Issue #68: adopt Trinity config (hidden=243, vocab=729-padded, GF16) #520

Task

Adopt Trinity configuration for 16MB Parameter Golf budget.

Config (downscaled from full Trinity for size)

• VOCAB_SIZE = 729 (3⁶, padded: active 0-255, logits 256-728 = -inf)
• HIDDEN_DIM = 243 (3⁵) ← downscaled from 729 for 16MB budget
• EMBED_DIM = 243 (3⁵, tied with hidden)
• CONTEXT_LEN = 81 (3⁴)
• NUM_BLOCKS = 9 (3²)
• HEADS = 9 (3²)
• HEAD_DIM = 27 (3³)
• FFN_HIDDEN = 729 (3×hidden, 3⁶)
• Quantization = GF16 (4-bit)
• Optimizer = LAMB, lr=3e-4, lamb_clamp=10.0, stable_ratio=0.02
• Schedule = cosine, lr_min=1e-5
• Batch = 66
• Grad clip = 1.0

Size Check

• Embed tied: 729×729×2B = 1.06MB ✓
• 9 blocks × (attn + ffn) = 6.5M params × 2B (GF16) = 13MB ✓
• Total: ~14.1MB < 16MB ✓

Acceptance Criteria

• All dims ∈ {3^k}
• ctx % head_dim == 0 && is_power_of_3(ctx / head_dim) (EXP-012)
• vocab=729 padded (bytes 0-255 active, 256-728 logits = -inf)
• GF16 quantization (4-bit)
• Model size < 16MB

Refs

• EXP-001: Trinity 3³ physics
• EXP-010: Ternary block alignment
• EXP-012: Square Attention theorem
• Issue Phase 5: Native Sessions for tri-api + tri-bot #64: simple_backward fix

Commit

"feat(config): adopt Trinity config (hidden=243, vocab=729-padded, GF16)"
Refs: EXP-001, EXP-010, EXP-012, #64