4-bit Recurrent-State Write-Back Raises GRU τ1 Error 70x in Fixed Fluorescence Lifetime Model

The intervention isolates recurrent-state write-back by holding weights fixed and quantizing only the returned hidden state at each timestep. In the compact GRU for fluorescence lifetime imaging, repeated sub-threshold updates accumulate without altering stored state, decoupling network proposals from actual dynamics. This produces systematic underestimation of both short and long lifetime components under high-noise conditions typical of quantitative molecular imaging. Error feedback, residual memory, and direction memory restore most accuracy by carrying suppressed update information across steps without retraining. Precision sweeps reveal that raising state bits can degrade performance when the write threshold remains mismatched to update magnitude. The same 4-bit intervention applied to an independently trained LSTM reproduces the failure mode, with greater degradation in cell state than hidden state. Standard quantization literature focuses on weight and activation bit-widths during training or forward passes; recurrent write-back interfaces receive little scrutiny despite their direct effect on temporal memory. The paper demonstrates that state-storage compatibility must be treated as an explicit design constraint for low-precision recurrent inference on edge hardware, where memory bandwidth dominates energy cost.