diff --git a/docs/FEATURES.md b/docs/FEATURES.md
index 556734d..f309ac6 100644
--- a/docs/FEATURES.md
+++ b/docs/FEATURES.md
@@ -137,6 +137,35 @@ covers internals). When you ship a feature, add it here.
 
 ---
 
+## Known bugs
+
+Defects found by code review, not yet fixed.
+
+1. **Adaptive playout delay never shrinks back in real conversation**
+   (`src/core/jitter.rs`) — `target_delay` grows +1 per disruption up to
+   `MAX_DELAY_FRAMES` (12 frames = 240 ms) but only shrinks after
+   `CLEAN_RUN_TO_SHRINK` = 250 consecutive cleanly-played frames, i.e. **5 s of
+   unbroken audio**. `clean_run` is reset in five places; two of them fire on
+   every natural pause in speech: the benign-underrun branch (`jitter.rs:201`,
+   talker went quiet) and the subsequent re-prime (`jitter.rs:181`). Because the
+   sender skips transmitting entirely while the noise gate is closed
+   (`src/core/mod.rs:2041`), a pause between sentences *always* underruns the
+   receiver and zeroes the clean run — twice.
+
+   Net effect: the controller is a one-way ratchet. A single burst of jitter
+   early in a call pins up to 240 ms of extra playout latency for the rest of
+   the session, because no conversational speaker talks for 5 continuous
+   seconds without the gate closing. The AIMD "decrease" half is effectively
+   unreachable under the workload the app is built for.
+
+   Likely fix: let `clean_run` survive a benign idle→re-prime transition rather
+   than resetting it. Silence is not evidence the link is bad, so it should not
+   count against the clean run. Distinguish "talker stopped" (benign) from
+   "playout broke" (real) at `jitter.rs:195-202`.
+
+   Found 2026-07-31 by code review. Not yet reproduced in a live call — pairs
+   with field-test debt item 5 below.
+
 ## Known field-test debt (the 🧪 rows above, collected)
 
 Re-run on a real desktop ↔ dopedart call before calling these done: