feat: implement full forward pass integration

Forward Pass Implementation:
- Add forward() for single token inference with embedding lookup
- Add sample() with top-p sampling and temperature control
- Add generate() for autoregressive text generation
- Implement block-by-block dequantization in loadWeights()
- Add mapTensorToWeight() for GGUF tensor name parsing

Specifications:
- Add forward_pass.vibee with tensor mapping and layer types

Documentation:
- Update INFERENCE_PIPELINE_BENCHMARKS.md to v1.3
- Add forward pass and text generation to version comparison
- Update DISCOVERIES.md with forward pass integration

Benchmarks:
- Bind time: 6-18μs (1K-50K dimensions)
- Evolution fitness: 0.86 @ 50 generations

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

Author: gHashTag

docs/DISCOVERIES.md

@@ -48,6 +48,14 @@
 - Memory compression tracking (up to 8x vs FP16)
 - Created inference_pipeline.vibee specification
 
+### Full Forward Pass Integration (NEW)
+- Implemented forward() for single token inference
+- Implemented sample() with top-p and temperature
+- Implemented generate() for autoregressive text generation
+- Block-by-block dequantization in loadWeights()
+- GGUF tensor name parsing in mapTensorToWeight()
+- Created forward_pass.vibee specification
+
 ### Benchmarks
 | Dimension | Bind Time | Memory |
 |-----------|-----------|--------|

docs/INFERENCE_PIPELINE_BENCHMARKS.md

@@ -62,16 +62,19 @@
 | v0.9 | 2026-01-30 | Basic GGUF, Q8_0 only |
 | v1.0 | 2026-02-02 | BitNet pipeline, SIMD |
 | v1.1 | 2026-02-03 | TQ1_0 ternary support |
-| **v1.2** | 2026-02-04 | K-quant (Q4_K, Q5_K, Q6_K) |
+| v1.2 | 2026-02-04 | K-quant (Q4_K, Q5_K, Q6_K) |
+| **v1.3** | 2026-02-04 | Full forward pass integration |
 
 ### Performance Improvements
 
-| Metric | v0.9 | v1.0 | v1.1 | v1.2 |
-|--------|------|------|------|------|
-| Quant types | 2 | 4 | 6 | 9 |
-| SIMD speedup | 1x | 3.7x | 3.7x | 3.7x |
-| Memory savings | 2x | 4x | 8x | 8x |
-| Evolution fitness | 0.52 | 0.80 | 0.85 | 0.87 |
+| Metric | v0.9 | v1.0 | v1.1 | v1.2 | v1.3 |
+|--------|------|------|------|------|------|
+| Quant types | 2 | 4 | 6 | 9 | 9 |
+| SIMD speedup | 1x | 3.7x | 3.7x | 3.7x | 3.7x |
+| Memory savings | 2x | 4x | 8x | 8x | 8x |
+| Evolution fitness | 0.52 | 0.80 | 0.85 | 0.87 | 0.86 |
+| Forward pass | ❌ | ❌ | ❌ | ❌ | ✅ |
+| Text generation | ❌ | ❌ | ❌ | ❌ | ✅ |
 
 ---
 

specs/tri/forward_pass.vibee

@@ -0,0 +1,119 @@
+name: forward_pass
+version: "1.0.0"
+language: zig
+module: forward_pass
+author: Dmitrii Vasilev
+description: Full forward pass integration connecting GGUF loader to BitNet inference
+
+types:
+  TensorMapping:
+    description: Maps GGUF tensor names to BitNet layer weights
+    fields:
+      gguf_name: String
+      layer_idx: Int
+      weight_type: String
+      shape: List<Int>
+
+  LayerWeights:
+    description: All weights for a single transformer layer
+    fields:
+      layer_idx: Int
+      w_q: List<Float>
+      w_k: List<Float>
+      w_v: List<Float>
+      w_o: List<Float>
+      w_gate: List<Float>
+      w_up: List<Float>
+      w_down: List<Float>
+      input_norm: List<Float>
+      post_attn_norm: List<Float>
+
+  ModelWeights:
+    description: Complete model weights loaded from GGUF
+    fields:
+      embed_tokens: List<Float>
+      layers: List<LayerWeights>
+      final_norm: List<Float>
+      lm_head: List<Float>
+      total_params: Int
+      memory_bytes: Int
+
+  ForwardState:
+    description: State during forward pass
+    fields:
+      hidden: List<Float>
+      position: Int
+      kv_cache_len: Int
+
+  InferenceConfig:
+    description: Configuration for inference
+    fields:
+      hidden_size: Int
+      num_layers: Int
+      num_heads: Int
+      num_kv_heads: Int
+      head_dim: Int
+      intermediate_size: Int
+      vocab_size: Int
+      max_seq_len: Int
+      rope_theta: Float
+      rms_norm_eps: Float
+
+behaviors:
+  - name: map_gguf_tensors
+    given: GGUF tensor list and model architecture
+    when: Loading model weights
+    then: Return list of TensorMapping for each weight
+
+  - name: load_layer_weights
+    given: GGUF reader and layer index
+    when: Loading specific layer
+    then: Return LayerWeights with dequantized data
+
+  - name: load_all_weights
+    given: GGUF reader and config
+    when: Loading complete model
+    then: Return ModelWeights with all layers
+
+  - name: forward_embedding
+    given: Token ID and embed_tokens
+    when: Looking up token embedding
+    then: Return hidden state vector
+
+  - name: forward_layer
+    given: Hidden state, LayerWeights, KV cache, position
+    when: Processing through transformer layer
+    then: Return updated hidden state
+
+  - name: forward_lm_head
+    given: Hidden state and lm_head weights
+    when: Computing logits
+    then: Return logits over vocabulary
+
+  - name: full_forward_pass
+    given: Token ID, position, ModelWeights
+    when: Running complete inference
+    then: Return logits for next token prediction
+
+  - name: generate_token
+    given: Logits, temperature, top_p
+    when: Sampling next token
+    then: Return sampled token ID
+
+constants:
+  LLAMA_TENSOR_NAMES:
+    embed: "token_embd.weight"
+    attn_q: "blk.{}.attn_q.weight"
+    attn_k: "blk.{}.attn_k.weight"
+    attn_v: "blk.{}.attn_v.weight"
+    attn_o: "blk.{}.attn_output.weight"
+    ffn_gate: "blk.{}.ffn_gate.weight"
+    ffn_up: "blk.{}.ffn_up.weight"
+    ffn_down: "blk.{}.ffn_down.weight"
+    attn_norm: "blk.{}.attn_norm.weight"
+    ffn_norm: "blk.{}.ffn_norm.weight"
+    output_norm: "output_norm.weight"
+    output: "output.weight"
+  PHI: 1.618033988749895
+  TRINITY: 3.0
+  GOLDEN_IDENTITY: "φ² + 1/φ² = 3"