perf(encode): emit in place, pool the buffer and cache interface flags

Grouped emission walks the entries slice in three passes instead of copying them into per-kind slices, error paths render their key path only when an error names it, inline tables are measured once because a measuring encoder no longer nests another measurement, clean string runs are written in one write, and date-times render through the single-buffer path. The output buffer comes from a sync.Pool and returns to it within a 1 MiB retention cap, the output handed to the caller as a copy; repeated marshals keep the live heap flat, verified over 500 runs. The encoder resolves Marshaler and the text interfaces through the same per-type flag cache and hint the decoder uses, and docs/ARCHITECTURE.md now lists the caches and the pool as the library's shared state.

Representative document: 177 to 141 allocations, 12 230 to 6170 bytes per operation; long document: 97 175 to 63 660 allocations, 6.97 to 2.79 MB per operation, about 7.8 to about 3.4 ms.

.gitea/workflows/test.yml

@@ -105,6 +105,8 @@ jobs:
         run: go build -o bin/interpres-decode ./cmd/interpres-decode
 
       - name: Compliance suite
-        # interpres implements TOML 1.0 and 1.1; the mode is pinned so an upstream
+        # interpres implements TOML 1.1, and the suite runs both directions: the decoder
-        # default change cannot silently move the corpus.
+        # on the valid and invalid corpora, the encoder on the tagged JSON of the valid
-        run: bin/toml-test test -decoder=bin/interpres-decode -toml=1.1
+        # one. The mode is pinned so an upstream default change cannot silently move the
+        # corpus.
+        run: bin/toml-test test -decoder=bin/interpres-decode -encoder='bin/interpres-decode -encode' -toml=1.1

CHANGELOG.md

@@ -9,7 +9,107 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
--
+- `encoding.TextMarshaler` and `encoding.TextUnmarshaler` are honoured by
+  default, with no option to switch them off. A type that implements them is
+  encoded as a TOML string and decoded from one: `net.IP` becomes
+  `"192.0.2.1"`, and a user type with `MarshalText` or `UnmarshalText` follows.
+  `MarshalTOML` and `UnmarshalTOML` still win over the text methods, and the
+  four date-time types keep their bare timestamp form instead of becoming a
+  quoted string. A struct type that implements the interface now encodes as a
+  string where it was a table before, which is the breaking part of the change.
+- `time.Duration` is encoded in its canonical Go form as a TOML string,
+  `1h30m0s`, because TOML has no duration type; the decoder reads that string
+  back and still accepts a bare integer as the nanosecond count.
+- `interpres-decode -encode`, the adapter's other direction: it reads the
+  toml-test tagged JSON from stdin and writes the TOML document it describes.
+  The compliance suite now runs the encoder as well as the decoder, 214
+  encoder cases against the tagged JSON of the valid corpus.
+- `Encoder.InlineTables(threshold)`: a sub-table whose single-line rendering is
+  at most `threshold` bytes is written as an inline table instead of a header
+  section, which shortens a document of small tables. An array of tables keeps
+  its header form, because its inline form would re-parse as a value array.
+- `Document` and `ParseMap`: `Parse` now returns a `*Document`, which holds the
+  values together with the key order, whether a table was written as an inline
+  table or under a header, and the comments, with `Keys`, `Entries`, `Get`,
+  `Comments` and `SetComments` to read and write them. `ParseMap` returns the
+  plain `map[string]any` tree, the shape `Parse` used to give. `Marshal` does
+  not accept a `Document`; it writes values, so `doc.Map()` is the way through.
+- `OffsetDateTime`, the Go type of the offset date-time kind, so that all four
+  TOML date-time kinds have one of their own. `Parse` and `Unmarshal` hand it
+  back where they produced a bare `time.Time` before, and `Marshal` accepts it.
+  Unmarshalling into a struct field of type `time.Time` keeps working, because
+  the plain type takes an offset date-time as it always did; code that asserts
+  the tree's type, and `UnmarshalTOML` implementations that expect a
+  `time.Time`, need the new type.
+- `Decoder.MaxDepth(depth)` and `Decoder.MaxInputSize(size)` bound the parse a
+  `Decode` performs, and every parse carries a nesting limit in any case
+  (10000 levels, which no hand-written document approaches): a document that
+  nests arrays or inline tables deeper used to run the stack out and is now
+  rejected with a `SyntaxError` naming the limit.
+
+### Changed
+
+- The output takes the TOML 1.1 form. A date-time writes its seconds only when
+  the value carries them and drops the trailing zeros of a fractional second,
+  so `07:32:00` is written `07:32` and half a second as `00.5`. Both are the
+  same value, and a document written without seconds now comes back without
+  them. `LocalDateTime.String()`, `LocalTime.String()` and the offset date-time
+  rendering follow the same rule.
+- An inline table that would pass the hundredth column is written across lines
+  with a trailing comma and one tab of indentation per nesting level, the shape
+  TOML 1.1 allows an inline table to take.
+- `MarshalTOML` reaches every array element and every field, whatever the Go
+  kind, and its result is normalised like any other value: an element rendering
+  itself as a table keeps the `[[header]]` form, one rendering itself as a
+  scalar turns the array into a value array, and the method runs once per
+  element. It is found on the addressable pointer as well, so a
+  pointer-receiver `MarshalTOML` is called for a field or an element, exactly
+  as `MarshalText` is.
+- TOML 1.1 is the acceptance contract, and TOML 1.0 is not. The compliance
+  suite runs the 1.1 corpus alone, and the promise that every 1.0 document
+  parses exactly as before is withdrawn. Nothing that parses today stops
+  parsing: the 1.0 valid corpus still passes in full. The documents whose
+  verdict changes are the ones 1.1 relaxed, such as the `\xHH` escape
+  sequences 1.0 rejected.
+- The module path carries the /v2 suffix the Go toolchain requires of
+  every major version 2 module: imports change to
+  `sourcedock.dev/petrbalvin/interpres/v2`.
+- Input that is not valid UTF-8 is now rejected where the parser's scan
+  meets the invalid byte, with a `SyntaxError` naming that line, instead of
+  a whole-input check that always reported line 1. Invalid input is still
+  rejected; the reported location is now the byte's own.
+
+**Performance**
+
+- Parsing is faster than in 1.1.0 while carrying the new document layer:
+  the suite's representative document decodes at about 79 MB/s with 104
+  allocations per call, and the long array-of-tables document at about
+  106 MB/s against 56 MB/s in 1.1.0, with allocations on that document
+  halved from 67 664 to 31 765. Date-time tokens are validated by a byte
+  scan instead of regular expressions, repeated keys share one string
+  across array-of-tables elements, and per-statement buffers are reused.
+- Typed decoding is 12 percent faster than in 1.1.0 on the representative
+  document (9792 ns against 11 147 ns) with 24 percent fewer allocations
+  (167 against 220); interface lookups resolve through a cached per-type
+  flag set instead of boxing every value into an interface to ask.
+- `Marshal` runs at the 1.1.0 speed while emitting the new TOML 1.1 output
+  form, at half the bytes per operation (6170 against 11 348 on the
+  representative document), and writes through a pooled output buffer with
+  a 1 MiB retention cap; repeated marshals keep the live heap flat.
+- Two benchmarks measure the shapes that drove the work:
+  `BenchmarkStrictDecodeLong` and `BenchmarkMarshalLong` run the 2000-entry
+  document at about 3.8 ms and 3.4 ms per call, at 63 772 and 63 660
+  allocations.
+
+### Fixed
+
+- Decoding into a defined type whose underlying kind is string or bool, such
+  as `type Name string`, panicked instead of storing the value, because a
+  value of the predeclared type is not assignable to a defined type and the
+  decoder assigned it without a conversion.
+- A top-level value the encoder could not normalise reported its path with
+  a leading dot, `interpres: .port: ...`; the message now reads
+  `interpres: port: ...`, the shape `EncodeError.Path` already used.
 
 ## [1.1.0] - 2026-09-18
 

CONTRIBUTING.md

@@ -45,8 +45,8 @@ just test
    formatting pass are three commits, never one.
 4. Record every user-visible change in `CHANGELOG.md` under `## [development]`.
 5. Add or update tests. Coverage stays at 80 percent or more; it is a hard
-   gate. Parser and decoder changes must also keep the toml-test suite at zero
+   gate. Parser, decoder and encoder changes must also keep both directions of
-   failures, checked with `just toml-test`.
+   the toml-test suite at zero failures, checked with `just toml-test`.
 6. Update the documentation when the public API, the configuration or the
    behaviour changes; the documents move in the same commit as the behaviour
    they describe.

README.md

@@ -1,14 +1,14 @@
 # interpres
 
-A TOML 1.0 and 1.1 parser and encoder for Go, written with the standard
+A TOML 1.1 parser and encoder for Go, written with the standard library
-library alone. `interpres` (Latin for *interpreter*) gives zero-dependency
+alone. `interpres` (Latin for *interpreter*) gives zero-dependency
 programs an `encoding/json`-style API for reading and writing TOML, and passes
 the entire official [toml-test](https://github.com/toml-lang/toml-test) suite:
-214 valid and 467 invalid cases, zero failures.
+214 valid, 467 invalid and 214 encoder cases, zero failures.
 
 ## Features
 
-- **Full TOML 1.0 and 1.1**: bare, quoted and dotted keys; tables and arrays of
+- **Full TOML 1.1**: bare, quoted and dotted keys; tables and arrays of
   tables; basic and literal strings including multiline, with the 1.1 `\e` and
   `\xHH` escapes; integers in the four radixes with `_` separators; floats with
   exponents, `inf` and `nan`; booleans; the four date-time kinds, seconds
@@ -18,18 +18,24 @@ the entire official [toml-test](https://github.com/toml-lang/toml-test) suite:
 - **Strict decoding**: `NewDecoder().DisallowUnknownFields()` rejects keys that
   match no destination field, at every struct depth.
 - **Custom types**: `Marshaler` and `Unmarshaler` let a type control its own
-  TOML representation in both directions.
+  TOML representation in both directions, and `encoding.TextMarshaler` and
+  `TextUnmarshaler` are honoured by default, so `net.IP`, `time.Duration` and
+  user types with text methods need no configuration.
 - **Cancellation**: every entry point has a `*Context` sibling that honours a
   `context.Context`.
+- **Ordered documents**: `Parse` gives a `*Document` that keeps the key order,
+tells an inline table from a header one, and carries the comments; `ParseMap`
+gives the plain `map[string]any` tree.
 - **Configurable emission**: `Encoder` options for declaration-order output,
-  omitting empty arrays, and literal multiline strings.
+  omitting empty arrays, literal multiline strings, and inlining small
+  sub-tables.
 
 ## Install
 
 As a library:
 
 ```sh
-go get sourcedock.dev/petrbalvin/interpres
+go get sourcedock.dev/petrbalvin/interpres/v2
 ```
 
