        }
+
+        fn finalize(&mut self) -> io::Result<()> {
+            Err(Self::err())
+        }
+
+        fn write_current_segment_metadata(&mut self) -> io::Result<()> {
+            Err(Self::err())
+        }
+
+        fn should_drain(&self) -> bool {
+            // Always want to drain so the post-drain error path is exercised
+            // every flush tick.
+            true
+        }
+
+        fn drained(&mut self) -> io::Result<bool> {
+            Err(Self::err())
+        }
+    }
+
+    // ── Counting subscriber ─────────────────────────────────────────────
+    //
+    // A minimal `tracing::Subscriber` that increments a shared counter
+    // on every WARN or ERROR event. We use `tracing::subscriber::with_default`
+    // to scope it to a single test invocation. We deliberately avoid
+    // depending on `tracing-subscriber` since that crate is gated behind
+    // the `tracing-layer` feature and isn't enabled under `_shuttle`.
+
+    struct CountingSubscriber {
+        warn_or_error_count: StdArc<StdAtomicU64>,
+    }
+
+    impl tracing::Subscriber for CountingSubscriber {
+        fn enabled(&self, metadata: &tracing::Metadata<'_>) -> bool {
+            matches!(
+                *metadata.level(),
+                tracing::Level::WARN | tracing::Level::ERROR
+            )
+        }
+
+        fn new_span(&self, _span: &tracing::span::Attributes<'_>) -> tracing::span::Id {
+            // We never actually use spans; return a fixed non-zero id.
+            tracing::span::Id::from_u64(1)
+        }
+
+        fn record(&self, _span: &tracing::span::Id, _values: &tracing::span::Record<'_>) {}
+
+        fn record_follows_from(&self, _span: &tracing::span::Id, _follows: &tracing::span::Id) {}
+
+        fn event(&self, event: &tracing::Event<'_>) {
+            let level = *event.metadata().level();
+            if level == tracing::Level::WARN || level == tracing::Level::ERROR {
+                self.warn_or_error_count.fetch_add(1, StdOrdering::Relaxed);
+            }
+        }
+
+        fn enter(&self, _span: &tracing::span::Id) {}
+
+        fn exit(&self, _span: &tracing::span::Id) {}
+    }
+
+    // ── Erroring-pipeline test body ─────────────────────────────────────
+
+    fn test_telemetry_core_erroring_pipeline() {
+        let _ts_guard =
+            metrique_timesource::set_time_source(metrique_timesource::TimeSource::custom(
+                metrique_timesource::fakes::StaticTimeSource::at_time(std::time::UNIX_EPOCH),
+            ));
+
+        let warn_count = StdArc::new(StdAtomicU64::new(0));
+        let subscriber = CountingSubscriber {
+            warn_or_error_count: warn_count.clone(),
+        };
+
+        tracing::subscriber::with_default(subscriber, || {
+            let num_threads = 3;
+            let next_id = Arc::new(AtomicU64::new(0));
+
+            let guard = TelemetryCore::builder()
+                .writer(AlwaysErroringWriter)
+                .build()
+                .unwrap();
+            guard.enable();
+            let handle = guard.handle();
+
+            let writers: Vec<_> = (0..num_threads)
+                .map(|thread_id| {
+                    let h = handle.clone();
+                    let next_id = next_id.clone();
+                    let thread_id = thread_id as u64;
+                    crate::primitives::thread::spawn(move || {
+                        let mut rng = shuttle::rand::thread_rng();
+                        let count = rng.gen_range(3u64..=10);
+                        let mut ts = rng.gen_range(1000u64..2000);
+                        for seq in 0..count {
+                            let id = next_id.fetch_add(1, Ordering::Relaxed);
+                            let timestamp_ns = next_timestamp(&mut ts);
+                            let ev = ValidationEvent {
+                                timestamp_ns,
+                                thread_id,
+                                seq,
+                                id,
+                            };
+                            // Errors are expected; whether the event is
+                            // recorded or dropped is not asserted here.
+                            h.record_encodable_event(&ev);
+                        }
+                    })
+                })
+                .collect();
+
+            for w in writers {
+                w.join().unwrap();
+            }
+            // Dropping the guard triggers the shutdown/finalize path,
+            // which should also be rate-limited if it logs on error.
+            drop(guard);
+        });
+
+        // Bound the warn+error count by the number of distinct rate-limited
+        // call sites. There are roughly 6 sites that might fire on every
+        // shuttle run; allow some slack but cap firmly below "one per loop
+        // iteration". If a `rate_limited!` wrapper is removed from the flush
+        // loop, this number explodes.
+        let total = warn_count.load(StdOrdering::Relaxed);
+        assert!(
+            total <= 10,
+            "rate limiting failed under persistent writer errors: \
+             observed {total} WARN/ERROR events, expected <= 10. \
+             A `rate_limited!` wrapper has likely been removed from a tight loop."
+        );
+    }
+
+    #[test]
+    fn determinism_check_erroring() {
+        shuttle::check_uncontrolled_nondeterminism(test_telemetry_core_erroring_pipeline, 10000);
+    }
+
+    #[test]
+    fn pct_erroring_pipeline() {
+        shuttle::check_pct(test_telemetry_core_erroring_pipeline, 10000, 3);
+    }
 }