Additive quantization enables extreme LLM compression while preserving lookup-table inference with free-form codebooks. At 2 bits per weight, however, greedy sequential initialisation can cause large quality losses that wider assignment search and PV-tuning do not fully erase. We show that this failure reflects an interaction between representational pressure and sequential codebook coupling. We use 𝜌=N/K^M as a structural pressure measure and propose OA-EM, a hard generalised-EM initialisation procedure using Hessian-weighted Mahalanobis distance to minimise output-aware reconstruction error. Across four models from 3B to 14B parameters, OA-EM consistently lowers matched post-PV perplexity. Controlled experiments at a fixed bitrate show that 𝜌 alone does not explain sensitivity to initialisation; the number and size of the additive codebooks also matter. A Euclidean-EM ablation and curvature diagnostics show that the gain is not explained by EM refinement alone: OA-EM reduces error in high-curvature directions. On Llama 3.2 3B, OA-EM also improves the measured quality-compute trade-off. Overall, the final quantized solution remains persistently dependent on initial codebook geometry. Code: https://github.com/kenno94-IK/aqlm-oaem.