ProCreations/repro-grok-grokking-ridge-regression

Add strengthened ReLU sweep

Claim 5: rerun the ReLU weight-decay sweep at 8 seeds and a doubled step budget; fitted exponent moves to -1.0879, within 8.6% of the exact linear-ridge value

Add exact ridge trajectory engine and ReLU grokking sweeps

Add Theorem 4.2 realizable-teacher experiments, the sample-size half of Figure 2, and two-layer ReLU sweeps for Figures 3-4

Update logbook: Repro - To Grok Grokking: Provable Grokking in Ridge Regression

initial commit