audit: 21 CUfunction conflicts — vaccine uses 15+ forward_child kernels ungraphed

docs/superpowers/audit/2026-04-18-cufunction-sharing-audit.md

@@ -0,0 +1,30 @@
+# CUfunction Sharing Audit (2026-04-18)
+
+## Summary
+- 104 CUfunction fields total
+- **21 active conflicts** (same CUfunction in graphed child + ungraphed op)
+- 3 disabled but latent conflicts
+- 4 already mitigated (separate CUmodule)
+- ~64 no conflict
+- ~12 dead fields
+
+## Critical Conflicts
+
+### Vaccine (15+ kernels): uses forward_child CUfunctions ungraphed
+c51_loss_kernel, c51_loss_reduce_kernel, c51_grad_kernel, mse_loss_kernel,
+mse_grad_kernel, fill_gamma_buf_kernel, scale_f32_kernel, saxpy_f32_kernel,
+curiosity_prepare_input_func, curiosity_bias_leaky_relu_func, curiosity_bias_mse_func,
+bn_tanh_concat_kernel, stochastic_depth_kernel, bn_tanh_backward_kernel,
+strided_accumulate_kernel
+
+### Causal intervention: uses CublasForward cloned kernels ungraphed
+vsn_kernel, kan_gate_combine_kernel, strided_scatter_kernel (clones from forward_child)
+
+### Cross-child sharing
+saxpy_f32_kernel: forward_child + aux_child + adam_grad_child (3-way)
+strided_accumulate_kernel: forward_child + aux_child
+
+### Eval path
+expected_q_kernel: aux_child + ungraphed (reduce_current_q_stats)
+
+## Fix: Move ALL into child graphs → zero ungraphed kernel launches