E2: V4 characterization + forward (DeepSeek-V4-Flash-4bit)

[tlkahn]

V4 characterization + forward (MODEL_DEEPSEEK_V4)

Inspect Flash weights before implementing or aliasing. Never assume product name implies V3/V3.2 architecture match.

Spec: doc/plans/engine-inference-core.md § E2

Part of: #105 (E2) · Epic: #56
Depends on: V3 forward path (likely base); family split already has MODEL_DEEPSEEK_V4

Tasks

• Inspect mlx-community/DeepSeek-V4-Flash-4bit config.json + weight names
• Document characterization outcome on this issue (match V3, match V3.2, or distinct)
• If distinct: implement MODEL_DEEPSEEK_V4 forward differences
• If true match: alias explicitly in code with a comment citing the characterization evidence - do not silent-map
• Support gate + reject strings stay accurate for deepseek_v4
• Smoke generate on Flash-4bit once path exists

Target checkpoint

mlx-community/DeepSeek-V4-Flash-4bit (safetensors)