fix: decode untagged embedded structs and maps inline

Assisted-by: GLM 5.3 Flash

docs/API.md

@@ -100,16 +100,23 @@ For a struct destination, a TOML key matches a field as follows:
 1. The `toml:"name"` tag, using the part before any comma. The literal `-`
    excludes the field.
 2. Without a tag, the lower-cased field name.
-3. The key itself is lower-cased before lookup, so the match is
+3. An anonymous (embedded) field without a tag is inlined: the decoder walks
+   into the embedded struct and matches its own fields against the same keys,
+   mirroring how the encoder flattens it. A nil embedded pointer struct is
+   allocated on demand. An untagged embedded map receives the keys no field
+   claims.
+4. The key itself is lower-cased before lookup, so the match is
    case-insensitive on both sides: `DATABASEURL` matches a field named
    `DatabaseUrl`.
 
 The match is exact after lower-casing. No separator is inserted, so a TOML key
 `database_url` does not match a field named `DatabaseUrl`; tag such a field
-(`toml:"database_url"`) or use the lower-cased name as the key. When two fields
-resolve to the same name, the one declared later wins.
+(`toml:"database_url"`) or use the lower-cased name as the key. When two
+fields resolve to the same name, the shallower one wins; at equal depth, the
+one declared later wins.
 
-Unknown keys are ignored by default; see [Strict decoding](#strict-decoding).
+Unknown keys are ignored by default, landing in an untagged embedded map when
+the struct has one; [Strict decoding](#strict-decoding) rejects them instead.
 
 ### Numeric conversion
 
@@ -257,10 +264,9 @@ type Config struct {
 Options combine after the name: `toml:"name,omitempty,omitzero"` is valid, and
 an unknown option is ignored.
 
-Note the asymmetry: the encoder inlines untagged embedded structs, while the
-decoder expects them under their lower-cased type name. A struct with an
-untagged embedded struct therefore does not round-trip through `Unmarshal` into
-the same type.
+Untagged embedded fields round-trip: the decoder inlines embedded structs and
+routes unclaimed keys into an embedded map exactly where the encoder flattened
+them.
 
 ### Group-by-kind layout