 Requires Go 1.27.1 or newer. The module imports only the standard library.
@@ -64,8 +70,9 @@ err := interpres.Unmarshal(data, &cfg)
 ```
 
 Fields match by the `toml:"name"` tag, or by the lower-cased field name when no
-tag is present; `toml:"-"` skips a field. `Parse` returns the untyped
+tag is present; `toml:"-"` skips a field. `Parse` returns a `*Document` that
-`map[string]any` tree instead, and `UnmarshalContext` accepts a context.
+also carries the key order and the comments, `ParseMap` returns the plain
+`map[string]any` tree, and `UnmarshalContext` accepts a context.
 
 ### Encode from a struct
 
@@ -161,7 +168,7 @@ See [docs/DEVELOPMENT.md](docs/DEVELOPMENT.md) for the full workflow, and
 
 - [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md): components and data flow
 - [docs/API.md](docs/API.md): the API reference, decoding and encoding rules
-- [docs/CLI.md](docs/CLI.md): the interpres-decode toml-test adapter and validator
+- [docs/CLI.md](docs/CLI.md): the interpres-decode adapter and validator
 
 ## Licence
 

SECURITY.md

@@ -7,7 +7,7 @@ releases do not receive them.
 
 | Version | Supported |
 |---|---|
-| 1.0.0 | yes |
+| 1.1.0 | yes |
 | older releases | no |
 
 ## Reporting a vulnerability

bench_test.go

@@ -87,14 +87,14 @@ func BenchmarkParse(b *testing.B) {
 	b.ReportAllocs()
 	b.SetBytes(int64(len(benchDoc)))
 	for b.Loop() {
-		if _, err := Parse(benchDoc); err != nil {
+		if _, err := ParseMap(benchDoc); err != nil {
 			b.Fatal(err)
 		}
 	}
 }
 
 func BenchmarkMarshal(b *testing.B) {
-	tree, err := Parse(benchDoc)
+	tree, err := ParseMap(benchDoc)
 	if err != nil {
 		b.Fatal(err)
 	}
@@ -122,7 +122,48 @@ func BenchmarkParseLong(b *testing.B) {
 	b.ReportAllocs()
 	b.SetBytes(int64(len(longDoc)))
 	for b.Loop() {
-		if _, err := Parse(longDoc); err != nil {
+		if _, err := ParseMap(longDoc); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+// benchLongEntry mirrors one [[entry]] element of longDoc for the typed
+// decode of the long document.
+type benchLongEntry struct {
+	Name   string    `toml:"name"`
+	Weight int       `toml:"weight"`
+	When   time.Time `toml:"when"`
+	Ratio  float64   `toml:"ratio"`
+	Tags   []string  `toml:"tags"`
+}
+
+type benchLongDoc struct {
+	Title string           `toml:"title"`
+	Entry []benchLongEntry `toml:"entry"`
+}
+
+func BenchmarkStrictDecodeLong(b *testing.B) {
+	dec := NewDecoder().DisallowUnknownFields()
+	b.ReportAllocs()
+	b.SetBytes(int64(len(longDoc)))
+	for b.Loop() {
+		var doc benchLongDoc
+		if err := dec.Decode(longDoc, &doc); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func BenchmarkMarshalLong(b *testing.B) {
+	tree, err := ParseMap(longDoc)
+	if err != nil {
+		b.Fatal(err)
+	}
+	b.ReportAllocs()
+	b.SetBytes(int64(len(longDoc)))
+	for b.Loop() {
+		if _, err := Marshal(tree); err != nil {
 			b.Fatal(err)
 		}
 	}

cmd/interpres-decode/main.go

@@ -4,14 +4,16 @@
 // Command interpres-decode is the toml-test harness adapter and a TOML
 // validator. Without flags it reads a TOML document from standard input and
 // writes the toml-test "tagged JSON" representation to standard output. With
-// -validate it checks the named documents, or standard input when none are
+// -encode it is the reverse: it reads tagged JSON and writes the TOML document
-// named, and exits non-zero on the first invalid one:
+// it describes. With -validate it checks the named documents, or standard
+// input when none are named, and exits non-zero on the first invalid one:
 //
 //	interpres-decode -validate config.toml
+//	interpres-decode -encode < case.json
 //
-// Run the official suite against the adapter with:
+// Run the official suite in both directions against the adapter with:
 //
-//	toml-test ./interpres-decode
+//	toml-test test -decoder=./interpres-decode -encoder='./interpres-decode -encode'
 package main
 
 import (
@@ -25,7 +27,7 @@ import (
 	"strconv"
 	"time"
 
-	"sourcedock.dev/petrbalvin/interpres"
+	"sourcedock.dev/petrbalvin/interpres/v2"
 )
 
 func main() {
@@ -39,12 +41,17 @@ func Run(args []string, stdin io.Reader, stdout, stderr io.Writer) int {
 	fs := flag.NewFlagSet("interpres-decode", flag.ContinueOnError)
 	fs.SetOutput(stderr)
 	validate := fs.Bool("validate", false, "validate the documents instead of emitting tagged JSON")
+	encode := fs.Bool("encode", false, "read tagged JSON from stdin and write TOML instead")
 	if err := fs.Parse(args); err != nil {
 		if errors.Is(err, flag.ErrHelp) {
 			return 0
 		}
 		return 2
 	}
+	if *validate && *encode {
+		fmt.Fprintln(stderr, "interpres-decode: -validate and -encode cannot be combined")
+		return 2
+	}
 	if *validate {
 		return validatePaths(fs.Args(), stdin, stderr)
 	}
@@ -52,12 +59,15 @@ func Run(args []string, stdin io.Reader, stdout, stderr io.Writer) int {
 		fmt.Fprintln(stderr, "interpres-decode: the adapter mode takes no arguments; name files with -validate")
 		return 2
 	}
+	if *encode {
+		return encodeJSON(stdin, stdout, stderr)
+	}
 	data, err := io.ReadAll(stdin)
 	if err != nil {
 		fmt.Fprintln(stderr, "read stdin:", err)
 		return 2
 	}
-	tree, err := interpres.Parse(data)
+	tree, err := interpres.ParseMap(data)
 	if err != nil {
 		fmt.Fprintln(stderr, err)
 		return 1
@@ -98,7 +108,7 @@ func validatePaths(paths []string, stdin io.Reader, stderr io.Writer) int {
 			fmt.Fprintf(stderr, "interpres-decode: %s: %v\n", name, err)
 			return 2
 		}
-		if _, err := interpres.Parse(data); err != nil {
+		if _, err := interpres.ParseMap(data); err != nil {
 			fmt.Fprintf(stderr, "%s: %v\n", name, err)
 			valid = false
 		}
@@ -109,6 +119,190 @@ func validatePaths(paths []string, stdin io.Reader, stderr io.Writer) int {
 	return 0
 }
 
+// encodeJSON reads a toml-test tagged JSON description from standard input and
+// writes the TOML document it describes to standard output.
+func encodeJSON(stdin io.Reader, stdout, stderr io.Writer) int {
+	data, err := io.ReadAll(stdin)
+	if err != nil {
+		fmt.Fprintln(stderr, "read stdin:", err)
+		return 2
+	}
+	var desc any
+	if err := json.Unmarshal(data, &desc); err != nil {
+		fmt.Fprintln(stderr, "decode JSON:", err)
+		return 2
+	}
+	tree, err := untag(desc)
+	if err != nil {
+		fmt.Fprintln(stderr, err)
+		return 2
+	}
+	doc, ok := tree.(map[string]any)
+	if !ok {
+		fmt.Fprintln(stderr, "interpres-decode: the description must be a JSON object at the top level")
+		return 2
+	}
+	out, err := interpres.Marshal(doc)
+	if err != nil {
+		fmt.Fprintln(stderr, err)
+		return 2
+	}
+	if _, err := stdout.Write(out); err != nil {
+		fmt.Fprintln(stderr, "write stdout:", err)
+		return 2
+	}
+	return 0
+}
+
+// untag converts a toml-test JSON description into the value tree Marshal
+// expects: a JSON object becomes a map[string]any, a JSON array becomes a
+// []any, and an object carrying exactly the keys "type" and "value" becomes
+// the Go value for that TOML type.
+func untag(v any) (any, error) {
+	switch x := v.(type) {
+	case map[string]any:
+		if typ, val, ok := taggedValue(x); ok {
+			return decodeTagged(typ, val)
+		}
+		out := make(map[string]any, len(x))
+		for k, e := range x {
+			u, err := untag(e)
+			if err != nil {
+				return nil, fmt.Errorf("%s: %w", k, err)
+			}
+			out[k] = u
+		}
+		return out, nil
+	case []any:
+		out := make([]any, len(x))
+		for i, e := range x {
+			u, err := untag(e)
+			if err != nil {
+				return nil, fmt.Errorf("[%d]: %w", i, err)
+			}
+			out[i] = u
+		}
+		return asTables(out), nil
+	default:
+		return nil, fmt.Errorf("unsupported JSON value %T", v)
+	}
+}
+
+// asTables returns the elements as a []map[string]any when there is at least
+// one and every element is a table, the shape the encoder renders as an array
+// of tables. The tagged JSON cannot tell an array of tables from a value array
+// of inline tables, and both parse back to the same value, so the header form
+// is chosen because it is the one the encoder otherwise never exercises. An
+// empty array stays a []any, because TOML has no empty array of tables.
+func asTables(items []any) any {
+	if len(items) == 0 {
+		return items
+	}
+	tbls := make([]map[string]any, len(items))
+	for i, e := range items {
+		tbl, ok := e.(map[string]any)
+		if !ok {
+			return items
+		}
+		tbls[i] = tbl
+	}
+	return tbls
+}
+
+// taggedValue reports whether m is a toml-test value object: a JSON object of
+// exactly the two string keys "type" and "value", carrying a type this adapter
+// knows. Any other object is a table.
+func taggedValue(m map[string]any) (typ, val string, ok bool) {
+	if len(m) != 2 {
+		return "", "", false
+	}
+	ts, ok := m["type"].(string)
+	if !ok || !knownType(ts) {
+		return "", "", false
+	}
+	vs, ok := m["value"].(string)
+	if !ok {
+		return "", "", false
+	}
+	return ts, vs, true
+}
+
+func knownType(typ string) bool {
+	switch typ {
+	case "string", "integer", "float", "bool",
+		"datetime", "datetime-local", "date-local", "time-local":
+		return true
+	}
+	return false
+}
+
+// decodeTagged returns the Go value for one tagged JSON value. Every type but
+// string is parsed by the library itself, so the adapter and the library agree
+// on what an integer, a float or a date-time is.
+func decodeTagged(typ, val string) (any, error) {
+	if typ == "string" {
+		return val, nil
+	}
+	v, err := parseAtom(val)
+	if err != nil {
+		return nil, fmt.Errorf("%s %q: %w", typ, val, err)
+	}
+	// A float with no fractional part and no exponent is described by a bare
+	// integer literal, so here the tag decides and not the literal.
+	if n, ok := v.(int64); ok && typ == "float" {
+		return float64(n), nil
+	}
+	if !typeMatches(typ, v) {
+		return nil, fmt.Errorf("%s %q parsed as %T", typ, val, v)
+	}
+	return v, nil
+}
+
+// parseAtom parses one bare TOML value, by handing `v = <val>` to the library's
+// parser and requiring the result to hold exactly that one statement, so a
+// value carrying a newline or a comment cannot smuggle a second one in.
+func parseAtom(val string) (any, error) {
+	tree, err := interpres.ParseMap([]byte("v = " + val + "\n"))
+	if err != nil {
+		return nil, err
+	}
+	if len(tree) != 1 {
+		return nil, errors.New("not a single bare value")
+	}
+	return tree["v"], nil
+}
+
+// typeMatches reports whether v is the Go value the tagged type names.
+func typeMatches(typ string, v any) bool {
+	switch typ {
+	case "integer":
+		_, ok := v.(int64)
+		return ok
+	case "float":
+		_, ok := v.(float64)
+		return ok
+	case "bool":
+		_, ok := v.(bool)
+		return ok
+	case "datetime":
+		switch v.(type) {
+		case time.Time, interpres.OffsetDateTime:
+			return true
+		}
+		return false
+	case "datetime-local":
+		_, ok := v.(interpres.LocalDateTime)
+		return ok
+	case "date-local":
+		_, ok := v.(interpres.LocalDate)
+		return ok
+	case "time-local":
+		_, ok := v.(interpres.LocalTime)
+		return ok
+	}
+	return false
+}
+
 // tag converts an interpres value into its toml-test tagged-JSON form. Tables
 // become JSON objects and arrays become JSON arrays; scalars are wrapped in a
 // {"type", "value"} object. An error is returned for value types the encoder
@@ -155,6 +349,8 @@ func tag(v any) (any, error) {
 		return tagged("float", formatFloat(x)), nil
 	case time.Time:
 		return tagged("datetime", x.Format(time.RFC3339Nano)), nil
+	case interpres.OffsetDateTime:
+		return tagged("datetime", x.Format(time.RFC3339Nano)), nil
 	case interpres.LocalDateTime:
 		return tagged("datetime-local", x.Format("2006-01-02T15:04:05.999999999")), nil
 	case interpres.LocalDate:

cmd/interpres-decode/main_test.go

@@ -8,11 +8,12 @@ import (
 	"encoding/json"
 	"errors"
 	"os"
+	"reflect"
 	"strings"
 	"testing"
 	"time"
 
-	"sourcedock.dev/petrbalvin/interpres"
+	"sourcedock.dev/petrbalvin/interpres/v2"
 )
 
 func TestRunParsesValidTOML(t *testing.T) {
@@ -283,3 +284,165 @@ func TestUnknownFlagReturnsTwo(t *testing.T) {
 		t.Fatalf("Run returned %d, want 2; stderr = %q", code, stderr.String())
 	}
 }
+
+// --- encoder mode ----------------------------------------------------------
+
+func TestRunEncoderScalars(t *testing.T) {
+	in := `{
+    "s":   {"type": "string", "value": "quote \" and backslash \\"},
+    "nl":  {"type": "string", "value": "line1\nline2"},
+    "i":   {"type": "integer", "value": "-9223372036854775808"},
+    "g":   {"type": "float", "value": "1.5"},
+    "f":   {"type": "float", "value": "inf"},
+    "b":   {"type": "bool", "value": "false"},
+    "dt":  {"type": "datetime", "value": "1979-05-27T07:32:00-07:00"},
+    "ldt": {"type": "datetime-local", "value": "1979-05-27T07:32:00"},
+    "ld":  {"type": "date-local", "value": "1979-05-27"},
+    "lt":  {"type": "time-local", "value": "07:32:00.999"}
+}
+`
+	var stdout, stderr bytes.Buffer
+	code := Run([]string{"-encode"}, strings.NewReader(in), &stdout, &stderr)
+	if code != 0 {
+		t.Fatalf("Run returned %d, stderr = %q", code, stderr.String())
+	}
+	want := "b = false\n" +
+		"dt = 1979-05-27T07:32-07:00\n" +
+		"f = inf\n" +
+		"g = 1.5\n" +
+		"i = -9223372036854775808\n" +
+		"ld = 1979-05-27\n" +
+		"ldt = 1979-05-27T07:32\n" +
+		"lt = 07:32:00.999\n" +
+		"nl = \"line1\\nline2\"\n" +
+		"s = \"quote \\\" and backslash \\\\\"\n"
+	if stdout.String() != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", stdout.String(), want)
+	}
+}
+
+func TestRunEncoderNested(t *testing.T) {
+	in := `{
+    "tbl": {"x": {"type": "bool", "value": "true"},
+            "sub": {"y": {"type": "integer", "value": "1"}}},
+    "items": [{"n": {"type": "string", "value": "a"}},
+              {"n": {"type": "string", "value": "b"}}],
+    "list": [{"type": "integer", "value": "1"}, {"type": "string", "value": "two"}],
+    "emptyTbl": {},
+    "emptyArr": []
+}
+`
+	var stdout, stderr bytes.Buffer
+	code := Run([]string{"-encode"}, strings.NewReader(in), &stdout, &stderr)
+	if code != 0 {
+		t.Fatalf("Run returned %d, stderr = %q", code, stderr.String())
+	}
+	want := "emptyArr = []\n" +
+		"list = [1, \"two\"]\n" +
+		"\n[emptyTbl]\n" +
+		"\n[tbl]\nx = true\n" +
+		"\n[tbl.sub]\ny = 1\n" +
+		"\n[[items]]\nn = \"a\"\n" +
+		"\n[[items]]\nn = \"b\"\n"
+	if stdout.String() != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", stdout.String(), want)
+	}
+}
+
+func TestRunEncoderFloatTagDecides(t *testing.T) {
+	// A float with no fraction is described by a bare integer literal, so the
+	// tag decides the type; the output must stay a float.
+	var stdout, stderr bytes.Buffer
+	in := `{"whole": {"type": "float", "value": "1"}, "exp": {"type": "float", "value": "5e+22"}}`
+	code := Run([]string{"-encode"}, strings.NewReader(in), &stdout, &stderr)
+	if code != 0 {
+		t.Fatalf("Run returned %d, stderr = %q", code, stderr.String())
+	}
+	if want := "exp = 5e+22\nwhole = 1.0\n"; stdout.String() != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", stdout.String(), want)
+	}
+}
+
+func TestRunEncoderRejectsBadInput(t *testing.T) {
+	cases := []struct {
+		name string
+		in   string
+		want string
+	}{
+		{"not-json", "not json", "decode JSON"},
+		{"top-level-array", `[{"type": "integer", "value": "1"}]`, "must be a JSON object"},
+		{"untagged-scalar", `{"x": 1}`, "unsupported JSON value"},
+		{"literal-mismatch", `{"x": {"type": "integer", "value": "1.5"}}`, "parsed as float64"},
+		{"offset-for-local", `{"x": {"type": "datetime-local", "value": "1979-05-27T07:32:00Z"}}`, "parsed as interpres.OffsetDateTime"},
+		{"bad-literal", `{"x": {"type": "date-local", "value": "nope"}}`, "date-local"},
+		{"smuggled-statement", `{"x": {"type": "integer", "value": "1\nx = 2"}}`, "not a single bare value"},
+	}
+	for _, c := range cases {
+		var stdout, stderr bytes.Buffer
+		code := Run([]string{"-encode"}, strings.NewReader(c.in), &stdout, &stderr)
+		if code != 2 {
+			t.Errorf("%s: Run returned %d, want 2; stderr = %q", c.name, code, stderr.String())
+			continue
+		}
+		if !strings.Contains(stderr.String(), c.want) {
+			t.Errorf("%s: stderr = %q, want it to mention %q", c.name, stderr.String(), c.want)
+		}
+		if stdout.Len() != 0 {
+			t.Errorf("%s: stdout should be empty, got %q", c.name, stdout.String())
+		}
+	}
+}
+
+func TestRunEncoderFlagConflicts(t *testing.T) {
+	var stdout, stderr bytes.Buffer
+	if code := Run([]string{"-encode", "-validate"}, strings.NewReader(""), &stdout, &stderr); code != 2 {
+		t.Errorf("Run returned %d, want 2 for the two modes together", code)
+	}
+	if !strings.Contains(stderr.String(), "cannot be combined") {
+		t.Errorf("stderr = %q, want it to explain the conflict", stderr.String())
+	}
+
+	stdout.Reset()
+	stderr.Reset()
+	if code := Run([]string{"-encode", "file.json"}, strings.NewReader(""), &stdout, &stderr); code != 2 {
+		t.Errorf("Run returned %d, want 2 for an argument", code)
+	}
+}
+
+func TestEncodeAfterDecodeRoundTrip(t *testing.T) {
+	doc := `title = "x"
+flt = 1.5
+whole = 7.0
+big = 9223372036854775807
+when = 1979-05-27T07:32:00-07:00
+day = 1979-05-27
+clock = 07:32:00.999
+list = [1, "two"]
+multi = "a\nb"
+
+[tbl]
+x = true
+
+[[items]]
+n = "a"
+`
+	var tagged, stderr bytes.Buffer
+	if code := Run(nil, strings.NewReader(doc), &tagged, &stderr); code != 0 {
+		t.Fatalf("decode returned %d, stderr = %q", code, stderr.String())
+	}
+	var out bytes.Buffer
+	if code := Run([]string{"-encode"}, bytes.NewReader(tagged.Bytes()), &out, &stderr); code != 0 {
+		t.Fatalf("encode returned %d, stderr = %q", code, stderr.String())
+	}
+	want, err := interpres.ParseMap([]byte(doc))
+	if err != nil {
+		t.Fatalf("parse of the original: %v", err)
+	}
+	got, err := interpres.ParseMap(out.Bytes())
+	if err != nil {
+		t.Fatalf("parse of the encoder output (%q): %v", out.String(), err)
+	}
+	if !reflect.DeepEqual(want, got) {
+		t.Errorf("round trip changed the document:\noriginal: %#v\nencoded:  %#v\noutput: %q", want, got, out.String())
+	}
+}

datetime.go

@@ -5,15 +5,19 @@ package interpres
 
 import (
 	"fmt"
-	"regexp"
-	"strconv"
 	"strings"
 	"time"
 )
 
-// TOML distinguishes four date-time kinds. interpres decodes an offset
+// TOML distinguishes four date-time kinds, and each has its own Go type:
-// date-time to a plain time.Time (it carries a zone), and uses the wrapper
+// OffsetDateTime for the offset kind, and the local wrappers below for the
-// types below for the local variants so callers can tell them apart.
+// three that carry no offset. A plain time.Time is accepted wherever an
+// offset date-time is, on both the encoding and the decoding side, so a
+// timestamp field does not have to name the wrapper.
+
+// OffsetDateTime is a TOML offset date-time, e.g. 1979-05-27T07:32:00-07:00.
+// The embedded time.Time is the instant, with the offset the document wrote.
+type OffsetDateTime struct{ time.Time }
 
 // LocalDateTime is a TOML local date-time with no offset, e.g.
 // 1979-05-27T07:32:00. The embedded time.Time is in UTC.
@@ -27,73 +31,210 @@ type LocalDate struct{ time.Time }
 // The embedded time.Time uses the zero date.
 type LocalTime struct{ time.Time }
 
+// String returns the TOML-canonical rendering of the offset date-time, e.g.
+// "1979-05-27T07:32Z" or "1979-05-27T07:32:00-07:00". The seconds appear only
+// when the value carries them, a fractional second drops its trailing zeros,
+// and an offset of zero is written "Z".
+func (odt OffsetDateTime) String() string { return offsetString(odt.Time) }
+
 // String returns the TOML-canonical rendering of the local date-time, e.g.
-// "1979-05-27T07:32:00" or "...:00.000000123" when the time has a fractional
+// "1979-05-27T07:32" or "1979-05-27T07:32:00.5" when the time carries a
-// second. The fractional component is zero-padded to nanosecond precision.
+// fractional second. TOML 1.1 makes the seconds optional, so they appear only
+// when they are non-zero, and a fraction drops its trailing zeros.
 func (ldt LocalDateTime) String() string {
-	base := ldt.Format("2006-01-02T15:04:05")
+	buf := ldt.Time.AppendFormat(make([]byte, 0, 32), "2006-01-02T")
-	if ns := ldt.Nanosecond(); ns > 0 {
+	return string(appendClock(buf, ldt.Time))
-		return base + "." + fmt.Sprintf("%09d", ns)
-	}
-	return base
 }
 
 // String returns the TOML-canonical rendering of the local date, e.g.
 // "1979-05-27".
 func (ld LocalDate) String() string { return ld.Format("2006-01-02") }
 
-// String returns the TOML-canonical rendering of the local time, e.g.
+// String returns the TOML-canonical rendering of the local time, e.g. "07:32"
-// "07:32:00" or "...:00.000000123" when the time has a fractional second.
+// or "07:32:00.5" when the time carries a fractional second.
-// The fractional component is zero-padded to nanosecond precision.
+func (lt LocalTime) String() string { return clockString(lt.Time) }
-func (lt LocalTime) String() string {
+
-	base := lt.Format("15:04:05")
+// appendClock appends the clock part of a TOML time to buf: HH:MM, seconds
-	if ns := lt.Nanosecond(); ns > 0 {
+// only when the value carries them, and a fraction with its trailing zeros
-		return base + "." + fmt.Sprintf("%09d", ns)
+// dropped, so half a second is ".5" and not ".500000000". Both are the same
+// value either way; the shorter form is the one TOML 1.1 allows. The whole
+// rendering is built in one buffer, because the encoder writes a date-time
+// per entry of a large document.
+func appendClock(buf []byte, t time.Time) []byte {
+	buf = t.AppendFormat(buf, "15:04")
+	if t.Second() != 0 || t.Nanosecond() != 0 {
+		buf = t.AppendFormat(buf, ":05")
 	}
-	return base
+	if ns := t.Nanosecond(); ns > 0 {
+		buf = append(buf, '.')
+		buf = append(buf, strings.TrimRight(fmt.Sprintf("%09d", ns), "0")...)
+	}
+	return buf
 }
 
-var (
+// clockString renders a time of day the way TOML writes it.
-	offsetDateTimeLayouts = []string{
+func clockString(t time.Time) string {
-		"2006-01-02T15:04:05.999999999Z07:00",
+	return string(appendClock(make([]byte, 0, 16), t))
-		"2006-01-02T15:04:05Z07:00",
-		"2006-01-02 15:04:05.999999999Z07:00",
-		"2006-01-02 15:04:05Z07:00",
-		// TOML 1.1 makes the seconds optional.
-		"2006-01-02T15:04Z07:00",
-		"2006-01-02 15:04Z07:00",
 }
-	localDateTimeLayouts = []string{
+
-		"2006-01-02T15:04:05.999999999",
+// offsetString renders an offset date-time, the fourth TOML kind, in the same
-		"2006-01-02T15:04:05",
+// shape: no zero seconds, no trailing zeros in the fraction, and the offset
-		"2006-01-02 15:04:05.999999999",
+// written as "Z" when it is zero.
-		"2006-01-02 15:04:05",
+func offsetString(t time.Time) string {
-		"2006-01-02T15:04",
+	buf := t.AppendFormat(make([]byte, 0, 32), "2006-01-02T")
-		"2006-01-02 15:04",
+	buf = appendClock(buf, t)
-	}
+	buf = t.AppendFormat(buf, "Z07:00")
-	localTimeLayouts = []string{
+	return string(buf)
-		"15:04:05.999999999",
-		"15:04:05",
-		"15:04",
 }
+
+// dateTimeKind names the date-time shape a bare token has, as the scanner
+// below classifies it.
+type dateTimeKind int
+
+const (
+	dateTimeNone dateTimeKind = iota
+	dateTimeOffset
+	dateTimeLocal
+	dateTimeDate
+	dateTimeClock
 )
 
-// dateTimeShape enforces the strict TOML grammar (two-digit components,
+// The layouts the time package parses each shape with. Parsing accepts a
-// seconds optional since 1.1, a fraction only after seconds) that time.Parse
+// fractional second even when the layout does not carry one, so each shape
-// would otherwise accept loosely (e.g. a single-digit hour).
+// needs a single layout, chosen by whether the token has seconds.
-var dateTimeShape = regexp.MustCompile(
+const (
-	`^\d{4}-\d{2}-\d{2}([Tt ]\d{2}:\d{2}(:\d{2}(\.\d+)?)?([Zz]|[+-]\d{2}:\d{2})?)?$` +
+	offsetDateTimeLayout = "2006-01-02T15:04:05Z07:00"
-		`|^\d{2}:\d{2}(:\d{2}(\.\d+)?)?$`,
+	offsetClockLayout    = "2006-01-02T15:04Z07:00"
+	localDateTimeLayout  = "2006-01-02T15:04:05"
+	localClockLayout     = "2006-01-02T15:04"
+	localTimeLayout      = "15:04:05"
+	localTimeClockLayout = "15:04"
+	localDateOnlyLayout  = "2006-01-02"
 )
 
-// offsetBounds extracts the numeric offset of a date-time. The ABNF bounds it
+// scanDateTimeShape validates a bare token against the strict TOML date-time
-// to 00:00 through 23:59, but time.Parse accepts values outside that range
+// grammar and reports which kind it is: two-digit components, seconds
-// and rolls them over (for example "+00:60" becomes "+01:00"), so the bounds
+// optional since TOML 1.1, a fraction only after seconds, an offset only
-// are enforced here.
+// after a time, and an offset bounded to 00:00 through 23:59. The grammar is
-var offsetBounds = regexp.MustCompile(`([+-])(\d{2}):(\d{2})$`)
+// a fixed byte shape, so the scan is a byte walk; the regular expressions
+// this replaced cost the parser measurably per token, and a shape that fails
+// the scan is simply not a date-time.
+func scanDateTimeShape(tok string) (kind dateTimeKind, seconds bool) {
+	// A local clock on its own: HH:MM[:SS[.fraction]].
+	if len(tok) >= 5 && tok[2] == ':' {
+		n, secs, ok := scanClock(tok, 0)
+		if !ok || n != len(tok) {
+			return dateTimeNone, false
+		}
+		return dateTimeClock, secs
+	}
+	// A date, optionally followed by a time and an offset.
+	if len(tok) < 10 || tok[4] != '-' || tok[7] != '-' {
+		return dateTimeNone, false
+	}
+	for _, i := range [8]int{0, 1, 2, 3, 5, 6, 8, 9} {
+		if !isDecDigit(tok[i]) {
+			return dateTimeNone, false
+		}
+	}
+	if len(tok) == 10 {
+		return dateTimeDate, false
+	}
+	if sep := tok[10]; sep != 'T' && sep != 't' && sep != ' ' {
+		return dateTimeNone, false
+	}
+	n, secs, ok := scanClock(tok, 11)
+	if !ok {
+		return dateTimeNone, false
+	}
+	if n == len(tok) {
+		return dateTimeLocal, secs
+	}
+	// The offset: Z/z, or a signed HH:MM bounded as the ABNF requires.
+	switch c := tok[n]; {
+	case c == 'Z' || c == 'z':
+		if n+1 != len(tok) {
+			return dateTimeNone, false
+		}
+	case c == '+' || c == '-':
+		if n+6 != len(tok) || tok[n+3] != ':' ||
+			!isDecDigit(tok[n+1]) || !isDecDigit(tok[n+2]) ||
+			!isDecDigit(tok[n+4]) || !isDecDigit(tok[n+5]) ||
+			tok[n+1] > '2' || (tok[n+1] == '2' && tok[n+2] > '3') ||
+			tok[n+4] > '5' {
+			return dateTimeNone, false
+		}
+	default:
+		return dateTimeNone, false
+	}
+	return dateTimeOffset, secs
+}
+
+// scanClock validates HH:MM[:SS[.fraction]] starting at i and returns the
+// position after the clock, whether seconds were present, and whether the
+// shape is valid.
+func scanClock(tok string, i int) (pos int, seconds bool, ok bool) {
+	if i+5 > len(tok) || tok[i+2] != ':' ||
+		!isDecDigit(tok[i]) || !isDecDigit(tok[i+1]) ||
+		!isDecDigit(tok[i+3]) || !isDecDigit(tok[i+4]) {
+		return 0, false, false
+	}
+	i += 5
+	if i == len(tok) || tok[i] != ':' {
+		return i, false, true
+	}
+	if i+3 > len(tok) || !isDecDigit(tok[i+1]) || !isDecDigit(tok[i+2]) {
+		return 0, false, false
+	}
+	i += 3
+	if i == len(tok) || tok[i] != '.' {
+		return i, true, true
+	}
+	i++
+	digits := i
+	for i < len(tok) && isDecDigit(tok[i]) {
+		i++
+	}
+	if i == digits {
+		return 0, false, false
+	}
+	return i, true, true
+}
+
+// normaliseDateTimeToken rewrites the date/time separator to 'T' and the
+// offset marker to 'Z', the characters the layouts above carry. A token that
+// already has them is returned as it is, without a copy.
+func normaliseDateTimeToken(tok string, kind dateTimeKind) string {
+	if kind == dateTimeDate || kind == dateTimeClock {
+		return tok
+	}
+	needs := false
+	for i := range len(tok) {
+		c := tok[i]
+		if c == 't' || c == 'z' || (c == ' ' && i == 10) {
+			needs = true
+			break
+		}
+	}
+	if !needs {
+		return tok
+	}
+	b := []byte(tok)
+	for i, c := range b {
+		switch {
+		case c == 't':
+			b[i] = 'T'
+		case c == 'z':
+			b[i] = 'Z'
+		case c == ' ' && i == 10:
+			b[i] = 'T'
+		}
+	}
+	return string(b)
+}
 
 // parseDateTime classifies and parses a bare token as a TOML date-time value.
-// It returns the decoded value (time.Time, LocalDateTime, LocalDate, or
+// It returns the decoded value (OffsetDateTime, LocalDateTime, LocalDate or
 // LocalTime) and whether the token was a date-time at all.
 func parseDateTime(tok string) (any, bool) {
 	if tok == "" || tok[0] < '0' || tok[0] > '9' {
@@ -102,36 +243,49 @@ func parseDateTime(tok string) (any, bool) {
 	if !strings.ContainsAny(tok, "-:") {
 		return nil, false
 	}
-	if !dateTimeShape.MatchString(tok) {
+	kind, seconds := scanDateTimeShape(tok)
+	if kind == dateTimeNone {
 		return nil, false
 	}
-	if m := offsetBounds.FindStringSubmatch(tok); m != nil {
+	norm := normaliseDateTimeToken(tok, kind)
-		hour, _ := strconv.Atoi(m[2])
+	switch kind {
-		minute, _ := strconv.Atoi(m[3])
+	case dateTimeOffset:
-		if hour > 23 || minute > 59 {
+		layout := offsetClockLayout
+		if seconds {
+			layout = offsetDateTimeLayout
+		}
+		t, err := time.Parse(layout, norm)
+		if err != nil {
 			return nil, false
 		}
+		return OffsetDateTime{t}, true
+	case dateTimeLocal:
+		layout := localClockLayout
+		if seconds {
+			layout = localDateTimeLayout
 		}
-	// The ABNF accepts lowercase "t"/"z"; time.Parse only matches uppercase.
+		t, err := time.Parse(layout, norm)
-	norm := strings.ToUpper(tok)
+		if err != nil {
-	for _, layout := range offsetDateTimeLayouts {
+			return nil, false
-		if t, err := time.Parse(layout, norm); err == nil {
-			return t, true
 		}
-	}
-	for _, layout := range localDateTimeLayouts {
-		if t, err := time.Parse(layout, norm); err == nil {
 		return LocalDateTime{t}, true
+	case dateTimeDate:
+		t, err := time.Parse(localDateOnlyLayout, norm)
+		if err != nil {
+			return nil, false
 		}
-	}
-	if t, err := time.Parse("2006-01-02", norm); err == nil {
 		return LocalDate{t}, true
+	case dateTimeClock:
+		layout := localTimeClockLayout
+		if seconds {
+			layout = localTimeLayout
+		}
+		t, err := time.Parse(layout, norm)
+		if err != nil {
+			return nil, false
 		}
-	for _, layout := range localTimeLayouts {
-		if t, err := time.Parse(layout, norm); err == nil {
 		return LocalTime{t}, true
 	}
-	}
 	return nil, false
 }
 

decode.go

@@ -4,11 +4,13 @@
 package interpres
 
 import (
+	"encoding"
 	"fmt"
 	"reflect"
 	"slices"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 )
 
@@ -21,6 +23,97 @@ func newDecoder() *decoder { return &decoder{} }
 
 var timeType = reflect.TypeFor[time.Time]()
 
+var (
+	unmarshalerType     = reflect.TypeFor[Unmarshaler]()
+	textUnmarshalerType = reflect.TypeFor[encoding.TextUnmarshaler]()
+)
+
+// The per-type flags record which interface lookups a decode into that type
+// can succeed at, so the hot path consults the cache instead of boxing every
+// value into an interface to ask. The bits name the receiver the method is
+// found on: the value itself, or its address.
+const (
+	flagUnmarshaler uint8 = 1 << iota
+	flagAddrUnmarshaler
+	flagTextUnmarshaler
+	flagAddrTextUnmarshaler
+)
+
+// typeFlagCache holds one flag entry per destination type. A set is immutable
+// once published, the same trade-off structSchemaCache makes; the cache grows
+// with the number of distinct types decoded, never per document. The hint
+// below re-points at these published entries, so a hot lookup allocates
+// nothing.
+var typeFlagCache sync.Map // reflect.Type -> *flagHintEntry
+
+// flagHintEntry pairs a type with its cached flags for the monomorphic hint
+// below. Both caches share the entry shape.
+type flagHintEntry struct {
+	typ   reflect.Type
+	flags uint8
+}
+
+// typeFlagHint remembers the entry resolved last, because a decode walks one
+// type across consecutive fields and elements. A lost race loses only the
+// hint: every value it can hold came from the cache.
+var typeFlagHint atomic.Pointer[flagHintEntry]
+
+func typeFlags(t reflect.Type) uint8 {
+	if e := typeFlagHint.Load(); e != nil && e.typ == t {
+		return e.flags
+	}
+	if v, ok := typeFlagCache.Load(t); ok {
+		entry := v.(*flagHintEntry)
+		typeFlagHint.Store(entry)
+		return entry.flags
+	}
+	var f uint8
+	if t.Implements(unmarshalerType) {
+		f |= flagUnmarshaler
+	}
+	pt := reflect.PointerTo(t)
+	if pt.Implements(unmarshalerType) {
+		f |= flagAddrUnmarshaler
+	}
+	// The date-time types are excluded from the text path: they carry
+	// time.Time's UnmarshalText through an embedded field while their only
+	// accepted form is a bare timestamp.
+	if !isDateTimeType(t) {
+		if t.Implements(textUnmarshalerType) {
+			f |= flagTextUnmarshaler
+		}
+		if pt.Implements(textUnmarshalerType) {
+			f |= flagAddrTextUnmarshaler
+		}
+	}
+	actual, _ := typeFlagCache.LoadOrStore(t, &flagHintEntry{t, f})
+	published := actual.(*flagHintEntry)
+	typeFlagHint.Store(published)
+	return published.flags
+}
+
+// unmarshalerOf resolves the Unmarshaler for dst through the flag cache, so
+// an interface value is built only where the cache says the assertion can
+// succeed. An interface destination is asked dynamically, because the value
+// it will hold may implement the interface even when the interface type
+// itself does not.
+func unmarshalerOf(dst reflect.Value) (Unmarshaler, bool) {
+	if dst.Kind() == reflect.Interface {
+		u, ok := dst.Interface().(Unmarshaler)
+		return u, ok
+	}
+	f := typeFlags(dst.Type())
+	if f&flagUnmarshaler != 0 {
+		u, ok := dst.Interface().(Unmarshaler)
+		return u, ok
+	}
+	if f&flagAddrUnmarshaler != 0 && dst.CanAddr() {
+		u, ok := dst.Addr().Interface().(Unmarshaler)
+		return u, ok
+	}
+	return nil, false
+}
+
 func (d *decoder) decode(tree map[string]any, v any) error {
 	rv := reflect.ValueOf(v)
 	if rv.Kind() != reflect.Pointer || rv.IsNil() {
@@ -63,6 +156,19 @@ func (d *decoder) assign(data any, dst reflect.Value) error {
 		}
 	}
 
+	// A TOML string fills a destination that implements
+	// encoding.TextUnmarshaler, the rule encoding/json follows. Every other
+	// value kind keeps its own rule, so an integer still reaches a numeric
+	// destination.
+	if s, isString := data.(string); isString {
+		if tu, ok := textUnmarshalerOf(dst); ok {
+			if err := tu.UnmarshalText([]byte(s)); err != nil {
+				return fmt.Errorf("unmarshal text: %w", err)
+			}
+			return nil
+		}
+	}
+
 	switch v := data.(type) {
 	case map[string]any:
 		return d.assignTable(v, dst)
@@ -71,6 +177,9 @@ func (d *decoder) assign(data any, dst reflect.Value) error {
 	case []any:
 		return d.assignSlice(v, dst)
 	case string:
+		if dst.Type() == durationType {
+			return setDuration(dst, v)
+		}
 		return setBasic(dst, reflect.ValueOf(v), "string")
 	case bool:
 		return setBasic(dst, reflect.ValueOf(v), "bool")
@@ -78,12 +187,10 @@ func (d *decoder) assign(data any, dst reflect.Value) error {
 		return setInt(dst, v)
 	case float64:
 		return setFloat(dst, v)
+	case OffsetDateTime:
+		return setOffsetDateTime(v, dst)
 	case time.Time:
-		if dst.Type() != timeType {
+		return setDateTime(v, dst)
-			return fmt.Errorf("interpres: cannot assign datetime to %s", dst.Type())
-		}
-		dst.Set(reflect.ValueOf(v))
-		return nil
 	default:
 		rv := reflect.ValueOf(data)
 		if rv.IsValid() && dst.Type() == rv.Type() {
@@ -94,6 +201,28 @@ func (d *decoder) assign(data any, dst reflect.Value) error {
 	}
 }
 
+// textUnmarshalerOf is the same resolution for encoding.TextUnmarshaler,
+// with the date-time types excluded for the reason typeFlags records.
+func textUnmarshalerOf(dst reflect.Value) (encoding.TextUnmarshaler, bool) {
+	if !dst.CanInterface() || isDateTimeType(dst.Type()) {
+		return nil, false
+	}
+	if dst.Kind() == reflect.Interface {
+		tu, ok := dst.Interface().(encoding.TextUnmarshaler)
+		return tu, ok
+	}
+	f := typeFlags(dst.Type())
+	if f&flagTextUnmarshaler != 0 {
+		tu, ok := dst.Interface().(encoding.TextUnmarshaler)
+		return tu, ok
+	}
+	if f&flagAddrTextUnmarshaler != 0 && dst.CanAddr() {
+		tu, ok := dst.Addr().Interface().(encoding.TextUnmarshaler)
+		return tu, ok
+	}
+	return nil, false
+}
+
 func (d *decoder) assignTable(tbl map[string]any, dst reflect.Value) error {
 	switch dst.Kind() {
 	case reflect.Struct:
@@ -112,6 +241,9 @@ func (d *decoder) assignStruct(tbl map[string]any, dst reflect.Value) error {
 		// deterministically: the smallest one.
 		unknown := ""
 		for key := range tbl {
+			if _, ok := schema.byName[key]; ok {
+				continue
+			}
 			if _, ok := schema.byName[strings.ToLower(key)]; ok {
 				continue
 			}
@@ -124,7 +256,12 @@ func (d *decoder) assignStruct(tbl map[string]any, dst reflect.Value) error {
 		}
 	}
 	for key, val := range tbl {
-		field, ok := schema.byName[strings.ToLower(key)]
+		// A key that is already lowercase, which document keys usually are,
+		// hits the map directly; only a miss pays for the case fold.
+		field, ok := schema.byName[key]
+		if !ok {
+			field, ok = schema.byName[strings.ToLower(key)]
+		}
 		if !ok {
 			if schema.embedMaps != nil {
 				// Leftover keys land in an untagged embedded map, the inverse
@@ -198,11 +335,57 @@ func (d *decoder) assignTableSlice(items []map[string]any, dst reflect.Value) er
 
 // --- low-level setters -----------------------------------------------------
 
+// setOffsetDateTime stores an offset date-time: in a wrapper destination as it
+// is, and in a plain time.Time, which takes the instant with the offset the
+// document wrote, so a timestamp field does not have to name the wrapper.
+func setOffsetDateTime(v OffsetDateTime, dst reflect.Value) error {
+	switch dst.Type() {
+	case offsetDateTimeType:
+		dst.Set(reflect.ValueOf(v))
+	case timeType:
+		dst.Set(reflect.ValueOf(v.Time))
+	default:
+		return fmt.Errorf("interpres: cannot assign datetime to %s", dst.Type())
+	}
+	return nil
+}
+
+// setDateTime stores a time.Time that reached the tree directly, which is the
+// shape a tree built by hand carries. Dates the parser produced arrive as
+// OffsetDateTime instead.
+func setDateTime(v time.Time, dst reflect.Value) error {
+	switch dst.Type() {
+	case timeType:
+		dst.Set(reflect.ValueOf(v))
+	case offsetDateTimeType:
+		dst.Set(reflect.ValueOf(OffsetDateTime{v}))
+	default:
+		return fmt.Errorf("interpres: cannot assign datetime to %s", dst.Type())
+	}
+	return nil
+}
+
 func setBasic(dst, val reflect.Value, kind string) error {
 	if dst.Kind() != val.Kind() {
 		return fmt.Errorf("interpres: cannot assign %s to %s", kind, dst.Type())
 	}
-	dst.Set(val)
+	// Convert rather than assign: a value of the predeclared type is not
+	// assignable to a defined type of the same kind, so a plain Set panics on
+	// a destination such as `type Name string`.
+	dst.Set(val.Convert(dst.Type()))
+	return nil
+}
+
+// setDuration reads a duration literal into a time.Duration destination. TOML
+// has no duration type, so the encoder writes the canonical Go form and the
+// decoder reads that back; a bare integer stays the nanosecond count it has
+// always been, and reaches the destination through setInt.
+func setDuration(dst reflect.Value, s string) error {
+	d, err := time.ParseDuration(s)
+	if err != nil {
+		return fmt.Errorf("interpres: invalid duration %q", s)
+	}
+	dst.SetInt(int64(d))
 	return nil
 }
 

decode_test.go

@@ -8,9 +8,11 @@ import (
 	"errors"
 	"fmt"
 	"math"
+	"net"
 	"slices"
 	"strings"
 	"testing"
+	"time"
 )
 
 func TestSyntaxErrorMessage(t *testing.T) {
@@ -22,19 +24,38 @@ func TestSyntaxErrorMessage(t *testing.T) {
 }
 
 func TestParseRejectsInvalidUTF8(t *testing.T) {
-	_, err := Parse([]byte("v = \"\xff\"\n"))
+	// The scan validates UTF-8 where it meets the byte, so the reported line
+	// is the invalid byte's own, wherever in the document it sits.
+	cases := []struct {
+		name string
+		doc  string
+		line int
+	}{
+		{"in a basic string", "v = \"\xff\"\n", 1},
+		{"in a literal string", "v = '\xff'\n", 1},
+		{"in a multiline string", "v = \"\"\"\n\xff\"\"\"\n", 2},
+		{"in a comment", "v = 1\n# caf\xe9\xff\n", 2},
+		{"in a bare key", "va\xfflue = 1\n", 1},
+		{"as a statement", "\xff = 1\n", 1},
+		{"in a bare value", "v = \xff1\n", 1},
+		{"after a value", "v = 1 \xff\n", 1},
+		{"after the first line", "a = 1\nb = \"\xff\"\n", 2},
+	}
+	for _, c := range cases {
+		_, err := ParseMap([]byte(c.doc))
 		if err == nil {
-		t.Fatal("expected a UTF-8 validation error")
+			t.Fatalf("%s: expected a UTF-8 validation error", c.name)
 		}
 		se, ok := err.(*SyntaxError)
 		if !ok {
-		t.Fatalf("err is %T, want *SyntaxError", err)
+			t.Fatalf("%s: err is %T, want *SyntaxError", c.name, err)
 		}
 		if !strings.Contains(se.Msg, "UTF-8") {
-		t.Errorf("Msg = %q, want it to mention UTF-8", se.Msg)
+			t.Errorf("%s: Msg = %q, want it to mention UTF-8", c.name, se.Msg)
+		}
+		if se.Line != c.line {
+			t.Errorf("%s: Line = %d, want %d", c.name, se.Line, c.line)
 		}
-	if se.Line != 1 {
-		t.Errorf("Line = %d, want 1", se.Line)
 	}
 }
 
@@ -128,7 +149,7 @@ func TestUnmarshalIntoNilAny(t *testing.T) {
 func TestParseContextHonoursCancellation(t *testing.T) {
 	ctx, cancel := context.WithCancel(context.Background())
 	cancel()
-	if _, err := ParseContext(ctx, []byte("a = 1\n")); !errors.Is(err, context.Canceled) {
+	if _, err := ParseMapContext(ctx, []byte("a = 1\n")); !errors.Is(err, context.Canceled) {
 		t.Fatalf("ParseContext returned %v, want context.Canceled", err)
 	}
 }
@@ -158,7 +179,7 @@ func TestContextRoundTrip(t *testing.T) {
 	in := []byte(`title = "x"
 count = 3
 `)
-	if _, err := ParseContext(context.Background(), in); err != nil {
+	if _, err := ParseMapContext(context.Background(), in); err != nil {
 		t.Fatalf("ParseContext: %v", err)
 	}
 	var out struct {
@@ -725,3 +746,354 @@ func TestDecodeErrorOnMapDestination(t *testing.T) {
 		t.Fatalf("Path = %v", de.Path)
 	}
 }
+
+func TestUnmarshalIntoDefinedScalarTypes(t *testing.T) {
+	// A defined type whose underlying kind is string or bool takes the value.
+	// A bare reflect Set panics on such a type, because a string is not
+	// assignable to a defined string type without a conversion.
+	type Name string
+	type Flag bool
+	type Cfg struct {
+		N Name `toml:"n"`
+		F Flag `toml:"f"`
+	}
+	var cfg Cfg
+	if err := Unmarshal([]byte("n = \"x\"\nf = true\n"), &cfg); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if cfg.N != "x" {
+		t.Errorf("N = %q, want \"x\"", cfg.N)
+	}
+	if !cfg.F {
+		t.Error("F = false, want true")
+	}
+}
+
+// --- encoding.TextUnmarshaler and time.Duration ----------------------------
+
+// textReceiver implements encoding.TextUnmarshaler on the pointer receiver.
+type textReceiver struct{ Text string }
+
+func (t *textReceiver) UnmarshalText(text []byte) error {
+	t.Text = "got:" + string(text)
+	return nil
+}
+
+// upperText is a defined string type whose UnmarshalText transforms the
+// content, so a plain string assignment would leave the wrong value behind.
+type upperText string
+
+func (u *upperText) UnmarshalText(text []byte) error {
+	*u = upperText(strings.ToUpper(string(text)))
+	return nil
+}
+
+// failingTextUnmarshaler fails the decode from UnmarshalText.
+type failingTextUnmarshaler struct{}
+
+func (f *failingTextUnmarshaler) UnmarshalText(_ []byte) error { return errors.New("text boom") }
+
+// textAndTOMLReceiver implements both decode interfaces; the TOML method wins.
+type textAndTOMLReceiver struct{ From string }
+
+func (t *textAndTOMLReceiver) UnmarshalTOML(any) error { t.From = "toml"; return nil }
+
+func (t *textAndTOMLReceiver) UnmarshalText([]byte) error { t.From = "text"; return nil }
+
+func TestTextUnmarshalerByPointer(t *testing.T) {
+	type Cfg struct {
+		R textReceiver `toml:"r"`
+	}
+	var cfg Cfg
+	if err := Unmarshal([]byte(`r = "hello"`), &cfg); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if cfg.R.Text != "got:hello" {
+		t.Errorf("Text = %q, want \"got:hello\"", cfg.R.Text)
+	}
+}
+
+func TestTextUnmarshalerWinsOverKindAssignment(t *testing.T) {
+	type Cfg struct {
+		U upperText `toml:"u"`
+	}
+	var cfg Cfg
+	if err := Unmarshal([]byte(`u = "abc"`), &cfg); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if cfg.U != "ABC" {
+		t.Errorf("U = %q, want \"ABC\"", cfg.U)
+	}
+}
+
+func TestTextUnmarshalerForNetIP(t *testing.T) {
+	type Cfg struct {
+		V4  net.IP   `toml:"v4"`
+		V6  net.IP   `toml:"v6"`
+		IPs []net.IP `toml:"ips"`
+	}
+	in := "v4 = \"192.0.2.1\"\nv6 = \"2001:db8::68\"\nips = [\"198.51.100.7\", \"203.0.113.9\"]\n"
+	var cfg Cfg
+	if err := Unmarshal([]byte(in), &cfg); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if got := cfg.V4.String(); got != "192.0.2.1" {
+		t.Errorf("V4 = %q, want \"192.0.2.1\"", got)
+	}
+	if got := cfg.V6.String(); got != "2001:db8::68" {
+		t.Errorf("V6 = %q, want \"2001:db8::68\"", got)
+	}
+	if len(cfg.IPs) != 2 || cfg.IPs[0].String() != "198.51.100.7" || cfg.IPs[1].String() != "203.0.113.9" {
+		t.Errorf("IPs = %v, want two addresses", cfg.IPs)
+	}
+}
+
+func TestTextUnmarshalerSeesStringsOnly(t *testing.T) {
+	// An integer keeps its own rule: the text method is not consulted, and the
+	// value does not reach the receiver.
+	type Cfg struct {
+		R textReceiver `toml:"r"`
+	}
+	var cfg Cfg
+	err := Unmarshal([]byte("r = 1\n"), &cfg)
+	if err == nil {
+		t.Fatal("expected an integer to be rejected for a text receiver")
+	}
+	if cfg.R.Text != "" {
+		t.Errorf("Text = %q, want it untouched", cfg.R.Text)
+	}
+}
+
+func TestUnmarshalTOMLWinsOverTextUnmarshaler(t *testing.T) {
+	type Cfg struct {
+		B textAndTOMLReceiver `toml:"b"`
+	}
+	var cfg Cfg
+	if err := Unmarshal([]byte(`b = "x"`), &cfg); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if cfg.B.From != "toml" {
+		t.Errorf("From = %q, want \"toml\"", cfg.B.From)
+	}
+}
+
+func TestTextUnmarshalerErrorCarriesPath(t *testing.T) {
+	type Inner struct {
+		F failingTextUnmarshaler `toml:"f"`
+	}
+	type Cfg struct {
+		Inner Inner `toml:"inner"`
+	}
+	var cfg Cfg
+	err := Unmarshal([]byte("[inner]\nf = \"x\"\n"), &cfg)
+	if err == nil {
+		t.Fatal("expected an error from UnmarshalText")
+	}
+	if !strings.Contains(err.Error(), "unmarshal text: text boom") {
+		t.Errorf("err = %v, want the text error wrapped", err)
+	}
+	de, ok := errors.AsType[*DecodeError](err)
+	if !ok {
+		t.Fatalf("expected a *DecodeError, got %T: %v", err, err)
+	}
+	if !slices.Equal(de.Path, []string{"inner", "f"}) {
+		t.Fatalf("Path = %v, want [inner f]", de.Path)
+	}
+}
+
+func TestTextUnmarshalerReportsBadText(t *testing.T) {
+	var cfg struct {
+		IP net.IP `toml:"ip"`
+	}
+	err := Unmarshal([]byte(`ip = "not-an-ip"`), &cfg)
+	if err == nil {
+		t.Fatal("expected an error for a malformed address")
+	}
+	if !strings.Contains(err.Error(), "unmarshal text:") {
+		t.Errorf("err = %v, want it wrapped as a text error", err)
+	}
+}
+
+func TestUnmarshalDurations(t *testing.T) {
+	type Cfg struct {
+		FromText time.Duration `toml:"from_text"`
+		FromInt  time.Duration `toml:"from_int"`
+		Fraction time.Duration `toml:"fraction"`
+	}
+	in := "from_text = \"1h30m\"\nfrom_int = 5400000000000\nfraction = \"1.5s\"\n"
+	var cfg Cfg
+	if err := Unmarshal([]byte(in), &cfg); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if cfg.FromText != 90*time.Minute {
+		t.Errorf("FromText = %v, want %v", cfg.FromText, 90*time.Minute)
+	}
+	if cfg.FromInt != 90*time.Minute {
+		t.Errorf("FromInt = %v, want %v", cfg.FromInt, 90*time.Minute)
+	}
+	if cfg.Fraction != 1500*time.Millisecond {
+		t.Errorf("Fraction = %v, want %v", cfg.Fraction, 1500*time.Millisecond)
+	}
+}
+
+func TestUnmarshalDurationRejectsMalformedText(t *testing.T) {
+	var cfg struct {
+		D time.Duration `toml:"d"`
+	}
+	err := Unmarshal([]byte("d = \"90\"\n"), &cfg)
+	if err == nil {
+		t.Fatal("expected an error for a duration without a unit")
+	}
+	if !strings.Contains(err.Error(), "invalid duration") {
+		t.Errorf("err = %v, want an invalid-duration message", err)
+	}
+}
+
+func TestQuotedStringNeverBecomesDateTime(t *testing.T) {
+	// The date-time types take a bare timestamp only, so the text path is
+	// excluded for them and a quoted string stays a string.
+	var stamp struct {
+		S time.Time `toml:"s"`
+	}
+	err := Unmarshal([]byte("s = \"2026-06-26T10:00:00Z\"\n"), &stamp)
+	if err == nil {
+		t.Fatal("expected a quoted string to be rejected for time.Time")
+	}
+	if !strings.Contains(err.Error(), "cannot assign string") {
+		t.Errorf("err = %v, want a cannot-assign message", err)
+	}
+
+	var day struct {
+		D LocalDate `toml:"d"`
+	}
+	if err := Unmarshal([]byte("d = \"1979-05-27\"\n"), &day); err == nil {
+		t.Fatal("expected a quoted string to be rejected for LocalDate")
+	}
+}
+
+func TestDecoderMaxDepth(t *testing.T) {
+	deep := func(n int) []byte {
+		return []byte("v = " + strings.Repeat("[", n) + strings.Repeat("]", n) + "\n")
+	}
+	var cfg struct {
+		V any `toml:"v"`
+	}
+	if err := NewDecoder().MaxDepth(4).Decode(deep(4), &cfg); err != nil {
+		t.Fatalf("at the limit: %v", err)
+	}
+	err := NewDecoder().MaxDepth(4).Decode(deep(5), &cfg)
+	if err == nil {
+		t.Fatal("expected a nesting error")
+	}
+	if !strings.Contains(err.Error(), "limit of 4") {
+		t.Errorf("err = %v, want it to name the limit", err)
+	}
+}
+
+func TestDecoderMaxInputSize(t *testing.T) {
+	doc := []byte("v = \"ab\"\n")
+	var cfg struct {
+		V string `toml:"v"`
+	}
+	if err := NewDecoder().MaxInputSize(len(doc)).Decode(doc, &cfg); err != nil {
+		t.Fatalf("at the limit: %v", err)
+	}
+	err := NewDecoder().MaxInputSize(len(doc)-1).Decode(doc, &cfg)
+	if err == nil {
+		t.Fatal("expected a size error")
+	}
+	if !strings.Contains(err.Error(), "over the limit of 8") {
+		t.Errorf("err = %v, want it to name the limit", err)
+	}
+	// Parse carries the nesting default and no size limit.
+	if _, err := ParseMap(doc); err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+}
+
+// --- OffsetDateTime --------------------------------------------------------
+
+func TestOffsetDateTimeIsTheParsedType(t *testing.T) {
+	// A document's offset date-time arrives as the wrapper, and a plain
+	// time.Time destination still takes it, so a timestamp field needs no
+	// change to keep working.
+	in := []byte("stamp = 2026-06-26T10:00:00-07:00\n")
+	want := time.Date(2026, 6, 26, 10, 0, 0, 0, time.FixedZone("", -7*3600))
+
+	var plain struct {
+		Stamp time.Time `toml:"stamp"`
+	}
+	if err := Unmarshal(in, &plain); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if !plain.Stamp.Equal(want) {
+		t.Errorf("time.Time destination = %v, want %v", plain.Stamp, want)
+	}
+
+	var wrapped struct {
+		Stamp OffsetDateTime `toml:"stamp"`
+	}
+	if err := Unmarshal(in, &wrapped); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if !wrapped.Stamp.Time.Equal(want) {
+		t.Errorf("OffsetDateTime destination = %v, want %v", wrapped.Stamp.Time, want)
+	}
+	if got := wrapped.Stamp.String(); got != "2026-06-26T10:00-07:00" {
+		t.Errorf("String() = %q, want 2026-06-26T10:00-07:00", got)
+	}
+}
+
+func TestOffsetDateTimeFromHandBuiltTree(t *testing.T) {
+	// A tree built by hand may carry a plain time.Time, which is the other
+	// source of the offset kind; both date-time destinations take it.
+	tree := map[string]any{"stamp": time.Date(2026, 6, 26, 10, 0, 0, 0, time.UTC)}
+	want := time.Date(2026, 6, 26, 10, 0, 0, 0, time.UTC)
+
+	var plain struct {
+		Stamp time.Time `toml:"stamp"`
+	}
+	if err := newDecoder().decode(tree, &plain); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	if !plain.Stamp.Equal(want) {
+		t.Errorf("time.Time destination = %v, want %v", plain.Stamp, want)
+	}
+
+	var wrapped struct {
+		Stamp OffsetDateTime `toml:"stamp"`
+	}
+	if err := newDecoder().decode(tree, &wrapped); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	if !wrapped.Stamp.Time.Equal(want) {
+		t.Errorf("OffsetDateTime destination = %v, want %v", wrapped.Stamp.Time, want)
+	}
+}
+
+// dateKindReceiver records the Go type UnmarshalTOML was handed.
+type dateKindReceiver struct{ Kind string }
+
+func (r *dateKindReceiver) UnmarshalTOML(data any) error {
+	r.Kind = fmt.Sprintf("%T", data)
+	return nil
+}
+
+func TestUnmarshalerReceivesOffsetDateTime(t *testing.T) {
+	// The interface sees the wrapper, which names the date-time kind on its
+	// own; the local kinds keep their own wrappers.
+	var cfg struct {
+		O dateKindReceiver `toml:"o"`
+		L dateKindReceiver `toml:"l"`
+	}
+	in := []byte("o = 2026-06-26T10:00:00Z\nl = 2026-06-26T10:00:00\n")
+	if err := Unmarshal(in, &cfg); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if cfg.O.Kind != "interpres.OffsetDateTime" {
+		t.Errorf("offset kind = %q, want interpres.OffsetDateTime", cfg.O.Kind)
+	}
+	if cfg.L.Kind != "interpres.LocalDateTime" {
+		t.Errorf("local kind = %q, want interpres.LocalDateTime", cfg.L.Kind)
+	}
+}

docs/API.md

@@ -1,36 +1,108 @@
 # API
 
-The library exports the surface below from the `sourcedock.dev/petrbalvin/interpres`
+The library exports the surface below from the `sourcedock.dev/petrbalvin/interpres/v2`
 package. The snippets assume:
 
