Learning is robust, seeing it is hard: attractor dynamics gate the readout of synaptic memory in a whole-brain spiking connectome model

No published work demonstrates reproducible associative learning inside a full connectome-constrained spiking model of the Drosophila brain; independent community attempts consistently report negative results. We show that the obstacle is not the learning rule but the readout. In the MaleCNS connectome (166,700 neurons, brain+VNC), a host-side dopaminergic depression of Kenyon-cell-to-MBON synapses writes robust, odour-specific synaptic memory after a single pairing (N=1), with consolidation that survives interference (protected-fraction ratio 1.737, 7/8 seeds, p=0.0039), a cross-seed depression-pattern cosine of 0.880, and an exactly zero no-reward control (8/8). A pre-registered ablation shows that type-specific spike-frequency adaptation is not required to write this memory (weight-level addressing is comparable within seed variability; 2.7x more edges are updated without it) but is required to see it: on the activity readout the same learning is dominated by attractor-basin switching (SPEC scatter 0.28-10.12). Even after physiological stabilization, the learned signal remains gated downstream: it decodes from synaptic weights (0.71-0.78) and from the postsynaptic drive (0.97-0.99) but only weakly from spikes (0.31-0.64), and seven candidate release mechanisms are negated with pre-registered controls. Degree-preserving null models abolish learning, working memory and benchmark performance, proving the effects are topologically constrained. The canon does not transfer to the FlyWire v783 connectome (connectome-specificity boundary). The framework parsimoniously explains the community's learning failures and defines the physiology required to read out synaptic memory in whole-brain models. Project channel: «Правила игры» — https://t.me/law_of_the_game