A survey of 291 open-weight language model repositories covering 280 distinct embedding tables reveals that 52.6% of untied gated and hybrid checkpoints (20 of 38) contain reachable input embeddings with norms below 10⁻⁶, whereas weight-tied models exhibit 0% occurrence. These near-zero rows emerge during pretraining when unseen rare or template tokens receive zero input gradients while AdamW weight decay shrinks untied embeddings; non-initial positions or a BOS token prevent this in standard Transformers. In gated sequence models (Mamba, Gated DeltaNet, xLSTM), a near-zero token amplifies input-embedding gradients by over 30 orders of magnitude (up to 35). At zero input, current-token readout factors (SiLU gates or linear queries) vanish, suppressing historical-state readout; residual RMS values then fall below the square root of the RMSNorm floor across consecutive layers, causing Pre-RMSNorm to amplify backward gradients by up to 10³ per layer. When a gradient entry exceeds 1.84×10¹⁹, FP32 global-norm clipping in PyTorch, Hugging Face Trainer, and FSDP2 overflows to infinity, zeroing the batch gradients. In Telugu continual pretraining of Nemotron-3-Nano-4B, this discards updates on ~20% of steps and leaves triggered rows frozen. Two remedies are proposed: GroupClip decouples input-embedding clipping using a 64-bit or rescaled norm, eliminating update loss and healing triggered rows in one step with negligible overhead; GateFloor (SiLU(z) + δ, δ ∈ [0.01, 0.1]) keeps the gate positive at zero input, reducing near-zero gradients by 28 orders of magnitude in experiments up to 370M parameters with comparable perplexity.