 ```go
-import "sourcedock.dev/petrbalvin/interpres"
+import "sourcedock.dev/petrbalvin/interpres/v2"
 ```
 
-The parser accepts TOML 1.0 documents plus the TOML 1.1 extensions: date-times
+The parser implements TOML 1.1: date-times and times without seconds, the
-and times without seconds, the `\e` and `\xHH` escape sequences, and
+`\e` and `\xHH` escape sequences, and multi-line inline tables with comments
-multi-line inline tables with comments and trailing commas. The encoder emits
+and trailing commas. The encoder emits TOML 1.1.
-TOML 1.0, which is valid under both versions.
 
 ## Functions
 
-### `func Parse(data []byte) (map[string]any, error)`
+### `func Parse(data []byte) (*Document, error)`
 
-Decodes a TOML document into an untyped tree, using the value mapping in the
+Decodes a TOML document into a [Document](#documents): the values, the order the
-[Decoding](#decoding) section below. Returns `*SyntaxError` on a malformed
+keys were written in, whether a table was written inline, and the comments.
-document. Input that is not valid UTF-8 is rejected before the parser runs.
+The values follow the mapping in the [Decoding](#decoding) section below.
-Equivalent to `ParseContext(context.Background(), data)`.
+Returns `*SyntaxError` on a malformed document. Input that is not valid UTF-8
+is rejected with a `SyntaxError` naming the line where the invalid byte
+appears, because validity is checked during the scan. Equivalent to
+`ParseContext(context.Background(), data)`.
 
