Cut per-token barriers in the naive recurrence from 12 to 3: - Removed the redundant pre-L2-norm barrier: each thread reads only its own ql[col]/kl[col] (written by itself) for the reduce_over_group input; reduce_over_group handles its own synchronization. - Replaced the 7-step rmsnorm sumsq tree reduction (7 barriers) with a single reduce_over_group + 1 SLM broadcast via pr (1 barrier). Reused pr for the sumsq so the lss local_accessor is gone. - Removed the redundant end-of-token barrier: the next iteration's first SLM op writes each thread's own ql/kl slot, and the first cross-lane read is guarded by the L2-norm broadcast barrier. Per-token barriers: 12 -> 3. SSM phase: 1.2 -> 1.1 ms/layer (-8%). Prefill (bench/session_4mb.txt): 8.95s -> 8.82s = 599 tok/s (was 591). Bit-exact correctness preserved (deltanet_naive_test 8/8 PASS, abs ~3e-6). All tests pass; output unchanged (' Paris. The').