Merge feat/parakeet-engine: Parakeet TDT 0.6B v3 English engine (spec 001 workstream 7)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

# Conflicts:
#	AGENTS.md
#	bun.lock
#	docs/Decisions.md
#	src/lib/models.ts
#	src/worker/inference.worker.ts
#	src/worker/protocol.ts
#	tests/models.test.ts

src/lib/models.ts

@@ -35,9 +35,32 @@ export interface DtypeDefault {
   note: string;
 }
 
+/**
+ * Which inference path a catalog entry runs on (D-WS7a). Whisper models use
+ * the Transformers.js ASR pipeline; Parakeet uses the worker's own TDT path
+ * (src/worker/parakeetEngine.ts). Absent ⇒ "whisper", so the Whisper entries
+ * carry no field at all.
+ */
+export type ModelFamily = "whisper" | "parakeet";
+
+/** Exact weight-file sizes per dtype, for entries whose files we know (D-WS7a). */
+export interface ExactModelBytes {
+  encoder: Partial<Record<WhisperDtype, number>>;
+  decoder: Partial<Record<WhisperDtype, number>>;
+  /** Tokenizer/vocab/config files. */
+  aux: number;
+}
+
 export interface WhisperModel {
   /** Hugging Face Hub id, e.g. "onnx-community/whisper-tiny.en". */
   id: string;
+  /** Inference path (D-WS7a); absent ⇒ "whisper". */
+  family?: ModelFamily;
+  /**
+   * Real file sizes; when present they replace the parameter-count estimate
+   * in `estimateDownloadBytes` / `estimatePeakLoadBytes`.
+   */
+  exactBytes?: ExactModelBytes;
   /** Human-readable name for the picker. */
   displayName: string;
   /** False for `.en` (English-only) variants. */
@@ -63,8 +86,23 @@ export interface WhisperModel {
    * CTranslate2 conversions on the Hub (`Systran/faster-whisper-*`;
    * large-v3-turbo → `mobiuslabsgmbh/faster-whisper-large-v3-turbo`), so the
    * pinned library version stays the single authority on the exact repo.
+   * Absent for browser-only families (Parakeet, D-WS7a): the server engine
+   * answers 400 for them and the picker hides them while it is selected.
    */
-  fasterWhisperModel: string;
+  fasterWhisperModel?: string;
+}
+
+/** The entry's inference path; absent family (or unknown model) ⇒ "whisper". */
+export function modelFamily(model: WhisperModel | undefined): ModelFamily {
+  return model?.family ?? "whisper";
+}
+
+/** Bytes of one part at one dtype: the exact file size when known, else the estimate. */
+function partBytes(model: WhisperModel, part: "encoder" | "decoder", dtype: WhisperDtype): number {
+  const exact = model.exactBytes?.[part][dtype];
+  if (exact !== undefined) return exact;
+  const params = part === "encoder" ? model.encoderParamsM : model.decoderParamsM;
+  return params * 1e6 * BYTES_PER_PARAM[dtype];
 }
 
 // ---------------------------------------------------------------------------
@@ -90,9 +128,9 @@ export function estimateDownloadBytes(
   const encDtype = typeof dtype === "string" ? dtype : dtype.encoder_model;
   const decDtype = typeof dtype === "string" ? dtype : dtype.decoder_model_merged;
   return Math.round(
-    model.encoderParamsM * 1e6 * BYTES_PER_PARAM[encDtype] +
-      model.decoderParamsM * 1e6 * BYTES_PER_PARAM[decDtype] +
-      AUX_FILES_BYTES,
+    partBytes(model, "encoder", encDtype) +
+      partBytes(model, "decoder", decDtype) +
+      (model.exactBytes?.aux ?? AUX_FILES_BYTES),
   );
 }
 
@@ -116,8 +154,8 @@ export function formatModelSize(bytes: number): string {
 export function estimatePeakLoadBytes(model: WhisperModel, dtype: DtypeSpec): number {
   const encDtype = typeof dtype === "string" ? dtype : dtype.encoder_model;
   const decDtype = typeof dtype === "string" ? dtype : dtype.decoder_model_merged;
-  const encoderBytes = model.encoderParamsM * 1e6 * BYTES_PER_PARAM[encDtype];
-  const decoderBytes = model.decoderParamsM * 1e6 * BYTES_PER_PARAM[decDtype];
+  const encoderBytes = partBytes(model, "encoder", encDtype);
+  const decoderBytes = partBytes(model, "decoder", decDtype);
   return Math.round(2 * Math.max(encoderBytes, decoderBytes));
 }
 
@@ -209,6 +247,10 @@ function danishRepos(name: string): Pick<WhisperModel, "id" | "fasterWhisperMode
 // The catalog (D-006: curated, not exhaustive)
 // ---------------------------------------------------------------------------
 
+/** Parakeet TDT 0.6B v3 ONNX repo and the revision its file sizes were taken from (D-WS7a). */
+export const PARAKEET_MODEL_ID = "ysdede/parakeet-tdt-0.6b-v3-onnx";
+export const PARAKEET_REVISION = "e5356bf6e24bdb7a269536b0e9e7869bc034b893";
+
 export const MODELS: readonly WhisperModel[] = [
   {
     id: "onnx-community/whisper-tiny.en",
@@ -325,6 +367,47 @@ export const MODELS: readonly WhisperModel[] = [
     decoderParamsM: 156,
     defaults: { wasm: DANISH_WASM_Q4, webgpu: WEBGPU_FP16_Q4, webnn: WEBNN_FP16 },
   },
+  {
+    // NVIDIA Parakeet TDT 0.6B v3 (CC-BY-4.0), NeMo→ONNX conversion by
+    // ysdede, run through parakeet.js on our onnxruntime-web (D-WS7a).
+    // English-only in this app (the model itself knows 25 European
+    // languages, but only English is measured here). Desktop-only: its
+    // ~1.3 GB load peak is far past every phone budget (D-027–D-029), so
+    // the picker hides it on mobile instead of just warning.
+    id: PARAKEET_MODEL_ID,
+    family: "parakeet",
+    displayName: "Parakeet v3 (English)",
+    multilingual: false,
+    canTranslate: false,
+    // Parameter counts only feed fallbacks; the sizes below are exact.
+    encoderParamsM: 609,
+    decoderParamsM: 17,
+    exactBytes: {
+      encoder: { q8: 652_183_999, fp16: 1_238_960_452 },
+      decoder: { q8: 18_202_004, fp16: 36_266_140 },
+      aux: 102_132, // vocab.txt
+    },
+    defaults: {
+      wasm: {
+        dtype: "q8",
+        verified: false,
+        note: "int8 encoder + decoder: the measured WASM path (D-WS7a).",
+      },
+      // Only used when WebGPU is chosen explicitly — "auto" never picks it
+      // for Parakeet (D-WS7a): int8 on WebGPU is reported broken, and fp16
+      // is a separate 1.24 GB download. The decoder stays int8 on WASM.
+      webgpu: {
+        dtype: { encoder_model: "fp16", decoder_model_merged: "q8" },
+        verified: false,
+        note: "fp16 encoder on WebGPU (needs shader-f16) + int8 decoder on WASM.",
+      },
+      webnn: {
+        dtype: "q8",
+        verified: false,
+        note: "No WebNN path: runs the WASM int8 setup.",
+      },
+    },
+  },
 ];
 
 export function getModel(id: string): WhisperModel | undefined {