 ```go
-tree, err := interpres.Parse([]byte("title = \"x\"\nport = 8080\n"))
+doc, err := interpres.Parse([]byte("title = \"x\"\nport = 8080\n"))
+tree := doc.Map()
 ```
 
-### `func ParseContext(ctx context.Context, data []byte) (map[string]any, error)`
+### `func ParseContext(ctx context.Context, data []byte) (*Document, error)`
 
 The cancellable variant of `Parse`. An already-cancelled context returns
 `ctx.Err()` before any work. During parsing the context is checked every 64
 top-level statements, so a long document aborts without running to completion.
 
+### `func ParseMap(data []byte) (map[string]any, error)`
+
+Decodes a TOML document into an untyped tree, the shape this package parsed
+into before [Document](#documents) existed: the order of the keys and the
+comments are not part of a map, so they are dropped. Use it when only the
+values matter, or when the extra bookkeeping of a document is not wanted.
+Equivalent to `ParseMapContext(context.Background(), data)`.
+
+```go
+tree, err := interpres.ParseMap([]byte("title = \"x\"\nport = 8080\n"))
+```
+
+### `func ParseMapContext(ctx context.Context, data []byte) (map[string]any, error)`
+
+The cancellable variant of `ParseMap`.
+
+## Documents
+
+`Parse` returns a `Document`: the value tree together with what a map cannot
+carry, which is the order the keys were written in, whether a table was written
+as an inline table or under a header, and the comments. `ParseMap` gives the
+plain tree when none of that is wanted.
+
+```go
+doc, err := interpres.Parse(data)
+if err != nil {
+	return err
+}
+root := doc.Root()
+for _, key := range root.Keys() { // written order, not sorted
+	entry, _ := root.Get(key)
+	fmt.Println(key, entry.Value())
+}
+```
+
+The values are shared with the tree `ParseMap` returns, so a value read from a
+document and from `doc.Map()` is the same value.
+
+| Type | Meaning |
+|---|---|
+| `Document` | the parsed document: `Root()` for the top-level table, `Map()` for the value tree, `Footer()` for a comment block at the end |
+| `Table` | one TOML table: `Keys()` and `Entries()` in written order, `Get(key)`, `Values()` for its part of the value tree, `Inline()` |
+| `Entry` | one key: `Value()`, `Inline()`, `Table()` when the value is a table, `Elements()` for the tables of an array value |
+
+`Elements()` holds one node per element of an array value: the tables of an
+array of tables, and the inline tables inside a value array, with `nil` for the
+elements that are not tables.
+
+### Comments
+
+A comment belongs to the line it precedes or follows, and to the node that line
+introduced:
+
+| Written | Carried by |
+|---|---|
+| lines above a key | that key's `Entry`, through `Comments()` |
+| a comment beside a key | that key's `Entry`, through `Trailing()` |
+| lines above a `[header]` or `[[header]]` | that `Table`, through `Comments()` |
+| a comment beside a header | that `Table`, through `Trailing()` |
+| a comment block after the last statement | the `Document`, through `Footer()` |
+
+`SetComments` and `SetTrailing` replace them. A line carries no leading `#`
+and no surrounding space, so `# note` is stored as `note` and a bare `#` as
+`""`.
+
+A `Document` is not a value to marshal: `Marshal` writes values, so it refuses
+one and points at `doc.Map()`. Writing a document back, with its order and its
+comments, belongs with the editing API.
+
 ### `func Unmarshal(data []byte, v any) error`
 
 Parses `data` and stores the result in the value pointed to by `v`, typically a
@@ -75,7 +147,7 @@ and every 64 fields during the reflection walk.
 | integer | `int64` |
 | float | `float64` |
 | boolean | `bool` |
-| offset date-time | `time.Time` |
+| offset date-time | `OffsetDateTime` |
 | local date-time | `LocalDateTime` |
 | local date | `LocalDate` |
 | local time | `LocalTime` |
@@ -134,20 +206,26 @@ The decoder converts to the destination type with explicit overflow checks:
 | `uint`, `uint8`, `uint16`, `uint32`, `uint64` | the value must be non-negative and must not overflow the destination's own width, `uint` on a 32-bit platform included; `uint64` accepts any non-negative `int64` |
 | `float32`, `float64` | copied verbatim, except that a finite value beyond the `float32` range is an overflow error rather than a silent infinity; an integer also coerces, so TOML `5` decodes into `5.0` |
 | `bool`, `string` | exact kind match only, no coercion across kinds |
-| `time.Time` | offset date-times only; no implicit conversion to or from the local variants |
+| `time.Time`, `OffsetDateTime` | offset date-times only; no implicit conversion to or from the local variants |
 
 A conversion that the rules do not allow produces an error wrapped with the
 offending key or index, for example `p: interpres: integer 300 overflows uint8`.
 
 ### Date-time values
 
-Offset date-times decode into `time.Time` and keep their offset. The local
+Offset date-times decode into `OffsetDateTime`, whose embedded `time.Time` is the
-variants decode into `LocalDateTime`, `LocalDate` and `LocalTime`, whose
+instant with the offset the document wrote; a destination of the plain
-embedded `time.Time` is normalised to UTC (midnight UTC for a local date, the
+`time.Time` takes the same value, so a timestamp field does not have to name the
-zero date for a local time). Every kind may omit the seconds as of TOML 1.1
+wrapper. The local variants decode into `LocalDateTime`, `LocalDate` and
-(`07:32`, `1979-05-27T07:32`); such a value carries a zero second, and the
+`LocalTime`, whose embedded `time.Time` is normalised to UTC (midnight UTC for a
-canonical rendering writes full seconds. There is no implicit conversion
+local date, the zero date for a local time). Every kind may omit the seconds as
-between the offset and local kinds; assigning one to the other is an error.
+of TOML 1.1 (`07:32`, `1979-05-27T07:32`); such a value carries a zero second,
+and the encoder writes the seconds only when the value carries them, so a
+document written without seconds comes back without them. There is no implicit
+conversion between the offset and local kinds; assigning one to the other is an
+error. The date-time types take a bare timestamp and never a quoted string, so a
+document that writes a date-time with quotes does not decode into them, and
+neither `encoding.TextUnmarshaler` nor the embedded `time.Time` changes that.
 
 ### Arrays of tables
 
@@ -167,7 +245,7 @@ type Unmarshaler interface {
 ```
 
 `data` is whatever the parser produced for that key: `string`, `bool`, `int64`,
-`float64`, `time.Time`, `LocalDateTime`, `LocalDate`, `LocalTime`, `[]any`, or
+`float64`, `OffsetDateTime`, `LocalDateTime`, `LocalDate`, `LocalTime`, `[]any`, or
 `map[string]any`. The method inspects the value and mutates its own receiver;
 the decoder keeps whatever state the receiver stored.
 
@@ -178,6 +256,38 @@ automatically, and a nil pointer destination is allocated first. An error
 returned from `UnmarshalTOML` halts the decode and propagates wrapped with the
 key path, for example `addr: unmarshal: not a string`.
 
+### Custom decoding: `encoding.TextUnmarshaler`
+
+A destination type that implements `encoding.TextUnmarshaler` receives a TOML
+string as its text content, the rule `encoding/json` follows:
+
+```go
+func (ip *IP) UnmarshalText(text []byte) error
+```
+
+The decoder looks for the method on the destination and on its address, so a
+pointer-receiver `UnmarshalText` is invoked on an addressable struct field, and
+the elements of a slice destination are reached the same way. The text path
+applies to TOML strings only: every other value kind keeps its own rule, so
+`r = 1` does not reach a receiver that expects text. An error from
+`UnmarshalText` halts the decode and propagates with the key path and the
+prefix `unmarshal text:`, for example `addr: unmarshal text: not an address`.
+
+[`UnmarshalTOML`](#custom-decoding-unmarshaler) wins over `UnmarshalText` when
+a type implements both, and the four [date-time
+types](#date-time-values) are excluded: a quoted string stays a string and
+never becomes an `OffsetDateTime` or one of the local wrappers.
+
+### Durations
+
+TOML has no duration type, so `time.Duration` has a rule of its own. The
+encoder writes the canonical Go form in a TOML string, `1h30m0s`, and the
+decoder reads that string back with `time.ParseDuration`. A bare integer is
+still the nanosecond count it has always been, so `from_int = 5400000000000`
+and `from_text = "1h30m"` decode to the same duration. Text that
+`time.ParseDuration` rejects, `d = "90"` among it, fails with
+`interpres: invalid duration "90"`.
+
 ### Strict decoding
 
 By default unknown keys are dropped silently. A `Decoder` built with
@@ -281,7 +391,7 @@ them.
 By default every table is emitted with its entries grouped by kind:
 
 1. scalars (`string`, `int64`, `float64`, `bool`, `time.Time`,
-   `LocalDateTime`, `LocalDate`, `LocalTime`)
+   `OffsetDateTime`, `LocalDateTime`, `LocalDate`, `LocalTime`)
 2. sub-tables (structs and `map[string]V` values)
 3. arrays of tables (`[]struct` and `[]map[string]V`)
 
@@ -316,11 +426,20 @@ type Marshaler interface {
 The returned value is encoded as if it had been passed in place of the
 receiver, so it may be a scalar, a slice, an array of tables, or another
 struct or map, including the `Marshaler` result of another type; the encoder
-recurses. An error returned from `MarshalTOML` fails the marshal wrapped with
+recurses, and the result is normalised like any other value, so a method may
-the key path, for example `interpres: server.port: bad timestamp`. A result
+return a plain `int` or a `time.Duration`.
-of `nil` with a nil error fails the same way with
+
-`MarshalTOML returned a nil value`: nil has no TOML representation, so
+An error returned from `MarshalTOML` fails the marshal wrapped with the key
-dropping the field silently is not an option.
+path, for example `interpres: server.port: bad timestamp`. A result of `nil` with
+a nil error fails the same way with `MarshalTOML returned a nil value`: nil has
+no TOML representation, so dropping the field silently is not an option.
+
+The method is reached for every value the walk meets, array elements included:
+an element that renders itself as a table keeps the `[[header]]` form, one that
+renders itself as a scalar turns the array into a value array, and the method
+runs once per element. It is looked up on the value and on its address, so a
+pointer-receiver method is called for a field or an element, exactly as
+`MarshalText` is.
 
 ```go
 type Port int
@@ -330,6 +449,27 @@ func (p Port) MarshalTOML() (any, error) {
 }
 ```
 
+### Custom encoding: `encoding.TextMarshaler`
+
+A type that implements `encoding.TextMarshaler` is encoded as a TOML string
+holding the text the method returns, which is the rule `encoding/json` follows:
+
+```go
+func (ip IP) MarshalText() ([]byte, error)
+```
+
+The encoder looks for the method on the value and on its address, so a
+pointer-receiver `MarshalText` is found on a struct field of an addressable
+value (pass a pointer to `Marshal`) and always on a slice element. `net.IP`,
+`netip.Addr` and user types follow this rule, and a struct that implements the
+interface becomes a string rather than a table. `MarshalTOML` wins when a type
+implements both, the four [date-time types](#date-time-values) keep their bare
+timestamp form, and text that is not valid UTF-8 is an error rather than a
+replacement character.
+
+A duration carries no text method of its own; see [Durations](#durations) for
+its rule.
+
 ### Arrays
 
 An array whose every element is a table (`[]struct`, `[]map[string]V`, after
@@ -357,10 +497,9 @@ omitted, because TOML forbids an empty `[[a]]`. Other empty arrays emit as
 ### Long strings
 
 By default every string is emitted as a basic `"..."` string with the escapes
-TOML requires, and a string containing a newline is emitted as an escaped
+TOML requires, a newline among them as `\n`. `UseLiteralMultiline(threshold)`
-multi-line basic string. `UseLiteralMultiline(threshold)` switches strings that
+switches strings that contain a newline and are at least `threshold` bytes long
-contain a newline and are at least `threshold` bytes long to the literal
+to the literal `'''...'''` form, which carries the newlines verbatim:
-`'''...'''` form, which carries the newlines verbatim:
 
 ```go
 out, err := interpres.NewEncoder().UseLiteralMultiline(80).Marshal(cfg)
@@ -372,6 +511,48 @@ carry verbatim (an embedded run of three single quotes, a control character
 other than tab or newline, or a carriage return outside a CRLF pair) also keeps
 the basic form, so the output always re-parses to the same value.
 
+### Inline tables
+
+A table element of a value array, and a sub-table inlined by
+[`InlineTables`](#compact-documents), is written as one `{a = 1, b = 2}` line
+while it fits. An inline table that would pass the hundredth column carries
+newlines and a trailing comma instead, which TOML 1.1 allows:
+
+```toml
+arr = [1, {
+	n = 1,
+	name = "a value long enough to push this line well past the one hundred column limit",
+}]
+```
+
+The closing brace and the entries are indented one tab per nesting level, a
+nested table is measured on its own line, and the output re-parses to the same
+value either way.
+
+### Compact documents
+
+`InlineTables(threshold)` writes a sub-table as an inline table when its
+single-line rendering is at most `threshold` bytes, and as a table header
+section when it is longer. A document of small tables therefore grows shorter:
+
+```go
+out, err := interpres.NewEncoder().InlineTables(60).Marshal(cfg)
+```
+
+With `60` and a table of three short entries, the same value is written
+
+```toml
+server = {host = "127.0.0.1", port = 9090, tls = {on = false}}
+```
+
+instead of three lines under a `[server]` header and a `[server.tls]` section.
+A nested sub-table takes part in the same way, and the whole option is off at
+`0` or less. Two limits are deliberate. An array of tables keeps the `[[a]]`
+header form, because its inline form re-parses as a value array and would change
+the value's Go type. And because an inlined table is a value line, every one of
+them precedes the first header of its document, so a table inlined next to a
+header is not read back as part of that header's section.
+
 ### Cancellation
 
 `MarshalContext` and `(*Encoder).MarshalContext` accept a `context.Context`. The
@@ -386,6 +567,8 @@ The output is not byte-identical to any document that produced the value:
 - map keys are emitted in sorted order
 - the choice between `[table]` headers and inline tables is not preserved
 - strings use the basic quoted form unless the literal option above applies
+- a date-time drops its zero seconds and the trailing zeros of its fraction, so
+  `07:32:00` is written `07:32`; both are the same value
 - floats always carry a `.` or an exponent, so a float `1` is emitted as `1.0`
   and stays distinguishable from the integer `1` across a round-trip; negative
   zero is normalised to `0.0`
@@ -413,9 +596,11 @@ sequenceDiagram
 
 ### `type SyntaxError struct{ Line int; Msg string }`
 
-Describes a malformed TOML document; `Line` is 1-based and `Error()` renders as
+Describes a document the parser rejected, with the 1-based `Line` at which it
-`interpres: line N: msg`. Read the structured fields with a type assertion or
+gave up and `Error()` rendering as `interpres: line N: msg`. A malformed
-`errors.AsType`:
+document is the usual cause; the nesting limit and an input that is not valid
+UTF-8 report through the same type. Read the structured fields with a type
+assertion or `errors.AsType`:
 
 ```go
 if se, ok := errors.AsType[*interpres.SyntaxError](err); ok {
@@ -451,6 +636,19 @@ with `DisallowUnknownFields`, then call `Decode` or `DecodeContext` any number
 of times. A configured `Decoder` holds no per-call state and is safe for
 concurrent use.
 
+| Method | Default | Effect |
+|---|---|---|
+| `DisallowUnknownFields()` | off | a key with no matching struct field is an error |
+| `MaxDepth(depth int)` | `10000` | bound how deeply arrays and inline tables may nest |
+| `MaxInputSize(size int)` | no limit | bound the size of the document, in bytes |
+
+The nesting limit protects the stack, because the parser is a recursive
+descent: a deeper document is rejected with a `SyntaxError` naming the limit
+rather than running the stack out. `Parse` and `ParseContext` carry that same
+default but take no options. The size limit is off by default, because the
+caller already holds the bytes and the size is therefore a policy, not a
+protection the library can impose on its own.
+
 ### `type Encoder`
 
 Configurable emission policy, constructed with `NewEncoder`. The option state
@@ -462,12 +660,14 @@ encoder:
 | `GroupByKind(v bool)` | `true` | group entries as scalars, then sub-tables, then arrays of tables; `false` preserves declaration order |
 | `OmitEmptyArrays()` | off | skip `key = []` for empty scalar arrays |
 | `UseLiteralMultiline(threshold int)` | `0` | emit multi-line strings of at least `threshold` bytes as literal `'''...'''` |
+| `InlineTables(threshold int)` | `0` | write a sub-table inline when its single-line form is at most `threshold` bytes |
 
 ```go
 out, err := interpres.NewEncoder().
 	GroupByKind(false).
 	OmitEmptyArrays().
 	UseLiteralMultiline(80).
+	InlineTables(60).
 	MarshalContext(ctx, cfg)
 ```
 
@@ -475,6 +675,11 @@ A configured `Encoder` holds no per-call state; each `Marshal` or
 `MarshalContext` call copies the options and is safe for concurrent use, as
 long as no setter races with a call.
 
+### `type Document`, `type Table`, `type Entry`
+
+See [Documents](#documents). A `Document` is what `Parse` returns, and it is
+not a value `Marshal` accepts.
+
 ### `type Marshaler interface{ MarshalTOML() (any, error) }`
 
 See [Custom encoding](#custom-encoding-marshaler).
@@ -486,24 +691,28 @@ See [Custom decoding](#custom-decoding-unmarshaler).
 ### Date-time wrappers
 
 ```go
+type OffsetDateTime struct{ time.Time } // 1979-05-27T07:32:00Z
 type LocalDateTime struct{ time.Time }  // 1979-05-27T07:32:00
 type LocalDate     struct{ time.Time }  // 1979-05-27
 type LocalTime     struct{ time.Time }  // 07:32:00.999999
 ```
 
-Each carries a `String()` method returning the TOML-canonical rendering, with
+Each carries a `String()` method returning the TOML-canonical rendering: the
-the fractional second zero-padded to nanosecond precision when present. The
+seconds appear only when the value carries them, and a fractional second drops
-types are produced by `Parse` and accepted by `Marshal`.
+its trailing zeros, so `07:32:00` renders as `07:32` and a half second as
+`00.5`. The types are produced by `Parse` and accepted by `Marshal`, which
+writes them through `String()`.
 
