Fix INT4 quantization references in WASM build files

Updated all INT8 references to INT4 to match the actual quantization
level used. The conversion scripts generate INT4 models and lib.rs
expects INT4 files, but download-models.mjs and README.md still
referenced INT8, causing build failures.

Author: jdalton

scripts/wasm/download-models.mjs

@@ -4,8 +4,8 @@
  * WHAT THIS DOWNLOADS:
  * 1. MiniLM model (int8 quantized, ~17MB)
  * 2. MiniLM tokenizer (~500KB)
- * 3. CodeT5 encoder (int8 quantized, ~60MB)
- * 4. CodeT5 decoder (int8 quantized, ~60MB)
+ * 3. CodeT5 encoder (int4 quantized, ~30MB)
+ * 4. CodeT5 decoder (int4 quantized, ~60MB)
  * 5. CodeT5 tokenizer (~500KB)
  * 6. ONNX Runtime WASM (~2-5MB)
  * 7. Yoga Layout WASM (~95KB) - copied from node_modules
@@ -50,14 +50,14 @@ const FILES = [
   // CodeT5 (needs manual conversion first - see convert-codet5.mjs).
   {
     copyFrom: null, // Set after conversion
-    description: 'CodeT5 encoder (int8)',
-    name: 'codet5-encoder-int8.onnx',
+    description: 'CodeT5 encoder (int4)',
+    name: 'codet5-encoder-int4.onnx',
     url: null, // Needs conversion first
   },
   {
     copyFrom: null,
-    description: 'CodeT5 decoder (int8)',
-    name: 'codet5-decoder-int8.onnx',
+    description: 'CodeT5 decoder (int4)',
+    name: 'codet5-decoder-int4.onnx',
     url: null, // Needs conversion first
   },
   {

wasm-bundle/README.md

@@ -5,11 +5,11 @@ Single WASM file containing all AI models and execution engines for Socket CLI.
 ## Architecture
 
 ```
-socket-ai.wasm (~145MB)
+socket-ai.wasm (~115MB)
   ├─ ONNX Runtime (~2-5MB) - ML execution engine
   ├─ MiniLM model (~17MB int8) - Semantic understanding
-  ├─ CodeT5 encoder (~60MB int8) - Code generation (encoder)
-  ├─ CodeT5 decoder (~60MB int8) - Code generation (decoder)
+  ├─ CodeT5 encoder (~30MB int4) - Code generation (encoder)
+  ├─ CodeT5 decoder (~60MB int4) - Code generation (decoder)
   ├─ Tokenizers (~1MB) - Text tokenization
   └─ Yoga Layout (~95KB) - Flexbox layout engine
 ```
@@ -41,8 +41,8 @@ node scripts/wasm/download-models.mjs
 - Yoga Layout WASM (from node_modules)
 
 **What needs conversion** (see next step):
-- CodeT5 encoder (PyTorch → ONNX int8)
-- CodeT5 decoder (PyTorch → ONNX int8)
+- CodeT5 encoder (PyTorch → ONNX int4)
+- CodeT5 decoder (PyTorch → ONNX int4)
 
 ### 3. Convert CodeT5 Models (One-Time)
 
@@ -57,7 +57,7 @@ node scripts/wasm/convert-codet5.mjs
 **What it does**:
 - Downloads `Salesforce/codet5-small` from HuggingFace
 - Exports PyTorch → ONNX format
-- Quantizes fp32 → int8
+- Quantizes fp32 → int4 (4-bit weights, 50% smaller than int8)
 - Saves to `.cache/models/`
 
 ### 4. Build Unified WASM
@@ -74,7 +74,7 @@ node scripts/wasm/build-unified-wasm.mjs
 5. Generates `external/socket-ai-sync.mjs`
 
 **Output**:
-- `wasm-bundle/pkg/socket_ai_bg.wasm` (~145MB)
+- `wasm-bundle/pkg/socket_ai_bg.wasm` (~115MB)
 - `external/socket-ai-sync.mjs` (brotli+base64 embedded)
 
 ## Distribution Pipeline
@@ -169,8 +169,8 @@ external/
 .cache/models/              # Downloaded models (gitignored)
 ├── minilm-int8.onnx
 ├── minilm-tokenizer.json
-├── codet5-encoder-int8.onnx
-├── codet5-decoder-int8.onnx
+├── codet5-encoder-int4.onnx
+├── codet5-decoder-int4.onnx
 ├── codet5-tokenizer.json
 ├── ort-wasm-simd-threaded.wasm
 └── yoga.wasm
@@ -181,8 +181,8 @@ external/
 ### 1. Create CodeT5 Conversion Script
 
 ```bash
-# TODO: Create scripts/wasm/convert-codet5.mjs
-# Uses Python + optimum-cli to convert PyTorch → ONNX int8
+# NOTE: Already created - see scripts/wasm/convert-codet5.mjs
+# Uses Python + optimum to convert PyTorch → ONNX int4
 ```
 
 ### 2. Update Native Stub for .bz Detection
@@ -261,7 +261,7 @@ node bin/bootstrap.js --version
 | **Initialization** | ~150-200ms | ~50-100ms |
 | **Memory layout** | Fragmented | Contiguous |
 | **Distribution** | Complex | Simple |
-| **Size (raw)** | ~140MB total | ~145MB (+5MB overhead) |
+| **Size (raw)** | ~140MB total | ~115MB (30MB savings with int4) |
 | **Size (brotli)** | N/A | ~50-70MB base64 |
 | **Size (final)** | ~10MB | ~20-30MB (estimated) |
 
@@ -299,10 +299,10 @@ node scripts/wasm/convert-codet5.mjs
 
 ### WASM Too Large
 
-The WASM file is ~145MB which is expected given:
-- CodeT5 encoder: 60MB
-- CodeT5 decoder: 60MB
-- MiniLM: 17MB
+The WASM file is ~115MB which is expected given:
+- CodeT5 encoder: 30MB (int4)
+- CodeT5 decoder: 60MB (int4)
+- MiniLM: 17MB (int8)
 - ONNX Runtime: 2-5MB
 - Yoga: <1MB