Generation now matches offline tests: full frame-aligned context (len%6==0, cap 1020), not seq[-48:]

daisychain.py

@@ -111,7 +111,13 @@ class DaisyChain:
         homopolymers). Sampling (greedy=False) trades that for variety; repetition_penalty
         then discourages the repeat loops these small specialists fall into."""
         m = self.models[domain]
-        ids = [self.bos] + (self.tok.encode(clean(prompt), add_special_tokens=False) if prompt else [])
+        # frame-align + cap: our 6-mer tokens tile cleanly only when the context length is a
+        # multiple of 6 (else the model generates out of phase). Trim the leading remainder and
+        # cap to the context window — this is what makes generation match the offline tests.
+        p = clean(prompt) if prompt else ""
+        p = p[-1020:]                 # 1020 = 170*6, within the 1024-token window
+        p = p[len(p) % 6:]            # drop leading remainder so the 6-mers align to the end
+        ids = [self.bos] + (self.tok.encode(p, add_special_tokens=False) if p else [])
         t = torch.tensor([ids], device=self.dev)
         bases, emitted = [], []
         while sum(len(b) for b in bases) < length: