⚡ Bolt: optimize runica training loops

.jules/bolt.md

@@ -0,0 +1,3 @@
+## 2025-05-15 - np.errstate and Indentation in runica training loops
+**Learning:** Moving `np.errstate` context managers outside of high-frequency iterative loops (like the ICA training blocks) and replacing internal helper wrappers (like `_matmul`) with native operators (like `@`) can yield a measurable (~10%) performance gain by eliminating function call and context entry/exit overhead. However, when wrapping large blocks of code with a new context manager, it's critical to verify the indentation of all nested logical blocks (e.g., restart conditions) to avoid bypassing core logic or breaking functionality.
+**Action:** When refactoring iterative numerical loops, prioritize moving context managers like `np.errstate` outside the loop, use native matrix operators, and perform a thorough indentation check on all nested `if/else` and `for` blocks.