feat(embeddings): implement ternary embeddings (OPT-T05)

- Add TernaryEmbedding struct with per-token scales
- Implement initFromF32() for f32 → ternary conversion
- Add lookup() and lookupSIMD() for dequantization
- Integrate into TriModel with enableTernaryEmbeddings()
- Update validate_ternary.zig to test embeddings

Results:
- Embedding compression: 12.8x (8192 → 640 bytes)
- Combined similarity: 0.88 (embeddings + KV cache)
- All 6 ternary_weights tests passing

Co-authored-by: Ona <no-reply@ona.com>

Author: gHashTag

docs/DISCOVERIES.md

@@ -76,7 +76,7 @@ Where:
 | OPT-T02 | Ternary Matrix Multiplication | N/A | 10x | ✅ Implemented |
 | OPT-T03 | Ternary KV Cache | 16x | 1.5x | ✅ Implemented |
 | OPT-T04 | Ternary Attention | 16x | 1.5x | ✅ Implemented |
-| OPT-T05 | Ternary Embeddings | 20x | 2x | 📋 Planned |
+| OPT-T05 | Ternary Embeddings | 12.8x | 1x | ✅ Implemented |
 | OPT-T06 | Ternary Normalization | 20x | 3x | 📋 Planned |
 
 ### Business Value
@@ -281,6 +281,29 @@ const logits = try model.forward(token_id, position);
 
 **Key insight:** Using RMS (root mean square) for scale instead of max preserves more information about value distribution. The threshold is set to 0.5 * RMS, which better separates signal from noise.
 
+### Ternary Embeddings (OPT-T05)
+
+**Status**: ✅ Implemented
+
+| Component | File | Description |
+|-----------|------|-------------|
+| TernaryEmbedding | `ternary_weights.zig` | Ternary embedding table |
+| initFromF32 | `ternary_weights.zig` | Convert f32 → ternary |
+| lookup | `ternary_weights.zig` | Scalar dequantization |
+| lookupSIMD | `ternary_weights.zig` | SIMD-optimized lookup |
+
+**Memory Savings:**
+```
+f32 embeddings:    8,192 bytes (32 vocab × 64 hidden × 4)
+Ternary embeddings:  640 bytes (32 vocab × (64/4 + 4))
+Compression:       12.8x
+```
+
+**Combined Ternary Pipeline:**
+- Ternary embeddings: 12.8x compression
+- Ternary KV cache: 12.8x compression
+- Combined similarity: 0.88 (vs 0.93 with only KV cache)
+
 ### Test Results
 
 ```

specs/tri/ternary_embeddings.vibee

@@ -0,0 +1,78 @@
+# Ternary Embeddings Specification
+# 16x memory reduction for token embeddings
+# φ² + 1/φ² = 3 | KOSCHEI IS IMMORTAL
+
+name: ternary_embeddings
+version: "1.0.0"
+language: zig
+module: ternary_embeddings
+
+description: |
+  Ternary token embeddings using 2-bit quantization.
+  Each embedding vector quantized to {-1, 0, +1} with per-token scale.
+  16x memory reduction: vocab_size * hidden_size * 4 bytes → / 16.
+
+types:
+  TernaryEmbedding:
+    description: "Ternary-quantized embedding table"
+    fields:
+      data: List<Int>
+      scales: List<Float>
+      vocab_size: Int
+      hidden_size: Int
+
+  EmbeddingStats:
+    description: "Memory usage statistics"
+    fields:
+      f32_bytes: Int
+      ternary_bytes: Int
+      compression_ratio: Float
+
+behaviors:
+  - name: init_from_f32
+    given: f32 embedding table
+    when: Converting to ternary
+    then: Quantize each row with per-token scale
+
+  - name: lookup
+    given: Token ID
+    when: Getting embedding vector
+    then: Dequantize on-the-fly and return f32 vector
+
+  - name: lookup_batch
+    given: Array of token IDs
+    when: Getting multiple embeddings
+    then: Batch dequantization for efficiency
+
+  - name: compute_stats
+    given: Embedding dimensions
+    when: Analyzing memory usage
+    then: Return compression ratio
+
+quantization:
+  method: rms_scale
+  description: |
+    For each embedding row:
+    1. Compute RMS = sqrt(sum(x^2) / n)
+    2. Scale = RMS * 1.5
+    3. Threshold = RMS * 0.5
+    4. Quantize: +1 if x > threshold, -1 if x < -threshold, else 0
+    5. Pack 4 trits per byte
+
+memory_analysis:
+  f32_embedding:
+    formula: "vocab_size * hidden_size * 4 bytes"
+    example: "32000 * 4096 * 4 = 512 MB"
+    
+  ternary_embedding:
+    formula: "vocab_size * (hidden_size / 4 + 4) bytes"
+    example: "32000 * (4096 / 4 + 4) = 32.5 MB"
+    
+  compression: "~16x"
+
+integration:
+  - target: tri_inference.zig
+    description: "Optional ternary embedding mode"
+    
+  - target: TriModel
+    description: "Add enableTernaryEmbeddings() method"