Gemma 4 [6/6]: validation — HF logit/argmax parity + perf + memory on Orin

[ai-hpc]

Part of #88. Depends on [4/6], [5/6].

Scope

• Greedy-decode logit/argmax parity vs HF transformers Gemma 4 E2B on a fixed prompt set (top-1 token match; bounded logit drift).
• Coherence check (interactive generation).
• tok/s (prefill + decode) and peak memory (model + KV + PLE table) within the 8 GB budget shared with voice.

Acceptance

• Top-1 parity vs HF on the prompt set; documented perf/memory; ready to port the math to TensorRT-Edge-LLM#72.