- New leaf package `audio` (ADR-0017): SpeechModel/SpeechProvider and
  TranscriptionModel/TranscriptionProvider with imagegen conventions
  (zero value = backend default, functional options + Apply, bytes
  in/out, never URLs). Root re-exports added.
- imagegen.Editor (ADR-0018): optional image-to-image interface —
  EditRequest carries the generation knobs plus Init image and
  denoising Strength; separate interface so existing Models keep
  compiling.
- provider/llamaswap implements all of it: POST /v1/audio/speech (JSON,
  raw-audio response, MIME from Content-Type with format fallback),
  POST /v1/audio/transcriptions (multipart, response_format=json),
  ListVoices (GET /v1/audio/voices?model=, tolerant of string-list and
  object-list shapes), POST /sdapi/v1/img2img (txt2img wire +
  init_images/denoising_strength, shared image decode), and Health(ctx)
  (GET /health) — a cheap liveness probe for often-offline hosts.
- Hermetic httptest coverage for every new wire shape and validation
  path; README sections + support-matrix footnote updated in the same
  commit (also corrects the stale /v1/images/generations claim — the
  image path has been SDAPI since the seed fix).

First consumer: mort's llamaswap media tool cluster (status / image /
TTS / STT agent tools against the netherstorm host).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AXQxVhXBw8PwFAtsVrXSmj

@@ -0,0 +10,4 @@
+	"mime/multipart"
+	"net/http"
+	"net/url"
+	"strings"

@@ -0,0 +20,4 @@
+// served by llama-swap (routed to a kokoro/chatterbox-style OpenAI-compatible
+// upstream). The id is passed through verbatim and selects which upstream
+// llama-swap loads.
+func (p *Provider) SpeechModel(id string, opts ...majaudio.SpeechModelOption) (majaudio.SpeechModel, error) {

@@ -0,0 +44,4 @@
+}
+
+// Speak implements audio.SpeechModel via POST {base}/v1/audio/speech.
+func (m *speechModel) Speak(ctx context.Context, req majaudio.SpeechRequest, opts ...majaudio.SpeechOption) (*majaudio.SpeechResult, error) {

@@ -0,0 +75,4 @@
+// audio/mpeg (the OpenAI endpoint's default container is mp3).
+func speechMIME(contentType, format string) string {
+	if mt, _, err := mime.ParseMediaType(contentType); err == nil {
+		if strings.HasPrefix(mt, "audio/") || strings.HasPrefix(mt, "video/") {

@@ -0,0 +90,4 @@
+		return "audio/aac"
+	case "flac":
+		return "audio/flac"
+	case "pcm":

@@ -0,0 +122,4 @@
+		return nil, fmt.Errorf("%w: transcription requires audio bytes", llm.ErrUnsupported)
+	}
+
+	var buf bytes.Buffer

@@ -0,0 +131,4 @@
+	if _, err := fw.Write(req.Audio); err != nil {
+		return nil, fmt.Errorf("llama-swap: build transcription form: %w", err)
+	}
+	fields := map[string]string{

@@ -0,0 +156,4 @@
+	var out struct {
+		Text string `json:"text"`
+	}
+	if err := json.Unmarshal(raw, &out); err != nil {

@@ -0,0 +168,4 @@
+	if req.Filename != "" {
+		return req.Filename
+	}
+	switch strings.ToLower(req.MIME) {

@@ -0,0 +173,4 @@
+		return "audio.mp3"
+	case "audio/wav", "audio/x-wav", "audio/wave":
+		return "audio.wav"
+	case "audio/ogg", "audio/opus":

@@ -0,0 +218,4 @@
+	}
+
+	var names []string
+	if err := json.Unmarshal(list, &names); err == nil {

@@ -0,0 +222,4 @@
+		return names, nil
+	}
+	// A failed decode above may have partially populated names — start fresh.
+	names = nil

@@ -0,0 +228,4 @@
+		Name    string `json:"name"`
+		VoiceID string `json:"voice_id"`
+	}
+	if err := json.Unmarshal(list, &objs); err == nil {

@@ -0,0 +248,4 @@
+// response body and its Content-Type — the sibling of doJSON for endpoints
+// whose success payload is not JSON (audio bytes) or whose shape varies.
+// contentType sets the request Content-Type when body is non-nil.
+func (p *Provider) doRaw(ctx context.Context, method, path, model, contentType string, body io.Reader) ([]byte, string, error) {

@@ -0,0 +270,4 @@
+	if resp.StatusCode/100 != 2 {
+		return nil, "", p.apiError(resp, model)
+	}
+	data, err := io.ReadAll(io.LimitReader(resp.Body, maxResponseBytes))

@@ -0,0 +15,4 @@
+
+func editInit(t *testing.T) imagegen.Image {
+	t.Helper()
+	raw, err := base64.StdEncoding.DecodeString(onePixelPNG)

@@ -0,0 +13,4 @@
+// request will take (a cold model swap can still block for minutes). Bound
+// the probe with a short context deadline; the client has no timeout by
+// design.
+func (p *Provider) Health(ctx context.Context) error {

@@ -0,0 +29,4 @@
+		return fmt.Errorf("llama-swap: health probe: %w", err)
+	}
+	defer resp.Body.Close()
+	_, _ = io.Copy(io.Discard, io.LimitReader(resp.Body, 1<<10))

@@ -0,0 +35,4 @@
+	}
+	return nil
+}
+

@@ -110,0 +144,4 @@
+		return nil, fmt.Errorf("%w: %v", llm.ErrUnsupported, err)
+	}
+
+	wire := img2imgRequest{