 ## Errors
 
 The entry points return:
 
-- `*SyntaxError` for a malformed document, with the 1-based line
+- `*SyntaxError` for a malformed document, with the 1-based line; the nesting
+  limit reports through it as well
 - `*DecodeError` for a decoding failure, with the key path in `Path`
 - `*EncodeError` for an encoding failure, with the key path in `Path`
 - a plain error for the rest: a non-pointer decode target, a cancelled
-  context, a key that is not valid UTF-8
+  context, a key that is not valid UTF-8, an input over the size limit
 
 Decode and encode failures carry the key path or element index in the typed
 wrappers above, so `errors.Is` and `errors.AsType` see through them and the

docs/ARCHITECTURE.md

@@ -6,10 +6,10 @@ source tree; nothing is aspirational.
 ## Overview
 
 interpres is one public library package, one command, and one example. The
-library implements the whole of TOML 1.0 and 1.1, decoding and encoding, in the
+library implements the whole of TOML 1.1, decoding and encoding, in the
-standard library alone; the command wraps the parser for the toml-test
+standard library alone; the command wraps the parser and the encoder for the
-compliance harness, against which it stands at 214 valid and 467 invalid cases
+toml-test compliance harness, against which it stands at 214 valid, 467 invalid
-with zero failures; the example demonstrates the API.
+and 214 encoder cases with zero failures; the example demonstrates the API.
 
 ```mermaid
 flowchart TD
@@ -36,14 +36,15 @@ strict validation.
 | Path | Responsibility |
 |---|---|
 | `.` (package `interpres`) | The whole library. `interpres.go` declares the exported surface (`Parse`, `Unmarshal`, `Marshal`, the `*Context` variants, `Decoder`, `Encoder`, `Marshaler`, `Unmarshaler`, `SyntaxError`, the local date-time types); everything below it is unexported. |
-| `cmd/interpres-decode` | The toml-test adapter. Reads TOML on stdin, writes tagged JSON on stdout. Owns no parsing logic. |
+| `cmd/interpres-decode` | The toml-test adapter, both directions. Reads TOML on stdin, writes tagged JSON on stdout; with `-encode` it reads tagged JSON and writes TOML. Owns no parsing logic and no emission logic. |
 | `examples/basic` | A runnable tour of the API. Documentation in executable form, not part of the library. |
 
 Inside the library package, one file owns one concern:
 
 | File | Responsibility |
 |---|---|
-| `parser.go` | The recursive-descent parser. Produces the `map[string]any` tree and enforces the structural rules of TOML 1.0 and 1.1 (table redefinitions, dotted keys, arrays of tables, multi-line inline tables). Reports a 1-based line on failure. |
+| `parser.go` | The recursive-descent parser. Produces the `map[string]any` tree, records the nodes a [Document](API.md#documents) is built from, and enforces the structural rules of TOML 1.1 (table redefinitions, dotted keys, arrays of tables, multi-line inline tables). Reports a 1-based line on failure. |
+| `document.go` | The parsed-document types: `Document`, `Table` and `Entry`, which carry the key order, whether a table was written inline, and the comments. The values they expose are the parser's own tree, not a copy. |
 | `number.go` | Strict numeric tokens: integers in the four radixes with `_` separators, and floats including `inf` and `nan`. Rejects leading zeros, misplaced underscores and malformed fractions. |
 | `datetime.go` | The three local date-time wrapper types and `parseDateTime`, which classifies a token into the four date-time kinds under the strict TOML grammar. |
 | `decode.go` | Maps the parsed tree onto Go values by reflection: struct fields, maps, slices, scalar conversion with overflow checks, `Unmarshaler` dispatch. |
@@ -101,15 +102,23 @@ sequenceDiagram
   setter methods are not, and must finish before the value is shared.
 - The parser is allocated per `ParseContext` call; the parser itself caches
   nothing between documents.
-- The one piece of shared state is the struct-schema cache in `decode.go`: a
+- The shared state is a set of caches and pools whose entries are immutable
-  `sync.Map` keyed by `reflect.Type`, holding the flattened field layout the
+  once published, each growing with the number of distinct types rather than
-  decoder and the encoder both consult. A schema is immutable once published,
+  with document size: the struct-schema cache in `decode.go` (a `sync.Map`
-  so concurrent callers only race to build an identical value, the same
+  keyed on `reflect.Type`, holding the flattened field layout the decoder and
-  trade-off `encoding/json`'s field cache makes. The cache grows with the
+  the encoder both consult), the per-type interface flag caches in `decode.go`
-  number of distinct struct types, never with document size.
+  and `encode.go` (recording where `Marshaler`, `Unmarshaler` and the text
+  interfaces can be found, so a walk builds an interface value only where the
+  assertion can succeed), each fronted by a monomorphic hint holding the type
+  resolved last, and the encoder's output-buffer pool in `encode.go`
+  (`sync.Pool`, buffers returned to it only within a 1 MiB retention cap). A
+  published schema or flag set never mutates, so concurrent callers only race
+  to build an identical value, the same trade-off `encoding/json`'s field
+  cache makes.
 - The date-time wrappers are values, not pointers, and are immutable in use.
-- Nothing in the library starts goroutines; apart from the schema cache above,
+- Nothing in the library starts goroutines; apart from the caches and the pool
-  which never mutates a published entry, there is no shared mutable state.
+  above, which never mutate a published entry, there is no shared mutable
+  state.
 
 ## Dependencies
 

docs/BENCHMARKING.md

@@ -9,10 +9,12 @@ The benchmarks live in `bench_test.go`, next to the code they measure:
 
 | Benchmark | What it measures |
 |---|---|
-| `BenchmarkParse` | `Parse` over a representative configuration document |
+| `BenchmarkParse` | `ParseMap` over a representative configuration document |
-| `BenchmarkMarshal` | `Marshal` of the tree `Parse` produced from the same document |
+| `BenchmarkMarshal` | `Marshal` of the tree `ParseMap` produced from the same document |
 | `BenchmarkStrictDecode` | `Decode` into a struct under `DisallowUnknownFields` |
-| `BenchmarkParseLong` | `Parse` over a generated document with about 2000 array-of-tables entries |
+| `BenchmarkParseLong` | `ParseMap` over a generated document with about 2000 array-of-tables entries |
+| `BenchmarkStrictDecodeLong` | `Decode` into a typed document under `DisallowUnknownFields`, over the same long document |
+| `BenchmarkMarshalLong` | `Marshal` of the tree `ParseMap` produced from the long document |
 
 ## Running
 

docs/CLI.md

@@ -1,25 +1,32 @@
 # Command line
 
 The reference below is taken from the program itself. `interpres-decode` is
-the toml-test harness adapter, and it also validates documents. Install it
+the toml-test harness adapter in both directions, decoding TOML into tagged
-with Go itself, no release assets involved:
+JSON and encoding tagged JSON back into TOML, and it also validates documents.
+Install it with Go itself, no release assets involved:
 
 ```sh
-go install sourcedock.dev/petrbalvin/interpres/cmd/interpres-decode@latest
+go install sourcedock.dev/petrbalvin/interpres/v2/cmd/interpres-decode@latest
 ```
 
 ## Synopsis
 
 ```sh
 interpres-decode [flags]
+interpres-decode -encode
 interpres-decode -validate [file ...]
 ```
 
-Without `-validate` the program is the toml-test adapter: it takes no
+Without `-validate` or `-encode` the program is the decoding half of the
-arguments, reads one TOML document from stdin, and writes the toml-test
+toml-test adapter: it takes no arguments, reads one TOML document from stdin,
-tagged-JSON form to stdout. Build it locally with `just build`, which
+and writes the toml-test tagged-JSON form to stdout. Build it locally with
-compiles it into `bin/interpres-decode`, or run it straight from the module
+`just build`, which compiles it into `bin/interpres-decode`, or run it
-directory with `just run`.
+straight from the module directory with `just run`.
+
+With `-encode` the direction is reversed: the program reads a tagged-JSON
+description from stdin and writes the TOML document it describes to stdout,
+which is the shape toml-test expects of an encoder command. It takes no
+arguments either, and `-validate` and `-encode` cannot be combined.
 
 With `-validate` the program parses each named file instead, or stdin when no
 file is named, and prints one line per invalid document to stderr. It is
@@ -31,15 +38,16 @@ means stdin.
 | Flag | Effect |
 |---|---|
 | `-validate` | validate the documents instead of emitting tagged JSON |
+| `-encode` | read tagged JSON from stdin and write TOML instead |
 | `-h` | print the usage |
 
 ## Exit codes
 
 | Code | Meaning |
 |---|---|
-| `0` | adapter: the document parsed and the tagged JSON was written; validate: every document parsed |
+| `0` | adapter: the document parsed and the tagged JSON was written; encode: the TOML was written; validate: every document parsed |
 | `1` | adapter: parse error; validate: at least one document is invalid |
-| `2` | a usage error, a read failure, or a value with no tagged representation |
+| `2` | a usage error, a read failure, malformed tagged JSON, or a value with no TOML representation |
 
 ## Wire format
 
@@ -66,6 +74,14 @@ wrapped in an object with a `type` and a `value`:
 | local date | `date-local` | `1979-05-27` |
 | local time | `time-local` | `07:32:00.999999` |
 
+The `-encode` mode reads exactly this form back. Two properties of it are
+worth knowing. A float whose value has no fraction and no exponent is written
+as a bare integer string, `{"type": "float", "value": "1"}`, so there the tag
+decides the type and not the literal. And the form cannot tell an array of
+tables from a value array of inline tables, so the adapter writes the header
+form, `[[a]]`, for an array whose every element is a JSON object; a mixed
+array keeps the value form.
+
 ## Examples
 
 Echo a small document through the adapter:
@@ -78,8 +94,18 @@ port = 9090
 ' | ./bin/interpres-decode
 ```
 
-The output is the equivalent value tree as one JSON object. Validate the
+The output is the equivalent value tree as one JSON object. Turn a description
-TOML files of another repository in CI:
+back into TOML with `-encode`:
+
+```sh
+echo '{"title": {"type": "string", "value": "hello"}}' | ./bin/interpres-decode -encode
+```
+
+```toml
+title = "hello"
+```
+
+Validate the TOML files of another repository in CI:
 
 ```sh
 interpres-decode -validate config.toml deploy/example.toml
@@ -101,5 +127,7 @@ just toml-test
 ```
 
 That recipe needs the `toml-test` binary on `PATH`, installed with
-`go install github.com/toml-lang/toml-test/v2/cmd/toml-test@v2.2.0`. The full
+`go install github.com/toml-lang/toml-test/v2/cmd/toml-test@v2.2.0`. It runs
+the suite in both directions: the decoder against the valid and invalid
+corpora, and the encoder against the tagged JSON of the valid one. The full
 reference for the library itself is [API.md](API.md).

docs/DEVELOPMENT.md

@@ -41,7 +41,7 @@ prints the same list.
 | `just run` | `go run ./cmd/interpres-decode`, reads TOML from stdin |
 | `just dev` | the same run, for iterating |
 | `just example` | `go run ./examples/basic`, the usage tour |
-| `just toml-test` | builds the adapter and runs the official toml-test compliance suite against it |
+| `just toml-test` | builds the adapter and runs the official toml-test compliance suite against it, decoder and encoder |
 | `just coverage-html` | `just test`, then `go tool cover -html` into `coverage.html` |
 | `just install` | builds, then copies the binary into `~/.local/bin` (`BINDIR` overrides) |
 | `just uninstall` | removes the installed binary |

document.go

@@ -0,0 +1,196 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: MIT
+
+package interpres
+
+// A Document is a parsed TOML document: the values, plus what the map shape
+// cannot carry, which is the order the keys were written in, whether a table
+// was written inline or under a header, and the comments.
+//
+// The values are the tree ParseMap returns, shared rather than copied, so a
+// value read from a Document and from that map is the same value. A comment
+// belongs to the statement it precedes: the lines above a key belong to the
+// key, the lines above a header belong to the header's table, and a comment
+// block after the last statement belongs to the Document.
+//
+// Comments inside array and inline-table values are not carried yet; the
+// parser skips them as it always has.
+type Document struct {
+	root   *Table
+	footer []string
+}
+
+// Root returns the document's root table.
+func (d *Document) Root() *Table { return d.root }
+
+// Map returns the value tree, the shape ParseMap gives. It is the tree the
+// document was parsed into, not a copy.
+func (d *Document) Map() map[string]any { return d.root.values }
+
+// Footer returns the comment lines that follow the last statement, and every
+// line of a document that holds no statement at all.
+func (d *Document) Footer() []string { return d.footer }
+
+// SetFooter replaces those lines.
+func (d *Document) SetFooter(lines []string) { d.footer = lines }
+
+// A Table is one TOML table: its values, its keys in written order, and the
+// comments around the header or the key that introduced it.
+type Table struct {
+	values  map[string]any
+	entries []*Entry
+	index   map[string]*Entry
+
+	// inline records that the table was written as an inline table, `{…}`,
+	// rather than under a header or as a dotted key.
+	inline bool
+
+	// comments are the lines above the table's header, trailing is the comment
+	// on the header's own line. Both are empty for a table a dotted key
+	// introduced, which has no line of its own.
+	comments []string
+	trailing string
+}
+
+func newTable(values map[string]any) *Table {
+	return &Table{values: values, index: map[string]*Entry{}}
+}
+
+// Keys returns the table's keys in the order they were written.
+func (t *Table) Keys() []string {
+	keys := make([]string, len(t.entries))
+	for i, e := range t.entries {
+		keys[i] = e.key
+	}
+	return keys
+}
+
+// Values returns the table's values, which is the map the value tree holds for
+// it.
+func (t *Table) Values() map[string]any { return t.values }
+
+// Entries returns the table's entries in written order.
+func (t *Table) Entries() []*Entry { return t.entries }
+
+// Get returns the entry for key, and whether the table has one.
+func (t *Table) Get(key string) (*Entry, bool) {
+	e, ok := t.index[key]
+	return e, ok
+}
+
+// Inline reports whether the table was written as an inline table, `{…}`,
+// rather than under a header or introduced by a dotted key.
+func (t *Table) Inline() bool { return t.inline }
+
+// Comments returns the comment lines above the table's header, or above the
+// key that introduced it. Lines carry no leading '#' and no surrounding space.
+func (t *Table) Comments() []string { return t.comments }
+
+// SetComments replaces those lines. Each line is written back with a "# " in
+// front of it, so a line should not carry one.
+func (t *Table) SetComments(lines []string) { t.comments = lines }
+
+// Trailing returns the comment on the header's own line, without the '#'.
+func (t *Table) Trailing() string { return t.trailing }
+
+// SetTrailing replaces that comment.
+func (t *Table) SetTrailing(line string) { t.trailing = line }
+
+// addValue records a key of the table, in written order.
+func (t *Table) addValue(key string, val any, inline bool) *Entry {
+	e := &Entry{table: t, key: key, inline: inline}
+	t.entries = append(t.entries, e)
+	t.index[key] = e
+	if node, ok := val.(map[string]any); ok {
+		e.child = newTable(node)
+	}
+	return e
+}
+
+// addTable records a key whose value is a table introduced by a header or a
+// dotted key, and returns the table's node.
+func (t *Table) addTable(key string, values map[string]any) *Table {
+	if e, ok := t.index[key]; ok {
+		// The key was seen before, as the leaf of an earlier dotted key.
+		if e.child == nil {
+			e.child = newTable(values)
+		}
+		return e.child
+	}
+	e := t.addValue(key, values, false)
+	e.child = newTable(values)
+	return e.child
+}
+
+// addElement records one element of an array of tables, and returns its node.
+func (t *Table) addElement(key string, values map[string]any) *Table {
+	e, ok := t.index[key]
+	if !ok {
+		e = t.addValue(key, nil, false)
+		e.elements = []*Table{}
+	}
+	el := newTable(values)
+	e.elements = append(e.elements, el)
+	return el
+}
+
+// child returns the node of a table-valued key, or nil.
+func (t *Table) child(key string) *Table {
+	if e, ok := t.index[key]; ok {
+		return e.child
+	}
+	return nil
+}
+
+// lastElement returns the node of the newest element of an array of tables.
+func (t *Table) lastElement(key string) *Table {
+	if e, ok := t.index[key]; ok && len(e.elements) > 0 {
+		return e.elements[len(e.elements)-1]
+	}
+	return nil
+}
+
+// An Entry is one key of a table: the value and the comments around the key.
+type Entry struct {
+	table    *Table
+	key      string
+	inline   bool
+	comments []string
+	trailing string
+
+	// child is the table the value is, and elements are the tables of an array
+	// of tables; one of them is set only when the value has that shape.
+	child    *Table
+	elements []*Table
+}
+
+// Key returns the key as it was written.
+func (e *Entry) Key() string { return e.key }
+
+// Value returns the value the key holds. It is read from the table's map, so
+// it stays current if that map is changed.
+func (e *Entry) Value() any { return e.table.values[e.key] }
+
+// Inline reports whether the value was written as an inline table, `{…}`.
+func (e *Entry) Inline() bool { return e.inline }
+
+// Table returns the table the value is, or nil when it is not a table.
+func (e *Entry) Table() *Table { return e.child }
+
+// Elements returns the tables of an array of tables, or nil when the value is
+// not one.
+func (e *Entry) Elements() []*Table { return e.elements }
+
+// Comments returns the comment lines above the key. Lines carry no leading '#'
+// and no surrounding space.
+func (e *Entry) Comments() []string { return e.comments }
+
+// SetComments replaces those lines. Each line is written back with a "# " in
+// front of it, so a line should not carry one.
+func (e *Entry) SetComments(lines []string) { e.comments = lines }
+
+// Trailing returns the comment on the key's own line, without the '#'.
+func (e *Entry) Trailing() string { return e.trailing }
+
+// SetTrailing replaces that comment.
+func (e *Entry) SetTrailing(line string) { e.trailing = line }

document_test.go

@@ -0,0 +1,314 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: MIT
+
+package interpres
+
+import (
+	"slices"
+	"strings"
+	"testing"
+)
+
+// mustEntry returns the entry a table must have, and fails the test when it
+// does not.
+func mustEntry(t *testing.T, tbl *Table, key string) *Entry {
+	t.Helper()
+	e, ok := tbl.Get(key)
+	if !ok {
+		t.Fatalf("%q is missing from the table", key)
+	}
+	return e
+}
+
+func TestDocumentKeepsKeyOrder(t *testing.T) {
+	doc, err := Parse([]byte(`
+b = 1
+a = 2
+inline = {y = 1, x = 2}
+
+[table]
+z = 3
+m = 4
+`))
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+
+	// The root's keys come back in written order, not sorted.
+	if got := doc.Root().Keys(); !slices.Equal(got, []string{"b", "a", "inline", "table"}) {
+		t.Errorf("root keys = %v, want [b a inline table]", got)
+	}
+
+	// So do the keys of an inline table, which the map shape loses.
+	inline, ok := doc.Root().Get("inline")
+	if !ok {
+		t.Fatal("inline is missing from the root")
+	}
+	if !inline.Inline() {
+		t.Error("inline is not marked inline")
+	}
+	if got := inline.Table().Keys(); !slices.Equal(got, []string{"y", "x"}) {
+		t.Errorf("inline keys = %v, want [y x]", got)
+	}
+
+	// And the keys of a table written under a header, which is not inline.
+	tbl, ok := doc.Root().Get("table")
+	if !ok {
+		t.Fatal("table is missing from the root")
+	}
+	if tbl.Inline() {
+		t.Error("table is marked inline")
+	}
+	if got := tbl.Table().Keys(); !slices.Equal(got, []string{"z", "m"}) {
+		t.Errorf("table keys = %v, want [z m]", got)
+	}
+}
+
+func TestDocumentValuesAreTheTree(t *testing.T) {
+	doc, err := Parse([]byte("n = 7\ns = \"x\"\n\n[t]\nk = true\n"))
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if got := mustEntry(t, doc.Root(), "n").Value(); got != int64(7) {
+		t.Errorf("n = %#v, want int64(7)", got)
+	}
+	tbl := mustEntry(t, doc.Root(), "t").Table()
+	if got := mustEntry(t, tbl, "k").Value(); got != true {
+		t.Errorf("t.k = %#v, want true", got)
+	}
+	// Map gives the tree ParseMap would have returned, the same values.
+	tree := doc.Map()
+	if tree["n"] != int64(7) || tree["s"] != "x" {
+		t.Errorf("Map = %#v", tree)
+	}
+	if tree["t"].(map[string]any)["k"] != true {
+		t.Errorf("Map[t] = %#v", tree["t"])
+	}
+	if tbl.Values()["k"] != true {
+		t.Errorf("t.Values() = %#v", tbl.Values())
+	}
+}
+
+func TestDocumentComments(t *testing.T) {
+	doc, err := Parse([]byte(`# above b
+b = 1 # trailing b
+
+# above the table
+[table] # trailing table
+# above m
+m = 2
+
+# footer
+`))
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+
+	b, ok := doc.Root().Get("b")
+	if !ok {
+		t.Fatal("b is missing")
+	}
+	if got := b.Comments(); !slices.Equal(got, []string{"above b"}) {
+		t.Errorf("b comments = %q, want [above b]", got)
+	}
+	if got := b.Trailing(); got != "trailing b" {
+		t.Errorf("b trailing = %q, want \"trailing b\"", got)
+	}
+
+	tbl, ok := doc.Root().Get("table")
+	if !ok {
+		t.Fatal("table is missing")
+	}
+	// A [header] line introduces the table, so the comments around it belong
+	// to the table node; the entry that names it stays bare.
+	if got := tbl.Table().Comments(); !slices.Equal(got, []string{"above the table"}) {
+		t.Errorf("table comments = %q, want [above the table]", got)
+	}
+	if got := tbl.Table().Trailing(); got != "trailing table" {
+		t.Errorf("table trailing = %q, want \"trailing table\"", got)
+	}
+	if got := tbl.Comments(); got != nil {
+		t.Errorf("entry comments = %q, want none", got)
+	}
+	m, ok := tbl.Table().Get("m")
+	if !ok {
+		t.Fatal("table.m is missing")
+	}
+	if got := m.Comments(); !slices.Equal(got, []string{"above m"}) {
+		t.Errorf("m comments = %q, want [above m]", got)
+	}
+
+	if got := doc.Footer(); !slices.Equal(got, []string{"footer"}) {
+		t.Errorf("footer = %q, want [footer]", got)
+	}
+}
+
+func TestDocumentCommentsAreWritable(t *testing.T) {
+	doc, err := Parse([]byte("# above\nk = 1\n"))
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	entry, ok := doc.Root().Get("k")
+	if !ok {
+		t.Fatal("k is missing")
+	}
+	entry.SetComments([]string{"first", "second"})
+	entry.SetTrailing("beside")
+	if got := entry.Comments(); !slices.Equal(got, []string{"first", "second"}) {
+		t.Errorf("comments = %q", got)
+	}
+	if got := entry.Trailing(); got != "beside" {
+		t.Errorf("trailing = %q", got)
+	}
+
+	tbl := doc.Root()
+	tbl.SetComments([]string{"above the root"})
+	if got := tbl.Comments(); !slices.Equal(got, []string{"above the root"}) {
+		t.Errorf("root comments = %q", got)
+	}
+	doc.SetFooter([]string{"end"})
+	if got := doc.Footer(); !slices.Equal(got, []string{"end"}) {
+		t.Errorf("footer = %q", got)
+	}
+}
+
+func TestDocumentArrayOfTables(t *testing.T) {
+	doc, err := Parse([]byte(`# first element
+[[item]]
+a = 1
+
+[[item]]
+b = 2 # beside b
+`))
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	entry, ok := doc.Root().Get("item")
+	if !ok {
+		t.Fatal("item is missing")
+	}
+	elems := entry.Elements()
+	if len(elems) != 2 {
+		t.Fatalf("elements = %d, want 2", len(elems))
+	}
+	if got := elems[0].Keys(); !slices.Equal(got, []string{"a"}) {
+		t.Errorf("first element keys = %v, want [a]", got)
+	}
+	if got := elems[0].Comments(); !slices.Equal(got, []string{"first element"}) {
+		t.Errorf("first element comments = %q", got)
+	}
+	if got := elems[1].Keys(); !slices.Equal(got, []string{"b"}) {
+		t.Errorf("second element keys = %v, want [b]", got)
+	}
+	if got := mustEntry(t, elems[1], "b").Trailing(); got != "beside b" {
+		t.Errorf("b trailing = %q, want \"beside b\"", got)
+	}
+	// The value keeps the map shape the decoder reads.
+	if _, ok := entry.Value().([]map[string]any); !ok {
+		t.Errorf("item value = %#v, want []map[string]any", entry.Value())
+	}
+}
+
+func TestDocumentDottedKeysAndValueArrays(t *testing.T) {
+	doc, err := Parse([]byte("a.b.c = 1\narr = [1, {x = 1}]\n"))
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+
+	// A dotted key builds tables, and they are not inline ones.
+	a, ok := doc.Root().Get("a")
+	if !ok {
+		t.Fatal("a is missing")
+	}
+	if a.Inline() {
+		t.Error("a is marked inline")
+	}
+	b, ok := a.Table().Get("b")
+	if !ok {
+		t.Fatal("a.b is missing")
+	}
+	if b.Inline() {
+		t.Error("a.b is marked inline")
+	}
+	if got := b.Table().Keys(); !slices.Equal(got, []string{"c"}) {
+		t.Errorf("a.b keys = %v, want [c]", got)
+	}
+
+	// An inline table inside a value array keeps its node in the elements
+	// slice; the scalar before it has none.
+	arr, ok := doc.Root().Get("arr")
+	if !ok {
+		t.Fatal("arr is missing")
+	}
+	elems := arr.Elements()
+	if len(elems) != 2 {
+		t.Fatalf("elements = %d, want 2", len(elems))
+	}
+	if elems[0] != nil {
+		t.Errorf("elements[0] = %#v, want nil for a scalar", elems[0])
+	}
+	if got := elems[1].Keys(); !slices.Equal(got, []string{"x"}) {
+		t.Errorf("elements[1] keys = %v, want [x]", got)
+	}
+}
+
+func TestDocumentWithoutStatements(t *testing.T) {
+	doc, err := Parse([]byte("# only a comment\n"))
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if got := doc.Root().Keys(); len(got) != 0 {
+		t.Errorf("keys = %v, want none", got)
+	}
+	if got := doc.Footer(); !slices.Equal(got, []string{"only a comment"}) {
+		t.Errorf("footer = %q, want [only a comment]", got)
+	}
+
+	empty, err := Parse(nil)
+	if err != nil {
+		t.Fatalf("parse of nothing: %v", err)
+	}
+	if len(empty.Root().Keys()) != 0 || len(empty.Footer()) != 0 {
+		t.Errorf("empty document = %v / %q", empty.Root().Keys(), empty.Footer())
+	}
+}
+
+func TestParseMapIsTheValueTree(t *testing.T) {
+	// ParseMap is the path that does not build a document, and it gives the
+	// tree the decoder reads.
+	tree, err := ParseMap([]byte("a = 1\n\n[t]\nb = \"x\"\n"))
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if tree["a"] != int64(1) {
+		t.Errorf("a = %#v", tree["a"])
+	}
+	if tree["t"].(map[string]any)["b"] != "x" {
+		t.Errorf("t = %#v", tree["t"])
+	}
+}
+
+func TestMarshalRejectsDocument(t *testing.T) {
+	// A Document is not a value to marshal: its order and comments would be
+	// dropped, and a struct walk would silently write nothing at all.
+	doc, err := Parse([]byte("a = 1\n"))
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if _, err := Marshal(doc); err == nil {
+		t.Fatal("expected an error for a Document")
+	} else if !strings.Contains(err.Error(), "Map()") {
+		t.Errorf("err = %v, want it to point at Map()", err)
+	}
+	if _, err := Marshal(*doc); err == nil {
+		t.Fatal("expected an error for a Document value")
+	}
+	// The tree marshals, which is the way through.
+	out, err := Marshal(doc.Map())
+	if err != nil {
+		t.Fatalf("marshal of the tree: %v", err)
+	}
+	if want := "a = 1\n"; string(out) != want {
+		t.Errorf("output = %q, want %q", out, want)
+	}
+}

