Docs: fix stale "Five weight-sets" heading; GPU-qualify the latency claim

The body and table already describe the two shared weight-sets; this fixes the
leftover section heading and notes the ~174s/300s figure is GPU timing (the rules'
evaluation guidance assumes GPU batching). Model, weights, and manifest unchanged.

README.md

@@ -47,7 +47,7 @@ The single-step function is **piecewise linear** (`2t + bit*b`, then subtract 0,
 `2p`), which is why it generalises across primes where the full bilinear map does not:
 held-out-prime validation accuracy tracks training accuracy throughout (no memorisation gap).
 
-## Five weight-sets, routed by prime size
+## Two weight-sets, routed by prime size
 
 The recurrence is exact only if the state is wide enough to hold the residue, so each cell is
 trained per bit-width — but because the dilated convolution is weight-shared across bit-positions
@@ -269,8 +269,9 @@ P(acc<0.90) 0.002% / worst-prime 0.933 (primary key held). Public benchmark: tie
 Per-tier at total=1100: tiers 1–10 all **1.00**
 (overall_accuracy is the mean over tiers 1-10). Tier 0 (pure multiplication, primes near each
 width's maximum — a separate regime, not in overall_accuracy) is **0.70** on this fixed public
-seed. Inference for all 1100 problems is ~174s, within the 300s budget (the 2048-step tier-10 scan
-is the bulk); artifact 0.04 GB.
+seed. Inference for all 1100 problems is ~174s **on GPU** (the 2048-step tier-10 scan is the bulk),
+within the 300s budget; the rules' evaluation guidance assumes GPU batching via
+`predict_digits_batch` (`rules/evaluation.md:229`). artifact 0.04 GB.
 
 ## Status under the rules