Keep every legal compound surname reachable, and tighten the README wording

README.md

@@ -17,7 +17,9 @@ lacked the locale coverage and format control we needed.
   id), never for what character classes and choices already do.
 - **Composable** — templates nest without limit: weighted choices, character
   classes and sub-templates combine to model any format.
-- **Reproducible** — seed a faker and it emits the same sequence every time.
+- **Reproducible** — seed a faker and it emits the same sequence every time, for a
+  given version of the data: changing how a value is composed shifts the stream for
+  that value and for everything drawn after it in the same faker.
   Every built-in draws only from that seed — no wall-clock, no `crypto/rand` —
   so determinism holds end to end.
 - **Zero dependencies** — standard library only.
@@ -337,11 +339,10 @@ rejected at `New`.
 may be a `{..path}` reference too (`{name|..en_US.person}`).
 
 Inside a `format`, `0 1 A a` are **always** character classes — so a fixed digit
-or letter must be escaped (`#1`, `#A`) or it becomes random. Only those four
+or letter must be escaped (`#1`, `#A`) or it becomes random. A format of
-characters are classes: digits `2`–`9` and every other letter are literal. A
+`100 Main St` renders e.g. `506 Mdin St` — the `1`, `0`, `0` and `a` were random,
-format of `100 Main St` renders e.g. `506 Mdin St` — the `1`, `0`, `0` and `a`
+the `M`, `in` and `St` were not. A half-fixed string is the trap: `555-0000`
-were random, the `M`, `in` and `St` were not. A half-fixed string is the trap:
+keeps `555-` and randomises the last four digits. For a value with no
-`555-0100` keeps `555-` and randomises the last four digits. For a value with no
 tokens at all, use a bare string node (`"100 Main St"`), emitted verbatim.
 
 **Putting it together** (`person.json`):

data/sv_SE/person.json

@@ -17,9 +17,15 @@
       },
       {
         "format": "{first}{last}",
-        "weight": 0.4,
+        "weight": 0.267,
-        "first": ["Hell", "Norr", "Wall"],
+        "first": ["Hell", "Wall"],
-        "last": ["berg", "blad", "crona", "dahl", "ed", "fors", "gren", "holm", "in", "kvist", "man", "mark", "qvist", "stedt", "sten", "strand", "ström", "vall"]
+        "last": ["berg", "blad", "crona", "dahl", "ed", "fors", "gren", "holm", "in", "kvist", "man", "mark", "qvist", "roth", "stedt", "sten", "strand", "ström", "vall"]
+      },
+      {
+        "format": "{first}{last}",
+        "weight": 0.133,
+        "first": ["Norr"],
+        "last": ["berg", "blad", "crona", "dahl", "ed", "fors", "gren", "holm", "in", "kvist", "löf", "lund", "man", "mark", "qvist", "stedt", "sten", "strand", "ström", "vall"]
       },
       ["Berg", "Blom", "Eismar", "Falk", "Holm", "Lind", "Norberg", "Strand", "Ström", "von Flemming", "Åberg", "Öberg"]
     ],

data_test.go

@@ -127,10 +127,9 @@ func TestShippedMiscReferenceData(t *testing.T) {
 	}
 }
 
-// TestSwedishNamesHaveNoTripleLetter pins an orthographic rule the shape regexes
+// TestSwedishPersonNamesHaveNoTripleLetter pins an orthographic rule the shape
-// miss: Swedish never triples a consonant, so no prefix/suffix pair in the name
+// regexes miss: Swedish never triples a consonant.
-// data may compose one (Jöns+sson, Wall+lund, Norr+roth).
+func TestSwedishPersonNamesHaveNoTripleLetter(t *testing.T) {
-func TestSwedishNamesHaveNoTripleLetter(t *testing.T) {
 	f := newFakes(t, "data/sv_SE", WithSeed(11))
 	for _, path := range []string{"person", "person.last"} {
 		for i := 0; i < 20000; i++ {

edge_test.go

@@ -231,7 +231,8 @@ func TestShippedStreetComposition(t *testing.T) {
 }
 
 func TestShippedLastNameComposition(t *testing.T) {
-	// last is a choice of patronymic {first}sson templates and literal surnames.
+	// last is a choice of patronymic {first}sson templates, compound
+	// {first}{last} templates and literal surnames.
 	f := newFakes(t, "data/sv_SE", WithSeed(6))
 	for i := 0; i < 300; i++ {
 		if s := fake(t, f, "person.last"); !swedishName.MatchString(s) {