feat(gguf-converter): add GGUF → TRI converter specification

- Create specs/tri/gguf_to_tri.vibee as single source of truth
- Support F32/F16/BF16/Q4/Q5/Q6/Q8/TQ1/TQ2 tensor types
- Add parallel quantization via thread pool
- Fix Zig syntax errors (packed keyword, while loop)
- Add 10 unit tests for converter
- Update docs with GGUF converter info

Compression ratios:
- F32 → Ternary: 16x
- F16 → Ternary: 8x
- Q8 → Ternary: 4x
- Q4 → Ternary: 2x

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

Author: gHashTag

docs/PERFORMANCE_COMPARISON.md

@@ -159,7 +159,24 @@ Speedup: 1.1x over baseline
 | llama.cpp | GGUF | Q4/Q8 | No | High |
 | vLLM | HF | FP16/INT8 | No | High |
 | TGI | HF | FP16/INT8 | No | High |
-| **Trinity** | **.tri** | **Ternary** | **Yes** | **Low** |
+| **Trinity** | **GGUF → .tri** | **Ternary** | **Yes** | **Low** |
+
+### 7.2 GGUF → TRI Converter
+
+Trinity now supports converting any GGUF model to ternary .tri format:
+
+| Input Format | Compression | Memory Savings |
+|--------------|-------------|----------------|
+| F32 → Ternary | 16x | 93.75% |
+| F16 → Ternary | 8x | 87.5% |
+| Q8 → Ternary | 4x | 75% |
+| Q4 → Ternary | 2x | 50% |
+
+**Supported GGUF tensor types:**
+- F32, F16, BF16 (full precision)
+- Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, Q8_1 (legacy quants)
+- Q4_K, Q5_K, Q6_K, Q8_K (K-quants)
+- TQ1_0, TQ2_0 (native ternary)
 
 ### 7.2 Performance Targets
 

docs/TECH_TREE_STRATEGY.md

@@ -37,7 +37,16 @@
 │  ✅ GQA (Grouped Query Attention) support                       │
 │  ✅ Ternary QKV projection integration                          │
 │                                                                 │
-│  COMPLETED (Phase 6 - E2E Verification) [NEW]                   │
+│  COMPLETED (Phase 5b - GGUF Converter)                          │
+│  ═════════════════════════════════════                          │
+│  ✅ GGUF → TRI converter specification (gguf_to_tri.vibee)      │
+│  ✅ Support for F32/F16/BF16/Q4/Q5/Q6/Q8 tensor types           │
+│  ✅ Per-group quantization (group_size=128)                     │
+│  ✅ Parallel quantization via thread pool                       │
+│  ✅ Metadata extraction (vocab, tokenizer)                      │
+│  ✅ CLI integration (vibeec convert)                            │
+│                                                                 │
+│  COMPLETED (Phase 6 - E2E Verification)                         │
 │  ════════════════════════════════════════════                   │
 │  ✅ GPU benchmarks (RTX 3090: 298K tokens/s)                    │
 │  ✅ 69 unit tests passing (100%)                                │