encode_test.go

@@ -8,6 +8,7 @@ import (
 	"context"
 	"errors"
 	"math"
+	"net"
 	"reflect"
 	"strings"
 	"testing"
@@ -288,7 +289,7 @@ func TestEncoderLiteralMultilineFallsBackWhenUnsafe(t *testing.T) {
 		if !bytes.HasPrefix(out, []byte("s = \"")) {
 			t.Errorf("%s: expected the basic quoted form, got:\n%s", c.name, out)
 		}
-		re, err := Parse(out)
+		re, err := ParseMap(out)
 		if err != nil {
 			t.Errorf("%s: re-parse: %v\ndoc:\n%s", c.name, err, out)
 			continue
@@ -325,7 +326,7 @@ func TestMarshalerReturningTime(t *testing.T) {
 	if err != nil {
 		t.Fatalf("marshal: %v", err)
 	}
-	want := "m = 2026-06-26T10:00:00Z\n"
+	want := "m = 2026-06-26T10:00Z\n"
 	if string(out) != want {
 		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
 	}
@@ -447,7 +448,7 @@ func TestMarshalDuplicateKeyResolvesToOneField(t *testing.T) {
 	if want := "name = \"outer\"\n"; string(out) != want {
 		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
 	}
-	if _, err := Parse(out); err != nil {
+	if _, err := ParseMap(out); err != nil {
 		t.Errorf("re-parse: %v\ndoc:\n%s", err, out)
 	}
 }
@@ -465,7 +466,7 @@ func TestMarshalEmbeddedScalarStruct(t *testing.T) {
 	if err != nil {
 		t.Fatalf("marshal: %v", err)
 	}
-	want := "name = \"x\"\ns = 2026-06-26T00:00:00\n"
+	want := "name = \"x\"\ns = 2026-06-26T00:00\n"
 	if string(out) != want {
 		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
 	}
@@ -540,7 +541,7 @@ func TestMarshalDateTime(t *testing.T) {
 	if err != nil {
 		t.Fatalf("marshal: %v", err)
 	}
-	want := "offset = 2026-06-26T10:00:00Z\nlocal = 2026-06-26T07:32:00\nday = 2026-06-26\nclock = 07:32:00\n"
+	want := "offset = 2026-06-26T10:00Z\nlocal = 2026-06-26T07:32\nday = 2026-06-26\nclock = 07:32\n"
 	if string(out) != want {
 		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
 	}
@@ -609,7 +610,7 @@ func TestMarshalMixedArrayWithInlineTable(t *testing.T) {
 	// Parse accepts a mixed array (TOML allows any value kinds in one array),
 	// so Marshal of the parsed tree must re-emit it. The table element has no
 	// header form inside a value array and renders inline.
-	tree, err := Parse([]byte("arr = [1, {a = 2}, \"x\"]\n"))
+	tree, err := ParseMap([]byte("arr = [1, {a = 2}, \"x\"]\n"))
 	if err != nil {
 		t.Fatalf("parse: %v", err)
 	}
@@ -621,7 +622,7 @@ func TestMarshalMixedArrayWithInlineTable(t *testing.T) {
 	if string(out) != want {
 		t.Fatalf("output mismatch:\ngot:  %q\nwant: %q", out, want)
 	}
-	re, err := Parse(out)
+	re, err := ParseMap(out)
 	if err != nil {
 		t.Fatalf("re-parse: %v", err)
 	}
@@ -639,7 +640,7 @@ func TestMarshalValueArrayOfTablesStaysInline(t *testing.T) {
 		"a = [{x = 1}, {x = 2}]\n",
 		"b = [{x = 1}, 2, \"three\"]\n",
 	} {
-		tree, err := Parse([]byte(doc))
+		tree, err := ParseMap([]byte(doc))
 		if err != nil {
 			t.Fatalf("%s: parse: %v", doc, err)
 		}
@@ -650,7 +651,7 @@ func TestMarshalValueArrayOfTablesStaysInline(t *testing.T) {
 		if bytes.HasPrefix(out, []byte("[[")) {
 			t.Errorf("%s: emitted the [[header]] form for a value array:\n%s", doc, out)
 		}
-		re, err := Parse(out)
+		re, err := ParseMap(out)
 		if err != nil {
 			t.Fatalf("%s: re-parse: %v\ndoc:\n%s", doc, err, out)
 		}
@@ -687,14 +688,14 @@ func TestMarshalInlineTableWithDatetime(t *testing.T) {
 	if err != nil {
 		t.Fatalf("marshal: %v", err)
 	}
-	want := "mix = [1979-05-27T07:32:00Z, {t = 1979-05-27T07:32:00}]\n"
+	want := "mix = [1979-05-27T07:32Z, {t = 1979-05-27T07:32}]\n"
 	if string(out) != want {
 		t.Fatalf("output mismatch:\ngot:  %q\nwant: %q", out, want)
 	}
 }
 
 func TestMarshalArrayOfTablesStaysHeaderForm(t *testing.T) {
-	tree, err := Parse([]byte("[[items]]\nname = \"a\"\n\n[[items]]\nname = \"b\"\n"))
+	tree, err := ParseMap([]byte("[[items]]\nname = \"a\"\n\n[[items]]\nname = \"b\"\n"))
 	if err != nil {
 		t.Fatalf("parse: %v", err)
 	}
@@ -721,7 +722,7 @@ func TestMarshalFloatExponentNoLeadingZero(t *testing.T) {
 	}
 	// Parse to check the output is valid TOML (for a strict parser that
 	// rejects leading zeros in exponents).
-	if _, err := Parse(out); err != nil {
+	if _, err := ParseMap(out); err != nil {
 		t.Fatalf("marshalled output is not valid TOML:\n%s\nerror: %v", out, err)
 	}
 	if string(out) != "large = 1e+6\nsmall = 1e-5\n" {
@@ -906,7 +907,7 @@ func TestMarshalTagOptionOmitZero(t *testing.T) {
 	if err != nil {
 		t.Fatalf("marshal: %v", err)
 	}
-	want = "name = \"x\"\ncount = 1\nratio = 0.5\nwhen = 2026-09-17T12:00:00Z\nalways = \"kept\"\n\n[server]\nhost = \"h\"\n"
+	want = "name = \"x\"\ncount = 1\nratio = 0.5\nwhen = 2026-09-17T12:00Z\nalways = \"kept\"\n\n[server]\nhost = \"h\"\n"
 	if string(out) != want {
 		t.Fatalf("output mismatch:\ngot:  %q\nwant: %q", out, want)
 	}
@@ -1160,7 +1161,7 @@ func TestMarshalThenParseRoundTrip(t *testing.T) {
 	if err != nil {
 		t.Fatalf("marshal: %v", err)
 	}
-	tree1, err := Parse(out)
+	tree1, err := ParseMap(out)
 	if err != nil {
 		t.Fatalf("parse of marshalled: %v\noutput:\n%s", err, out)
 	}
@@ -1195,8 +1196,10 @@ qty = 2
 
 [meta]
 created = 2026-06-26T10:00:00Z
+
+mixed = [1, {n = 1, name = "a value long enough to push this line well past the one hundred column limit"}]
 `)
-	tree1, err := Parse(src)
+	tree1, err := ParseMap(src)
 	if err != nil {
 		t.Fatalf("parse src: %v", err)
 	}
@@ -1204,7 +1207,7 @@ created = 2026-06-26T10:00:00Z
 	if err != nil {
 		t.Fatalf("marshal: %v", err)
 	}
-	tree2, err := Parse(out)
+	tree2, err := ParseMap(out)
 	if err != nil {
 		t.Fatalf("re-parse marshalled: %v\noutput:\n%s", err, out)
 	}
@@ -1257,24 +1260,35 @@ func TestLocalDateString(t *testing.T) {
 }
 
 func TestLocalDateTimeString(t *testing.T) {
+	// The rendering drops zero seconds and the trailing zeros of a fraction,
+	// which TOML 1.1 allows and which keeps a value written without seconds
+	// written without them.
 	ldt := LocalDateTime{Time: time.Date(1979, 5, 27, 7, 32, 0, 0, time.UTC)}
-	if got := ldt.String(); got != "1979-05-27T07:32:00" {
+	if got := ldt.String(); got != "1979-05-27T07:32" {
-		t.Errorf("LocalDateTime.String() = %q, want 1979-05-27T07:32:00", got)
+		t.Errorf("LocalDateTime.String() = %q, want 1979-05-27T07:32", got)
 	}
 	ldt2 := LocalDateTime{Time: time.Date(1979, 5, 27, 7, 32, 0, 5, time.UTC)}
 	if got := ldt2.String(); got != "1979-05-27T07:32:00.000000005" {
 		t.Errorf("LocalDateTime.String() = %q, want 1979-05-27T07:32:00.000000005", got)
 	}
 	ldt3 := LocalDateTime{Time: time.Date(1979, 5, 27, 7, 32, 0, 500, time.UTC)}
-	if got := ldt3.String(); got != "1979-05-27T07:32:00.000000500" {
+	if got := ldt3.String(); got != "1979-05-27T07:32:00.0000005" {
-		t.Errorf("LocalDateTime.String() = %q, want 1979-05-27T07:32:00.000000500", got)
+		t.Errorf("LocalDateTime.String() = %q, want 1979-05-27T07:32:00.0000005", got)
+	}
+	ldt4 := LocalDateTime{Time: time.Date(1979, 5, 27, 7, 32, 30, 500000000, time.UTC)}
+	if got := ldt4.String(); got != "1979-05-27T07:32:30.5" {
+		t.Errorf("LocalDateTime.String() = %q, want 1979-05-27T07:32:30.5", got)
 	}
 }
 
 func TestLocalTimeString(t *testing.T) {
 	lt := LocalTime{Time: time.Date(0, 1, 1, 7, 32, 0, 0, time.UTC)}
-	if got := lt.String(); got != "07:32:00" {
+	if got := lt.String(); got != "07:32" {
-		t.Errorf("LocalTime.String() = %q, want 07:32:00", got)
+		t.Errorf("LocalTime.String() = %q, want 07:32", got)
+	}
+	lt2 := LocalTime{Time: time.Date(0, 1, 1, 7, 32, 15, 250000000, time.UTC)}
+	if got := lt2.String(); got != "07:32:15.25" {
+		t.Errorf("LocalTime.String() = %q, want 07:32:15.25", got)
 	}
 }
 
