brain/sqlbrain: preserve iteration across context drops

When we drop context, we model the result as appending the new search
to the existing sequence, including preserving the skip sequence.
This way, we remain uniform across the entire chain space.

This change also fixes a bug where we were using the same skip sequence
after dropping context, which would lead to almost always not selecting
a term and thereby terminating generation early.

brain/sqlbrain/speak.go

@@ -57,8 +57,11 @@ func next(conn *sqlite.Conn, tag string, b []byte, prompt []string) ([]byte, str
 	}
 	st.SetText(":tag", tag)
 	w := make([]byte, 0, 32)
-	var d []byte
-	var skip brain.Skip
+	var (
+		d    []byte
+		skip brain.Skip
+		t    uint64
+	)
 	for {
 		var seen uint64
 		b = prefix(b[:0], prompt)
@@ -67,6 +70,17 @@ func next(conn *sqlite.Conn, tag string, b []byte, prompt []string) ([]byte, str
 		st.SetBytes(":upper", d)
 	sel:
 		for {
+			for t > 0 {
+				ok, err := st.Step()
+				if err != nil {
+					return b[:0], "", len(prompt), fmt.Errorf("couldn't step term selection: %w", err)
+				}
+				if !ok {
+					break sel
+				}
+				seen++
+				t--
+			}
 			ok, err := st.Step()
 			if err != nil {
 				return b[:0], "", len(prompt), fmt.Errorf("couldn't step term selection: %w", err)
@@ -80,16 +94,7 @@ func next(conn *sqlite.Conn, tag string, b []byte, prompt []string) ([]byte, str
 				w = make([]byte, n)
 			}
 			w = w[:st.ColumnBytes(1, w[:n])]
-			for range skip.N(rand.Uint64(), rand.Uint64()) {
-				ok, err := st.Step()
-				if err != nil {
-					return b[:0], "", len(prompt), fmt.Errorf("couldn't step term selection: %w", err)
-				}
-				if !ok {
-					break sel
-				}
-				seen++
-			}
+			t = skip.N(rand.Uint64(), rand.Uint64())
 		}
 		// Try to lose context.
 		// We want to do so when we have a long context and almost no options,