Fix stale docstring: identity-scan prediction was pre-tower; scan decorative at LM inference (internal + Watson independent replication) — see paper §3.2

model_v2.py

@@ -14,9 +14,14 @@ Operator-native attention (per block = per head):
 
 Because rotors preserve η, score(t,s) = ⟨s_t, (R_q⁻¹R_k)·s_s⟩_η — attention is
 the alignment of the current state with a LEARNED-RELATIVELY-ROTATED past state.
-Everything is operator-derived → operator-only bottleneck holds → the R_state→
-identity bypass test still collapses to unigram (all states=s0 ⇒ attention over
-identical values ⇒ constant readout).
+Everything is operator-derived → operator-only bottleneck holds. STALE-PREDICTION
+NOTE (2026-09-10, kept for the record): "the R_state→identity bypass collapses to
+unigram" was written for the PRE-TOWER readout and was true of it; with the grade
+tower (shipped config) the readout has a direct route to B_t..B_{t-2} and identity-
+scan costs only ~1.2x (a LayerNorm distribution artifact — time-shuffling S costs
+~1.0x; measured internally and independently replicated by N. Watson 2026-09-10).
+At LM inference the transported state is causally decorative; in sole-channel/tape
+configs it is load-bearing. See paper §3.2.
 
 Scan + attention in fp32 (bf16 proven-unsafe on the recurrence).
 """