@@ -1372,3 +1386,563 @@ func TestEncodeErrorHeterogeneousArrayPath(t *testing.T) {
 		t.Fatalf("Path = %q, want %q", ee.Path, "items[0]")
 	}
 }
+
+// --- encoding.TextMarshaler and time.Duration ------------------------------
+
+// textTag is a value-receiver encoding.TextMarshaler, so the encoder finds the
+// method on the value itself.
+type textTag string
+
+func (t textTag) MarshalText() ([]byte, error) { return []byte("tag:" + string(t)), nil }
+
+// textPointer carries MarshalText on the pointer receiver only, so the encoder
+// has to look at the address of an addressable field.
+type textPointer struct{ V string }
+
+func (t *textPointer) MarshalText() ([]byte, error) { return []byte(strings.ToUpper(t.V)), nil }
+
+// textAndTOML implements both encoding interfaces; the TOML method wins.
+type textAndTOML struct{}
+
+func (textAndTOML) MarshalTOML() (any, error)    { return "toml", nil }
+func (textAndTOML) MarshalText() ([]byte, error) { return []byte("text"), nil }
+
+// brokenText fails the marshal from MarshalText.
+type brokenText struct{}
+
+func (brokenText) MarshalText() ([]byte, error) { return nil, errors.New("text boom") }
+
+// notUTF8 renders bytes that no TOML string can carry.
+type notUTF8 struct{}
+
+func (notUTF8) MarshalText() ([]byte, error) { return []byte{0xff, 0xfe}, nil }
+
+// textTagBoth renders itself with a prefix and strips it again on decode, so
+// the round trip through a TOML string is lossless.
+type textTagBoth string
+
+func (t textTagBoth) MarshalText() ([]byte, error) { return []byte("tag:" + string(t)), nil }
+
+func (t *textTagBoth) UnmarshalText(text []byte) error {
+	trimmed, ok := strings.CutPrefix(string(text), "tag:")
+	if !ok {
+		return errors.New("textTagBoth: missing the tag prefix")
+	}
+	*t = textTagBoth(trimmed)
+	return nil
+}
+
+func TestMarshalTextValues(t *testing.T) {
+	// The pointer receiver is reachable only through an addressable field, so
+	// the whole value is marshalled through a pointer here.
+	type Cfg struct {
+		IP       net.IP        `toml:"ip"`
+		Duration time.Duration `toml:"duration"`
+		Tag      textTag       `toml:"tag"`
+		Pointer  textPointer   `toml:"pointer"`
+		Both     textAndTOML   `toml:"both"`
+	}
+	out, err := Marshal(&Cfg{
+		IP:       net.IPv4(192, 0, 2, 1),
+		Duration: 90 * time.Minute,
+		Tag:      "x",
+		Pointer:  textPointer{V: "abc"},
+	})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	want := "ip = \"192.0.2.1\"\nduration = \"1h30m0s\"\ntag = \"tag:x\"\npointer = \"ABC\"\nboth = \"toml\"\n"
+	if string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+}
+
+func TestMarshalTextValuesInContainers(t *testing.T) {
+	// Slice elements are addressable, so a pointer-receiver MarshalText is used
+	// there too, and an array of such values stays a value array: each element's
+	// TOML form is a string, so the [[header]] form cannot carry it.
+	type Cfg struct {
+		Map   map[string]net.IP `toml:"map"`
+		Durs  []time.Duration   `toml:"durs"`
+		Ptrs  []textPointer     `toml:"ptrs"`
+		Empty []textPointer     `toml:"empty"`
+	}
+	out, err := Marshal(Cfg{
+		Map:   map[string]net.IP{"a": net.IPv4(10, 0, 0, 1)},
+		Durs:  []time.Duration{0, 250 * time.Millisecond},
+		Ptrs:  []textPointer{{V: "a"}, {V: "b"}},
+		Empty: []textPointer{},
+	})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	want := "durs = [\"0s\", \"250ms\"]\nptrs = [\"A\", \"B\"]\nempty = []\n\n[map]\na = \"10.0.0.1\"\n"
+	if string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+}
+
+func TestMarshalTextLeavesDateTimesAlone(t *testing.T) {
+	// The four date-time types carry time.Time's text methods through an
+	// embedded field; their TOML form is a bare atom, never a quoted string.
+	stamp := time.Date(2026, 6, 26, 10, 0, 0, 0, time.UTC)
+	type Cfg struct {
+		Stamp time.Time     `toml:"stamp"`
+		Ptr   *time.Time    `toml:"ptr"`
+		Day   LocalDate     `toml:"day"`
+		At    LocalDateTime `toml:"at"`
+		Clock LocalTime     `toml:"clock"`
+	}
+	out, err := Marshal(&Cfg{
+		Stamp: stamp,
+		Ptr:   &stamp,
+		Day:   LocalDate{time.Date(1979, 5, 27, 0, 0, 0, 0, time.UTC)},
+		At:    LocalDateTime{time.Date(1979, 5, 27, 7, 32, 0, 0, time.UTC)},
+		Clock: LocalTime{time.Date(0, 1, 1, 7, 32, 0, 0, time.UTC)},
+	})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	want := "stamp = 2026-06-26T10:00Z\nptr = 2026-06-26T10:00Z\nday = 1979-05-27\nat = 1979-05-27T07:32\nclock = 07:32\n"
+	if string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+}
+
+func TestMarshalTextNilPointerOmitted(t *testing.T) {
+	type Cfg struct {
+		P *textPointer `toml:"p"`
+		K string       `toml:"k"`
+	}
+	out, err := Marshal(&Cfg{K: "x"})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	if want := "k = \"x\"\n"; string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+}
+
+func TestMarshalTextErrorCarriesPath(t *testing.T) {
+	type Inner struct {
+		F brokenText `toml:"f"`
+	}
+	type Cfg struct {
+		Inner Inner `toml:"inner"`
+	}
+	_, err := Marshal(Cfg{})
+	if err == nil {
+		t.Fatal("expected an error from MarshalText")
+	}
+	if !strings.Contains(err.Error(), "text boom") {
+		t.Errorf("err = %v, want substring \"text boom\"", err)
+	}
+	ee, ok := errors.AsType[*EncodeError](err)
+	if !ok {
+		t.Fatalf("expected an *EncodeError, got %T: %v", err, err)
+	}
+	if ee.Path != "inner.f" {
+		t.Fatalf("Path = %q, want %q", ee.Path, "inner.f")
+	}
+}
+
+func TestMarshalTextRejectsInvalidUTF8(t *testing.T) {
+	// A TOML string holds UTF-8 only, so text that is not gets an error rather
+	// than replacement characters.
+	_, err := Marshal(struct {
+		V notUTF8 `toml:"v"`
+	}{})
+	if err == nil {
+		t.Fatal("expected an error for text that is not valid UTF-8")
+	}
+	if !strings.Contains(err.Error(), "UTF-8") {
+		t.Errorf("err = %v, want a UTF-8 message", err)
+	}
+}
+
+func TestMarshalTextValuesRoundTrip(t *testing.T) {
+	type Cfg struct {
+		Duration time.Duration `toml:"duration"`
+		IP       net.IP        `toml:"ip"`
+		Tag      textTagBoth   `toml:"tag"`
+	}
+	in := Cfg{Duration: 90 * time.Minute, IP: net.IPv4(198, 51, 100, 7), Tag: "y"}
+	out, err := Marshal(&in)
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	var back Cfg
+	if err := Unmarshal(out, &back); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if back.Duration != in.Duration {
+		t.Errorf("Duration = %v, want %v", back.Duration, in.Duration)
+	}
+	if !back.IP.Equal(in.IP) {
+		t.Errorf("IP = %v, want %v", back.IP, in.IP)
+	}
+	if back.Tag != in.Tag {
+		t.Errorf("Tag = %q, want %q", back.Tag, in.Tag)
+	}
+}
+
+// --- TOML 1.1 output forms -------------------------------------------------
+
+func TestMarshalDateTimeRendering(t *testing.T) {
+	// The seconds are written only when the value carries them, and a fraction
+	// drops its trailing zeros. Both are the same value either way; the shorter
+	// form is the one TOML 1.1 allows.
+	base := time.Date(2026, 6, 26, 10, 0, 0, 0, time.UTC)
+	cases := []struct {
+		name string
+		val  any
+		want string
+	}{
+		{"offset-zero-seconds", base, "v = 2026-06-26T10:00Z\n"},
+		{"offset-wrapper", OffsetDateTime{Time: base}, "v = 2026-06-26T10:00Z\n"},
+		{"offset-seconds", base.Add(30 * time.Second), "v = 2026-06-26T10:00:30Z\n"},
+		{"offset-fraction", base.Add(500 * time.Millisecond), "v = 2026-06-26T10:00:00.5Z\n"},
+		{"offset-zone", time.Date(2026, 6, 26, 10, 0, 0, 0, time.FixedZone("", -7*3600)), "v = 2026-06-26T10:00-07:00\n"},
+		{"local-zero-seconds", LocalDateTime{Time: base}, "v = 2026-06-26T10:00\n"},
+		{"local-fraction", LocalDateTime{Time: base.Add(2500 * time.Millisecond)}, "v = 2026-06-26T10:00:02.5\n"},
+		{"date", LocalDate{Time: base}, "v = 2026-06-26\n"},
+		{"time-zero-seconds", LocalTime{Time: base}, "v = 10:00\n"},
+		{"time-seconds", LocalTime{Time: base.Add(15 * time.Second)}, "v = 10:00:15\n"},
+		{"time-nanoseconds", LocalTime{Time: base.Add(123456789 * time.Nanosecond)}, "v = 10:00:00.123456789\n"},
+	}
+	for _, c := range cases {
+		out, err := Marshal(map[string]any{"v": c.val})
+		if err != nil {
+			t.Errorf("%s: marshal: %v", c.name, err)
+			continue
+		}
+		if string(out) != c.want {
+			t.Errorf("%s: output mismatch:\ngot:  %q\nwant: %q", c.name, out, c.want)
+		}
+	}
+}
+
+func TestMarshalInlineTableBreaksWhenLong(t *testing.T) {
+	// A table element of a value array is written inline; a long one carries
+	// newlines and a trailing comma instead of running past the line limit,
+	// which TOML 1.1 allows an inline table to do.
+	const long = "a-very-long-value-that-pushes-the-line-well-past-the-one-hundred-column-limit"
+	type Cfg struct {
+		Arr []any `toml:"arr"`
+	}
+	out, err := Marshal(Cfg{Arr: []any{int64(1), map[string]any{"n": int64(1), "name": long}}})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	want := "arr = [1, {\n\tn = 1,\n\tname = \"" + long + "\",\n}]\n"
+	if string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+
+	// The same values without the long string stay on one line.
+	out, err = Marshal(Cfg{Arr: []any{int64(1), map[string]any{"n": int64(1), "name": "short"}}})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	if want := "arr = [1, {n = 1, name = \"short\"}]\n"; string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+
+	// The broken form parses back to the same tree.
+	tree, err := ParseMap(out)
+	if err != nil {
+		t.Fatalf("parse of the encoder output: %v", err)
+	}
+	if got := len(tree["arr"].([]any)); got != 2 {
+		t.Fatalf("arr has %d elements, want 2", got)
+	}
+}
+
+func TestMarshalNestedInlineTableBreaksIndependently(t *testing.T) {
+	// A nested table breaks on its own measure, so a table whose entries stay
+	// short keeps the one-line form inside a parent that broke.
+	const long = "a-very-long-value-that-pushes-the-line-well-past-the-one-hundred-column-limit"
+	type Cfg struct {
+		Arr []any `toml:"arr"`
+	}
+	out, err := Marshal(Cfg{Arr: []any{int64(1), map[string]any{"n": int64(1), "sub": map[string]any{"name": long}}}})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	want := "arr = [1, {\n\tn = 1,\n\tsub = {name = \"" + long + "\"},\n}]\n"
+	if string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+}
+
+type inlineTLS struct {
+	On bool `toml:"on"`
+}
+
+type inlineServer struct {
+	Host string    `toml:"host"`
+	Port int       `toml:"port"`
+	TLS  inlineTLS `toml:"tls"`
+}
+
+type inlineBig struct {
+	A int `toml:"a"`
+	B int `toml:"b"`
+	C int `toml:"c"`
+}
+
+func TestEncoderInlineTables(t *testing.T) {
+	type Cfg struct {
+		Server inlineServer `toml:"server"`
+		Big    inlineBig    `toml:"big"`
+	}
+	cfg := Cfg{Server: inlineServer{Host: "127.0.0.1", Port: 9090}, Big: inlineBig{A: 1, B: 2, C: 3}}
+
+	// The default keeps every sub-table a header section.
+	headerForm, err := Marshal(cfg)
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	want := "[server]\nhost = \"127.0.0.1\"\nport = 9090\n\n[server.tls]\non = false\n\n[big]\na = 1\nb = 2\nc = 3\n"
+	if string(headerForm) != want {
+		t.Errorf("default output mismatch:\ngot:  %q\nwant: %q", headerForm, want)
+	}
+
+	// With the option both fit the threshold and become inline tables, nested
+	// ones included.
+	out, err := NewEncoder().InlineTables(60).Marshal(cfg)
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	want = "server = {host = \"127.0.0.1\", port = 9090, tls = {on = false}}\nbig = {a = 1, b = 2, c = 3}\n"
+	if string(out) != want {
+		t.Errorf("compact output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+
+	// A threshold below the rendering keeps the header form.
+	out, err = NewEncoder().InlineTables(10).Marshal(cfg)
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	if string(out) != string(headerForm) {
+		t.Errorf("small threshold output mismatch:\ngot:  %q\nwant: %q", out, headerForm)
+	}
+}
+
+func TestEncoderInlineTablesOrderAndRoundTrip(t *testing.T) {
+	// An inlined sub-table is a value line, so it precedes every header of the
+	// document; written after a header it would be read back as part of that
+	// table. The compact form and the header form parse to the same tree.
+	type Four struct {
+		A int `toml:"a"`
+		B int `toml:"b"`
+		C int `toml:"c"`
+		D int `toml:"d"`
+	}
+	type Cfg struct {
+		Small inlineTLS `toml:"small"`
+		Big   Four      `toml:"big"`
+	}
+	cfg := Cfg{Small: inlineTLS{On: true}, Big: Four{A: 1, B: 2, C: 3, D: 4}}
+
+	headerForm, err := Marshal(cfg)
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	compact, err := NewEncoder().InlineTables(20).Marshal(cfg)
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	want := "small = {on = true}\n\n[big]\na = 1\nb = 2\nc = 3\nd = 4\n"
+	if string(compact) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", compact, want)
+	}
+
+	got, err := ParseMap(compact)
+	if err != nil {
+		t.Fatalf("parse of the compact output: %v", err)
+	}
+	ref, err := ParseMap(headerForm)
+	if err != nil {
+		t.Fatalf("parse of the header output: %v", err)
+	}
+	if !reflect.DeepEqual(got, ref) {
+		t.Errorf("the compact form changed the tree:\ncompact: %#v\nheaders: %#v", got, ref)
+	}
+	if _, ok := got["big"].(map[string]any); !ok {
+		t.Errorf("big = %#v, want a table", got["big"])
+	}
+}
+
+func TestEncoderInlineTablesKeepsArraysOfTables(t *testing.T) {
+	// An array of tables has no inline form that keeps the value's type, so the
+	// option leaves it alone and the tree keeps its []map[string]any shape.
+	type Item struct {
+		N int `toml:"n"`
+	}
+	type Cfg struct {
+		Items []Item    `toml:"items"`
+		Small inlineTLS `toml:"small"`
+	}
+	cfg := Cfg{Items: []Item{{N: 1}}, Small: inlineTLS{On: true}}
+	out, err := NewEncoder().InlineTables(60).Marshal(cfg)
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	want := "small = {on = true}\n\n[[items]]\nn = 1\n"
+	if string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+	tree, err := ParseMap(out)
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if _, ok := tree["items"].([]map[string]any); !ok {
+		t.Errorf("items = %#v, want []map[string]any", tree["items"])
+	}
+}
+
+// --- Marshaler on array elements -------------------------------------------
+
+// countingMarshaler reports how often its method ran, so a test can check that
+// the encoder resolves an element once.
+type countingMarshaler struct{ calls *int }
+
+func (c countingMarshaler) MarshalTOML() (any, error) {
+	*c.calls++
+	return map[string]any{"n": int64(*c.calls)}, nil
+}
+
+// ptrMarshaler carries MarshalTOML on the pointer receiver only.
+type ptrMarshaler struct{ V string }
+
+func (p *ptrMarshaler) MarshalTOML() (any, error) { return map[string]any{"v": p.V}, nil }
+
+func TestMarshalerElementOfArrayOfTables(t *testing.T) {
+	// An element is classified by what MarshalTOML returns, so methods that
+	// render tables keep the [[header]] form the Go kind would have given them.
+	type Cfg struct {
+		Items []marshalerFunc `toml:"items"`
+	}
+	out, err := Marshal(Cfg{Items: []marshalerFunc{
+		func() (any, error) { return map[string]any{"k": "a"}, nil },
+		func() (any, error) { return map[string]any{"k": "b"}, nil },
+	}})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	want := "[[items]]\nk = \"a\"\n\n[[items]]\nk = \"b\"\n"
+	if string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+}
+
+func TestMarshalerElementScalarResultMakesValueArray(t *testing.T) {
+	// One element rendering itself as a scalar turns the whole array into a
+	// value array, with the table elements written inline.
+	type Cfg struct {
+		Items []marshalerFunc `toml:"items"`
+	}
+	out, err := Marshal(Cfg{Items: []marshalerFunc{
+		func() (any, error) { return map[string]any{"k": "a"}, nil },
+		func() (any, error) { return "x", nil },
+	}})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	want := "items = [{k = \"a\"}, \"x\"]\n"
+	if string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+}
+
+func TestMarshalerElementRunsOnce(t *testing.T) {
+	// Classification and emission share one result, so the method runs exactly
+	// once per element even when it decides the array's form.
+	calls := 0
+	type Cfg struct {
+		Items []countingMarshaler `toml:"items"`
+	}
+	out, err := Marshal(Cfg{Items: []countingMarshaler{{calls: &calls}, {calls: &calls}, {calls: &calls}}})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	if calls != 3 {
+		t.Errorf("MarshalTOML ran %d times, want 3", calls)
+	}
+	want := "[[items]]\nn = 1\n\n[[items]]\nn = 2\n\n[[items]]\nn = 3\n"
+	if string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+}
+
+func TestMarshalerPointerReceiverOnElement(t *testing.T) {
+	// A slice element is addressable, so a pointer-receiver MarshalTOML is
+	// found there, and the method's table keeps the [[header]] form.
+	type Cfg struct {
+		Items []ptrMarshaler `toml:"items"`
+	}
+	out, err := Marshal(&Cfg{Items: []ptrMarshaler{{V: "a"}, {V: "b"}}})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	want := "[[items]]\nv = \"a\"\n\n[[items]]\nv = \"b\"\n"
+	if string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+
+	// The same method is found on an addressable struct field, whose result is
+	// a table and so keeps a header section.
+	type Field struct {
+		F ptrMarshaler `toml:"f"`
+	}
+	out, err = Marshal(&Field{F: ptrMarshaler{V: "c"}})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	if want := "[f]\nv = \"c\"\n"; string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+}
+
+func TestMarshalerElementErrorCarriesPath(t *testing.T) {
+	type Cfg struct {
+		Items []any `toml:"items"`
+	}
+	_, err := Marshal(Cfg{Items: []any{map[string]any{"k": "a"}, failingMarshalerFunc{}}})
+	if err == nil {
+		t.Fatal("expected an error from MarshalTOML")
+	}
+	if !strings.Contains(err.Error(), "oops") {
+		t.Errorf("err = %v, want substring \"oops\"", err)
+	}
+	ee, ok := errors.AsType[*EncodeError](err)
+	if !ok {
+		t.Fatalf("expected an *EncodeError, got %T: %v", err, err)
+	}
+	if ee.Path != "items[1]" {
+		t.Fatalf("Path = %q, want %q", ee.Path, "items[1]")
+	}
+}
+
+func TestMarshalerResultIsNormalised(t *testing.T) {
+	// A result is normalised like any other value, so a method may return a
+	// plain int or a duration where the Go kind alone would not encode.
+	type Cfg struct {
+		Plain    marshalerFunc `toml:"plain"`
+		Duration marshalerFunc `toml:"duration"`
+		Elements []any         `toml:"elements"`
+	}
+	out, err := Marshal(Cfg{
+		Plain:    func() (any, error) { return 7, nil },
+		Duration: func() (any, error) { return 90 * time.Minute, nil },
+		Elements: []any{marshalerFunc(func() (any, error) { return 8, nil })},
+	})
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	want := "plain = 7\nduration = \"1h30m0s\"\nelements = [8]\n"
+	if string(out) != want {
+		t.Errorf("output mismatch:\ngot:  %q\nwant: %q", out, want)
+	}
+}

examples/basic/main.go

@@ -13,7 +13,7 @@ import (
 	"os"
 	"time"
 
-	"sourcedock.dev/petrbalvin/interpres"
+	"sourcedock.dev/petrbalvin/interpres/v2"
 )
 
 // document is a small but realistic configuration: it has scalars, a

fuzz_test.go

@@ -41,7 +41,7 @@ func FuzzParse(f *testing.F) {
 		f.Add([]byte(s))
 	}
 	f.Fuzz(func(t *testing.T, data []byte) {
-		tree, err := Parse(data)
+		tree, err := ParseMap(data)
 		if err != nil {
 			return
 		}
@@ -49,7 +49,7 @@ func FuzzParse(f *testing.F) {
 		if err != nil {
 			t.Fatalf("marshal of a parsed tree failed: %v\ntree: %#v", err, tree)
 		}
-		re, err := Parse(out)
+		re, err := ParseMap(out)
 		if err != nil {
 			t.Fatalf("re-parse of the emitted document failed: %v\ndoc:\n%s", err, out)
 		}

go.mod

@@ -1,3 +1,3 @@
-module sourcedock.dev/petrbalvin/interpres
+module sourcedock.dev/petrbalvin/interpres/v2
 
 go 1.27.1

interpres.go

@@ -11,9 +11,10 @@
 //
 //	out, err := interpres.Marshal(cfg)
 //
-// or, for an untyped tree:
+// or, for the document with its key order and comments:
 //
-//	tree, err := interpres.Parse(data)
+//	doc, err := interpres.Parse(data)
+//	tree := doc.Map()
 //
 // A Decoder allows strict decoding that rejects keys without a matching
 // struct field, mirroring (*json.Decoder).DisallowUnknownFields.
@@ -23,7 +24,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
-	"unicode/utf8"
+	"slices"
 )
 
 // A SyntaxError describes a malformed TOML document, including the 1-based
@@ -89,31 +90,77 @@ func (e *EncodeError) Error() string { return "interpres: " + e.Path + ": " + e.
 // Unwrap returns the failure the path points at.
 func (e *EncodeError) Unwrap() error { return e.Err }
 
-// Parse decodes a TOML document into a nested map[string]any.
+// Parse decodes a TOML document into a Document: the values, the order the
+// keys were written in, whether a table was written inline, and the comments.
+// ParseMap gives the plain value tree instead.
 //
 // Values are mapped to Go types as follows: strings to string, integers to
-// int64, floats to float64, booleans to bool, date-times to time.Time, arrays
+// int64, floats to float64, booleans to bool, offset date-times to
-// to []any, and tables (including inline tables) to map[string]any.
+// OffsetDateTime, the local date-time kinds to their wrappers, arrays to
+// []any, and tables (including inline tables) to map[string]any.
 //
 // Parse is equivalent to ParseContext with context.Background.
-func Parse(data []byte) (map[string]any, error) {
+func Parse(data []byte) (*Document, error) {
 	return ParseContext(context.Background(), data)
 }
 
-// ParseContext decodes a TOML document into a nested map[string]any, obeying
+// ParseContext decodes a TOML document into a Document, obeying ctx. The
-// ctx. The context is checked between top-level statements so cancellation is
+// context is checked between top-level statements so cancellation is honoured
-// honoured before the parser has done substantial work.
+// before the parser has done substantial work.
-func ParseContext(ctx context.Context, data []byte) (map[string]any, error) {
+func ParseContext(ctx context.Context, data []byte) (*Document, error) {
-	if err := ctx.Err(); err != nil {
+	_, doc, err := parseWithOptions(ctx, data, parseOptions{}, true)
-		return nil, err
+	return doc, err
 }
-	if !utf8.Valid(data) {
+
-		return nil, &SyntaxError{Line: 1, Msg: "input is not valid UTF-8"}
+// ParseMap decodes a TOML document into a nested map[string]any, the value
+// tree without the order and the comments a Document carries. It is the shape
+// this package parsed into before [Document] existed.
+//
+// ParseMap is equivalent to ParseMapContext with context.Background.
+func ParseMap(data []byte) (map[string]any, error) {
+	return ParseMapContext(context.Background(), data)
+}
+
+// ParseMapContext is the cancellable variant of ParseMap.
+func ParseMapContext(ctx context.Context, data []byte) (map[string]any, error) {
+	tree, _, err := parseWithOptions(ctx, data, parseOptions{}, false)
+	return tree, err
+}
+
+// parseOptions bound the work one parse may do. A zero field takes the
+// default.
+type parseOptions struct {
+	maxDepth     int
+	maxInputSize int
+}
+
+// parseWithOptions parses data, building the node tree of a Document when
+// wantDoc asks for it, and returns both the value tree and that document.
+func parseWithOptions(ctx context.Context, data []byte, opts parseOptions, wantDoc bool) (map[string]any, *Document, error) {
+	if err := ctx.Err(); err != nil {
+		return nil, nil, err
+	}
+	if opts.maxInputSize > 0 && len(data) > opts.maxInputSize {
+		return nil, nil, fmt.Errorf("interpres: input is %d bytes, over the limit of %d", len(data), opts.maxInputSize)
+	}
+	// UTF-8 validity is not checked in a pass of its own: the scanner
+	// validates the multi-byte sequences where it meets them, so an invalid
+	// byte is reported on its own line instead of always on line 1.
+	maxDepth := opts.maxDepth
+	if maxDepth <= 0 {
+		maxDepth = maxNestingDepth
 	}
 	// The parser scans data in place; it only reads the buffer, and every
 	// string it stores in the tree is copied out of it.
-	p := &parser{src: data, line: 1, ctx: ctx}
+	p := &parser{src: data, line: 1, ctx: ctx, maxDepth: maxDepth, wantDoc: wantDoc}
-	return p.parse()
+	tree, err := p.parse()
+	if err != nil {
+		return nil, nil, err
+	}
+	if !wantDoc {
+		return tree, nil, nil
+	}
+	return tree, &Document{root: p.doc, footer: p.footer}, nil
 }
 
 // Unmarshal parses a TOML document and stores the result in the value pointed
@@ -123,6 +170,11 @@ func ParseContext(ctx context.Context, data []byte) (map[string]any, error) {
 // case-insensitive match on the field name when no tag is present. A tag of
 // "-" skips the field.
 //
+// A destination implementing Unmarshaler receives the parsed value as it is,
+// a TOML string fills a destination implementing encoding.TextUnmarshaler, and
+// a time.Duration destination takes a duration literal such as `1h30m` or a
+// bare integer as its nanosecond count.
+//
 // Unmarshal is equivalent to UnmarshalContext with context.Background.
 func Unmarshal(data []byte, v any) error {
 	return UnmarshalContext(context.Background(), data, v)
@@ -130,7 +182,7 @@ func Unmarshal(data []byte, v any) error {
 
 // UnmarshalContext is the cancellable variant of Unmarshal.
 func UnmarshalContext(ctx context.Context, data []byte, v any) error {
-	tree, err := ParseContext(ctx, data)
+	tree, err := ParseMapContext(ctx, data)
 	if err != nil {
 		return err
 	}
@@ -138,9 +190,11 @@ func UnmarshalContext(ctx context.Context, data []byte, v any) error {
 }
 
 // A Decoder decodes a TOML document into a Go value with configurable
-// strictness.
+// strictness and configurable limits on the parse it performs.
 type Decoder struct {
 	disallowUnknown bool
+	maxDepth        int
+	maxInputSize    int
 }
 
 // NewDecoder returns a Decoder.
@@ -153,6 +207,28 @@ func (d *Decoder) DisallowUnknownFields() *Decoder {
 	return d
 }
 
+// MaxDepth bounds how deeply arrays and inline tables may nest in a document
+// this decoder accepts. The parser is a recursive descent, so a document that
+// nests without bound would exhaust the stack; one that nests deeper than the
+// limit is rejected with a SyntaxError naming it instead. Use 0 or any
+// negative value for the default of 10000, which no hand-written document
+// approaches.
+func (d *Decoder) MaxDepth(depth int) *Decoder {
+	d.maxDepth = depth
+	return d
+}
+
+// MaxInputSize bounds the size of a document this decoder accepts, in bytes; a
+// larger one is rejected before parsing starts. Use 0 or any negative value for
+// no limit, which is the default: the caller already holds the bytes, so the
+// size is a policy the caller sets rather than a protection the library
+// imposes on its own. Parse and ParseContext take no limit beyond the nesting
+// default.
+func (d *Decoder) MaxInputSize(size int) *Decoder {
+	d.maxInputSize = size
+	return d
+}
+
 // Decode parses data and stores the result in the value pointed to by v,
 // honouring the decoder's strictness settings.
 //
@@ -163,7 +239,10 @@ func (d *Decoder) Decode(data []byte, v any) error {
 
 // DecodeContext is the cancellable variant of Decode.
 func (d *Decoder) DecodeContext(ctx context.Context, data []byte, v any) error {
-	tree, err := ParseContext(ctx, data)
+	tree, _, err := parseWithOptions(ctx, data, parseOptions{
+		maxDepth:     d.maxDepth,
+		maxInputSize: d.maxInputSize,
+	}, false)
 	if err != nil {
 		return err
 	}
@@ -177,6 +256,10 @@ func (d *Decoder) DecodeContext(ctx context.Context, data []byte, v any) error {
 // then encodes as if the returned value had been passed in its place, which
 // is useful for emitting a Go type as a different TOML shape (for example, a
 // struct as an inline table or a primitive alias as a richer value).
+//
+// MarshalTOML wins over encoding.TextMarshaler when a type implements both.
+// A type that implements only encoding.TextMarshaler is encoded as a TOML
+// string holding its text, and needs no method here.
 type Marshaler interface {
 	MarshalTOML() (any, error)
 }
@@ -184,22 +267,28 @@ type Marshaler interface {
 // Unmarshaler is the inverse of Marshaler: a type that wants control over
 // how it is decoded from a TOML value may implement UnmarshalTOML. The data
 // argument is whatever the parser produced for that key: one of string,
-// bool, int64, float64, time.Time, LocalDateTime, LocalDate, LocalTime,
+// bool, int64, float64, OffsetDateTime, LocalDateTime, LocalDate, LocalTime,
-// []any, or map[string]any. UnmarshalTOML may parse, inspect, or transform
+// []any, or map[string]any. A tree built by hand may carry a plain time.Time
-// the value however it likes, then store the result by mutating its
+// where the parser would put an OffsetDateTime.
-// receiver through the standard pointer-indirection rules of the reflect
+//
-// package (i.e. via reflect.Value.Set or by reassigning fields through a
+// UnmarshalTOML may parse, inspect, or transform the value however it likes,
-// pointer the receiver holds).
+// then store the result by mutating its receiver through the standard
+// pointer-indirection rules of the reflect package (i.e. via
+// reflect.Value.Set or by reassigning fields through a pointer the receiver
+// holds).
 //
 // UnmarshalTOML is invoked from (*Decoder).Decode / Unmarshal when the
 // destination type implements the interface. The decoder does not need to
 // consult the concrete return value; whatever the receiver stores is kept.
+//
+// UnmarshalTOML wins over encoding.TextUnmarshaler when a type implements
+// both. A type that implements only encoding.TextUnmarshaler is filled from a
+// TOML string holding its text, and needs no method here.
 type Unmarshaler interface {
 	UnmarshalTOML(data any) error
 }
 
-// Marshal returns the TOML encoding of v. The output stays within TOML 1.0,
+// Marshal returns the TOML encoding of v. The output is valid TOML 1.1.
-// so it is valid under both TOML 1.0 and 1.1.
 //
 // Marshal traverses v using reflection and applies the following rules:
 //
@@ -219,10 +308,17 @@ type Unmarshaler interface {
 //     value array (for example an inline table in a mixed array) emits as an
 //     inline table.
 //   - Scalars encode as TOML scalars: bool, int64, float64, string, time.Time
-//     (offset date-time), and LocalDateTime/LocalDate/LocalTime (local
+//     and OffsetDateTime (offset date-time), and LocalDateTime/LocalDate/
-//     variants).
+//     LocalTime (local variants). A date-time writes its seconds only when the value carries
+//     them, and drops the trailing zeros of a fractional second.
+//   - A table element of a value array, and a sub-table inlined by
+//     Encoder.InlineTables, is written as an inline table, across lines when it
+//     does not fit one.
 //   - Values implementing Marshaler are encoded by calling MarshalTOML and
 //     using its result.
+//   - Values implementing encoding.TextMarshaler, and not one of the
+//     date-time types, encode as a TOML string holding the text the method
+//     returns. time.Duration is written in its canonical Go form, `1h30m0s`.
 //   - nil pointer fields are omitted.
 //
 // Marshal cannot encode cyclic data structures; passing one will loop until
@@ -246,14 +342,16 @@ func MarshalContext(ctx context.Context, v any) ([]byte, error) {
 
 // An Encoder encodes Go values into TOML.
 //
-// All options default to behaviour that preserves byte-for-byte compatibility
+// All options default to the behaviour that passes the toml-test compliance
-// with previous releases and passes the toml-test compliance suite:
+// suite in both directions:
 //
 //	GroupByKind:        true  (scalars first, then tables, then arrays of tables)
 //	OmitEmptyArrays:    false (a nil/empty []string slice emits [] as a value;
 //	                          a nil/empty []Item struct slice is still skipped)
-//	LiteralMultilineAt: 0     (always emit basic multi-line strings with
+//	LiteralMultilineAt: 0     (always emit the escaped basic form, never a
-//	                          escape sequences, never literal ones)
+//	                          literal one)
+//	InlineTablesAt:     0     (always emit a table header, never an inline
+//	                          table)
 //
 // Use the chainable option methods to opt out. The option state is private;
 // callers that need the underlying knobs reach for the methods rather than
@@ -262,6 +360,7 @@ type Encoder struct {
 	groupByKind        bool // default true; set via (*Encoder).GroupByKind
 	omitEmptyArrays    bool // default false; set via (*Encoder).OmitEmptyArrays
 	literalMultilineAt int  // default 0; set via (*Encoder).UseLiteralMultiline
+	inlineTablesAt     int  // default 0; set via (*Encoder).InlineTables
 }
 
 // NewEncoder returns an Encoder with default options.
@@ -294,6 +393,24 @@ func (e *Encoder) UseLiteralMultiline(threshold int) *Encoder {
 	return e
 }
 
+// InlineTables sets the size limit, in bytes of the single-line rendering, at
+// which a sub-table is written as an inline table instead of a table header,
+// which makes a document of small tables shorter. Use 0 or any negative value
+// to disable (always emit a header).
+//
+// A sub-table is inlined only when doing so keeps every value's type: an array
+// of tables keeps its header form, because its inline form would re-parse as a
+// value array. An inlined table that does not fit the line is written across
+// lines, which TOML 1.1 allows.
+//
+// With GroupByKind(false) the layout is already for presentation only, and an
+// inlined table follows the same rule as any other value line: it lands in the
+// section of the header that precedes it.
+func (e *Encoder) InlineTables(threshold int) *Encoder {
+	e.inlineTablesAt = threshold
+	return e
+}
+
 // Marshal encodes v to TOML bytes. It is equivalent to calling Marshal with v.
 //
 // Marshal is equivalent to MarshalContext with context.Background.
@@ -307,7 +424,12 @@ func (e *Encoder) MarshalContext(ctx context.Context, v any) ([]byte, error) {
 	enc.ctx = ctx
 	enc.opts = *e
 	if err := enc.encode(v); err != nil {
+		enc.release()
 		return nil, err
 	}
-	return enc.bytes(), nil
+	// The output leaves the pooled buffer as a copy, so the next Marshal
+	// reuses the buffer without touching what the caller holds.
+	out := slices.Clone(enc.buf.Bytes())
+	enc.release()
+	return out, nil
 }

interpres_test.go

@@ -4,6 +4,7 @@
 package interpres
 
 import (
+	"errors"
 	"math"
 	"strings"
 	"testing"
@@ -11,7 +12,7 @@ import (
 )
 
 func TestParseScalars(t *testing.T) {
-	tree, err := Parse([]byte(`
+	tree, err := ParseMap([]byte(`
 title = "interpres"
 count = 42
 ratio = 3.14
@@ -49,7 +50,7 @@ expv = 1e3
 }
 
 func TestParseInfNan(t *testing.T) {
-	tree, err := Parse([]byte("pos = inf\nneg = -inf\nbad = nan\n"))
+	tree, err := ParseMap([]byte("pos = inf\nneg = -inf\nbad = nan\n"))
 	if err != nil {
 		t.Fatalf("parse: %v", err)
 	}
@@ -65,7 +66,7 @@ func TestParseInfNan(t *testing.T) {
 }
 
 func TestParseStrings(t *testing.T) {
-	tree, err := Parse([]byte(`
+	tree, err := ParseMap([]byte(`
 basic = "a\tb\nc"
 literal = 'C:\path\no\escape'
 quote = "say \"hi\""
@@ -89,7 +90,7 @@ unicode = "\u00e9"
 }
 
 func TestParseMultilineString(t *testing.T) {
-	tree, err := Parse([]byte("text = \"\"\"\nfirst\nsecond\"\"\"\n"))
+	tree, err := ParseMap([]byte("text = \"\"\"\nfirst\nsecond\"\"\"\n"))
 	if err != nil {
 		t.Fatalf("parse: %v", err)
 	}
@@ -99,7 +100,7 @@ func TestParseMultilineString(t *testing.T) {
 }
 
 func TestParseMultilineLineEndingBackslash(t *testing.T) {
-	tree, err := Parse([]byte("text = \"\"\"\\\n    one \\\n    two\"\"\"\n"))
+	tree, err := ParseMap([]byte("text = \"\"\"\\\n    one \\\n    two\"\"\"\n"))
 	if err != nil {
 		t.Fatalf("parse: %v", err)
 	}
@@ -109,7 +110,7 @@ func TestParseMultilineLineEndingBackslash(t *testing.T) {
 }
 
 func TestParseTablesAndDottedKeys(t *testing.T) {
-	tree, err := Parse([]byte(`
+	tree, err := ParseMap([]byte(`
 owner.name = "Petr"
 
 [server]
@@ -137,7 +138,7 @@ enabled = true
 }
 
 func TestParseArrayOfTables(t *testing.T) {
-	tree, err := Parse([]byte(`
+	tree, err := ParseMap([]byte(`
 [[forms]]
 name = "contact"
 
@@ -157,7 +158,7 @@ name = "feedback"
 }
 
 func TestParseArraysAndInlineTables(t *testing.T) {
-	tree, err := Parse([]byte(`
+	tree, err := ParseMap([]byte(`
 ports = [80, 443]
 mixed = [
   "a",
@@ -183,7 +184,7 @@ point = { x = 1, y = 2 }
 }
 
 func TestParseDateTime(t *testing.T) {
-	tree, err := Parse([]byte(`
+	tree, err := ParseMap([]byte(`
 offset = 1979-05-27T07:32:00Z
 local = 1979-05-27T07:32:00
 day = 1979-05-27
@@ -192,7 +193,7 @@ clock = 07:32:00
 	if err != nil {
 		t.Fatalf("parse: %v", err)
 	}
-	if off, ok := tree["offset"].(time.Time); !ok || off.Year() != 1979 || off.Hour() != 7 {
+	if off, ok := tree["offset"].(OffsetDateTime); !ok || off.Year() != 1979 || off.Hour() != 7 {
 		t.Errorf("offset = %#v (%T)", tree["offset"], tree["offset"])
 	}
 	if ldt, ok := tree["local"].(LocalDateTime); !ok || ldt.Year() != 1979 || ldt.Hour() != 7 {
@@ -207,15 +208,15 @@ clock = 07:32:00
 }
 
 func TestDateTimeFormats(t *testing.T) {
-	tree, err := Parse([]byte("a = 1987-07-05 17:45:00Z\nb = 1987-07-05t17:45:00z\nc = 1977-12-21T10:32:00.555\n"))
+	tree, err := ParseMap([]byte("a = 1987-07-05 17:45:00Z\nb = 1987-07-05t17:45:00z\nc = 1977-12-21T10:32:00.555\n"))
 	if err != nil {
 		t.Fatalf("parse: %v", err)
 	}
-	if _, ok := tree["a"].(time.Time); !ok {
+	if _, ok := tree["a"].(OffsetDateTime); !ok {
-		t.Errorf("a is %T, want time.Time", tree["a"])
+		t.Errorf("a is %T, want OffsetDateTime", tree["a"])
 	}
-	if _, ok := tree["b"].(time.Time); !ok {
+	if _, ok := tree["b"].(OffsetDateTime); !ok {
-		t.Errorf("b is %T, want time.Time", tree["b"])
+		t.Errorf("b is %T, want OffsetDateTime", tree["b"])
 	}
 	if _, ok := tree["c"].(LocalDateTime); !ok {
 		t.Errorf("c is %T, want LocalDateTime", tree["c"])
@@ -352,7 +353,7 @@ func TestSkippedFieldTag(t *testing.T) {
 }
 
 func TestSyntaxErrorReportsLine(t *testing.T) {
-	_, err := Parse([]byte("a = 1\nb = \nc = 3\n"))
+	_, err := ParseMap([]byte("a = 1\nb = \nc = 3\n"))
 	if err == nil {
 		t.Fatal("expected a syntax error")
 	}
@@ -366,7 +367,7 @@ func TestSyntaxErrorReportsLine(t *testing.T) {
 }
 
 func TestComments(t *testing.T) {
-	tree, err := Parse([]byte(`
+	tree, err := ParseMap([]byte(`
 # a leading comment
 key = "value" # trailing comment
 # another
@@ -380,7 +381,7 @@ key = "value" # trailing comment
 }
 
 func TestDuplicateKeyRejected(t *testing.T) {
-	_, err := Parse([]byte("a = 1\na = 2\n"))
+	_, err := ParseMap([]byte("a = 1\na = 2\n"))
 	if err == nil {
 		t.Fatal("expected duplicate key error")
 	}
@@ -395,7 +396,7 @@ func TestRejectsInvalidNumbers(t *testing.T) {
 		"0x", "0o", "0b", "0b2", "0o8", "0xG",
 		"+0x1",
 	} {
-		if _, err := Parse([]byte("v = " + tok + "\n")); err == nil {
+		if _, err := ParseMap([]byte("v = " + tok + "\n")); err == nil {
 			t.Errorf("%q: expected an error, got none", tok)
 		}
 	}
@@ -408,22 +409,22 @@ func TestParseRejectsOffsetOutOfRange(t *testing.T) {
 		"1979-05-27T07:32:00+24:00",
 		"1979-05-27T07:32:00+99:99",
 	} {
-		if _, err := Parse([]byte("v = " + tok + "\n")); err == nil {
+		if _, err := ParseMap([]byte("v = " + tok + "\n")); err == nil {
 			t.Errorf("%q: expected an error, got none", tok)
 		}
 	}
 }
 
 func TestParseAcceptsOffsetBounds(t *testing.T) {
-	tree, err := Parse([]byte("a = 1979-05-27T07:32:00+23:59\nb = 1979-05-27T07:32:00-23:59\n"))
+	tree, err := ParseMap([]byte("a = 1979-05-27T07:32:00+23:59\nb = 1979-05-27T07:32:00-23:59\n"))
 	if err != nil {
 		t.Fatalf("parse: %v", err)
 	}
-	a := tree["a"].(time.Time)
+	a := tree["a"].(OffsetDateTime)
 	if _, offset := a.Zone(); offset != 23*3600+59*60 {
 		t.Fatalf("a offset = %d, want %d", offset, 23*3600+59*60)
 	}
-	b := tree["b"].(time.Time)
+	b := tree["b"].(OffsetDateTime)
 	if _, offset := b.Zone(); offset != -(23*3600 + 59*60) {
 		t.Fatalf("b offset = %d", offset)
 	}
@@ -449,7 +450,7 @@ func TestAcceptsNumberEdgeCases(t *testing.T) {
 		"-2.5E-3":     -2.5e-3,
 	}
 	for tok, want := range cases {
-		tree, err := Parse([]byte("v = " + tok + "\n"))
+		tree, err := ParseMap([]byte("v = " + tok + "\n"))
 		if err != nil {
 			t.Errorf("%q: %v", tok, err)
 			continue
@@ -461,14 +462,14 @@ func TestAcceptsNumberEdgeCases(t *testing.T) {
 }
 
 func TestRejectsTableRedefinition(t *testing.T) {
-	_, err := Parse([]byte("[a]\nx = 1\n\n[a]\ny = 2\n"))
+	_, err := ParseMap([]byte("[a]\nx = 1\n\n[a]\ny = 2\n"))
 	if err == nil {
 		t.Fatal("expected a table-redefinition error")
 	}
 }
 
 func TestAllowsImplicitThenExplicitTable(t *testing.T) {
-	tree, err := Parse([]byte("[a.b]\nx = 1\n\n[a]\ny = 2\n"))
+	tree, err := ParseMap([]byte("[a.b]\nx = 1\n\n[a]\ny = 2\n"))
 	if err != nil {
 		t.Fatalf("parse: %v", err)
 	}
@@ -482,13 +483,13 @@ func TestAllowsImplicitThenExplicitTable(t *testing.T) {
 }
 
 func TestRejectsControlCharInString(t *testing.T) {
-	if _, err := Parse([]byte("v = \"a\x01b\"\n")); err == nil {
+	if _, err := ParseMap([]byte("v = \"a\x01b\"\n")); err == nil {
 		t.Fatal("expected a control-character error")
 	}
 }
 
 func TestAllowsEscapedControlChar(t *testing.T) {
-	tree, err := Parse([]byte(`v = "\u0000"`))
+	tree, err := ParseMap([]byte(`v = "\u0000"`))
 	if err != nil {
 		t.Fatalf("parse: %v", err)
 	}
@@ -498,7 +499,7 @@ func TestAllowsEscapedControlChar(t *testing.T) {
 }
 
 func TestMultilineQuotesAtDelimiter(t *testing.T) {
-	tree, err := Parse([]byte("a = '''''two quotes'''''\n"))
+	tree, err := ParseMap([]byte("a = '''''two quotes'''''\n"))
 	if err != nil {
 		t.Fatalf("parse: %v", err)
 	}
@@ -517,7 +518,7 @@ func TestRejectsInlineTableExtension(t *testing.T) {
 		"by nested array header": "a = { b = {} }\n[[a.b.c]]\nx = 2\n",
 	}
 	for name, doc := range cases {
-		if _, err := Parse([]byte(doc)); err == nil {
+		if _, err := ParseMap([]byte(doc)); err == nil {
 			t.Errorf("%s: expected an inline-table extension error", name)
 		}
 	}
@@ -532,7 +533,7 @@ func TestArrayOfTablesFreshScopePerElement(t *testing.T) {
 		"dotted key":             "[[a]]\nb.c = 1\n[[a]]\n[a.b]\nd = 2\n",
 	}
 	for name, doc := range cases {
-		tree, err := Parse([]byte(doc))
+		tree, err := ParseMap([]byte(doc))
 		if err != nil {
 			t.Errorf("%s: %v", name, err)
 			continue
@@ -547,7 +548,7 @@ func TestArrayOfTablesFreshScopePerElement(t *testing.T) {
 		"header over dotted in one element": "[[a]]\nb.c = 1\n[a.b]\nd = 2\n",
 		"table over nested array":           "[[a]]\n[[a.b]]\n[a.b]\nx = 1\n",
 	} {
-		if _, err := Parse([]byte(doc)); err == nil {
+		if _, err := ParseMap([]byte(doc)); err == nil {
 			t.Errorf("%s: expected an error, got none", name)
 		}
 	}
@@ -565,14 +566,14 @@ func TestRejectsSpecInvalid(t *testing.T) {
 		// makes the seconds optional.
 	}
 	for name, doc := range cases {
-		if _, err := Parse([]byte(doc)); err == nil {
+		if _, err := ParseMap([]byte(doc)); err == nil {
 			t.Errorf("%s: expected an error", name)
 		}
 	}
 }
 
 func TestArrayOfTablesPerElementSubtable(t *testing.T) {
-	tree, err := Parse([]byte(`
+	tree, err := ParseMap([]byte(`
 [[forms]]
 name = "a"
 
@@ -603,7 +604,7 @@ host = "h2"
 // --- TOML 1.1 --------------------------------------------------------------
 
 func TestParseAcceptsNoSecondsDatetimes(t *testing.T) {
-	tree, err := Parse([]byte(`t = 13:37
+	tree, err := ParseMap([]byte(`t = 13:37
 dt = 1979-05-27T07:32
 odt1 = 1979-05-27 07:32Z
 odt2 = 1979-05-27 07:32-07:00
@@ -611,26 +612,28 @@ odt2 = 1979-05-27 07:32-07:00
 	if err != nil {
 		t.Fatalf("parse: %v", err)
 	}
-	if got := tree["t"].(LocalTime).String(); got != "13:37:00" {
+	// A value written without seconds comes back without them: the seconds are
-		t.Errorf("t = %q, want %q", got, "13:37:00")
+	// only written when the value carries them.
+	if got := tree["t"].(LocalTime).String(); got != "13:37" {
+		t.Errorf("t = %q, want %q", got, "13:37")
 	}
-	if got := tree["dt"].(LocalDateTime).String(); got != "1979-05-27T07:32:00" {
+	if got := tree["dt"].(LocalDateTime).String(); got != "1979-05-27T07:32" {
-		t.Errorf("dt = %q, want %q", got, "1979-05-27T07:32:00")
+		t.Errorf("dt = %q, want %q", got, "1979-05-27T07:32")
 	}
-	if got := tree["odt1"].(time.Time).Format(time.RFC3339Nano); got != "1979-05-27T07:32:00Z" {
+	if got := tree["odt1"].(OffsetDateTime).Format(time.RFC3339Nano); got != "1979-05-27T07:32:00Z" {
 		t.Errorf("odt1 = %q", got)
 	}
-	if got := tree["odt2"].(time.Time).Format(time.RFC3339Nano); got != "1979-05-27T07:32:00-07:00" {
+	if got := tree["odt2"].(OffsetDateTime).Format(time.RFC3339Nano); got != "1979-05-27T07:32:00-07:00" {
 		t.Errorf("odt2 = %q", got)
 	}
 	// The fraction still requires the seconds it belongs to.
-	if _, err := Parse([]byte("a = 07:32.5\n")); err == nil {
+	if _, err := ParseMap([]byte("a = 07:32.5\n")); err == nil {
 		t.Error("07:32.5: expected an error, got none")
 	}
 }
 
 func TestParseAcceptsEscapeAndHexEscapes(t *testing.T) {
-	tree, err := Parse([]byte(`esc = "\e"
+	tree, err := ParseMap([]byte(`esc = "\e"
 hex = "\x20\x7f\xf8"
 nul = "\x00"
 multi = """\x68\x65"""
@@ -657,14 +660,14 @@ lit = '\x20'
 	}
 	// Two digits exactly; a short or non-hex escape is an error.
 	for _, doc := range []string{`a = "\x4"`, `a = "\x"`, `a = "\xgg"`} {
-		if _, err := Parse([]byte(doc)); err == nil {
+		if _, err := ParseMap([]byte(doc)); err == nil {
 			t.Errorf("%s: expected an error, got none", doc)
 		}
 	}
 }
 
 func TestParseAcceptsMultilineInlineTables(t *testing.T) {
-	tree, err := Parse([]byte("tbl = {\n\thello = \"world\",\n\tarr = [1,\n\t\t2,\n\t],\n\tsub = {\n\t\tk = 1,\n\t},\n\tbare = 2}\n"))
+	tree, err := ParseMap([]byte("tbl = {\n\thello = \"world\",\n\tarr = [1,\n\t\t2,\n\t],\n\tsub = {\n\t\tk = 1,\n\t},\n\tbare = 2}\n"))
 	if err != nil {
 		t.Fatalf("parse: %v", err)
 	}
@@ -679,7 +682,7 @@ func TestParseAcceptsMultilineInlineTables(t *testing.T) {
 		t.Errorf("sub = %#v", tbl["sub"])
 	}
 	// Comments inside the table, and a trailing comma at both depths.
-	tree, err = Parse([]byte("m = { # one\n\t# two\n\ta = 1, # three\n\t# four\n}\n"))
+	tree, err = ParseMap([]byte("m = { # one\n\t# two\n\ta = 1, # three\n\t# four\n}\n"))
 	if err != nil {
 		t.Fatalf("parse with comments: %v", err)
 	}
@@ -687,7 +690,7 @@ func TestParseAcceptsMultilineInlineTables(t *testing.T) {
 		t.Errorf("m = %#v", m)
 	}
 	// The old single-line shapes keep working, with and without the comma.
-	if _, err := Parse([]byte("a = { b = 1, c = 2 }\n")); err != nil {
+	if _, err := ParseMap([]byte("a = { b = 1, c = 2 }\n")); err != nil {
 		t.Errorf("single line: %v", err)
 	}
 	// Still rejected: two commas, a missing value, and an unclosed table.
@@ -696,8 +699,30 @@ func TestParseAcceptsMultilineInlineTables(t *testing.T) {
 		"missing value": "a = {\n\tb =\n}\n",
 		"unterminated":  "a = { b = 1,\n",
 	} {
-		if _, err := Parse([]byte(doc)); err == nil {
+		if _, err := ParseMap([]byte(doc)); err == nil {
 			t.Errorf("%s: expected an error, got none", name)
 		}
 	}
 }
+
+func TestParseNestingLimit(t *testing.T) {
+	// The parser is a recursive descent, so a document that nests without bound
+	// is rejected instead of exhausting the stack.
+	deep := func(n int) []byte {
+		return []byte("v = " + strings.Repeat("[", n) + strings.Repeat("]", n) + "\n")
+	}
+	if _, err := ParseMap(deep(100)); err != nil {
+		t.Fatalf("a document well inside the limit: %v", err)
+	}
+	_, err := ParseMap(deep(maxNestingDepth + 1))
+	if err == nil {
+		t.Fatal("expected a nesting error")
+	}
+	var se *SyntaxError
+	if !errors.As(err, &se) {
+		t.Fatalf("expected a *SyntaxError, got %T: %v", err, err)
+	}
+	if !strings.Contains(se.Msg, "nesting") {
+		t.Errorf("Msg = %q, want it to name the nesting limit", se.Msg)
+	}
+}

justfile

@@ -92,9 +92,9 @@ run:
 dev:
     go run -buildvcs=true {{package}}
 
-# Runs the official toml-test compliance suite against the built adapter; toml-test must be on PATH (go install github.com/toml-lang/toml-test/v2/cmd/toml-test@v2.2.0); not standard because no canonical recipe covers a domain compliance suite.
+# Runs the official toml-test compliance suite in both directions, decoder and encoder, against the built adapter; toml-test must be on PATH (go install github.com/toml-lang/toml-test/v2/cmd/toml-test@v2.2.0); not standard because no canonical recipe covers a domain compliance suite.
 toml-test: build
-    toml-test test -decoder=bin/interpres-decode -toml=1.1
+    toml-test test -decoder=bin/interpres-decode -encoder='bin/interpres-decode -encode' -toml=1.1
 
 # Coverage report as an HTML map from the gate's profile; not standard because the gate needs only the numeric floor, and a browser artefact is exploration, not a gate.
 coverage-html: test

number.go

@@ -41,6 +41,15 @@ func decodeDecimalInt(tok string) (any, error) {
 	if err := checkNoLeadingZero(digits); err != nil {
 		return nil, err
 	}
+	// An unsigned token parses in place; only a sign needs the concatenated
+	// copy, and concatenating an empty sign still allocated.
+	if sign == "" {
+		i, err := strconv.ParseInt(digits, 10, 64)
+		if err != nil {
+			return nil, fmt.Errorf("integer %q out of range", tok)
+		}
+		return i, nil
+	}
 	i, err := strconv.ParseInt(sign+digits, 10, 64)
 	if err != nil {
 		return nil, fmt.Errorf("integer %q out of range", tok)