Implement experiment 3: first thought

[Ved235]

Experiment 3 was previously only represented indirectly through Probe 4 in Experiment 6. This PR implements it as a standalone experiment with its own training, probe, control, and results pipeline.

The new experiment keeps the same CfC substrate and training setup as Experiment 6 so the results remain comparable, but narrows the scope to the first-thought question: is the observer’s first prediction better than the prediction it reaches after repeated self-feedback? The PR also has implemented K sweep and stability analysis based on prediction drift across passes.

[kk-genesis]

This is a fascinating experiment direction. The question — whether the observer's first prediction is better than the one reached after repeated self-feedback — gets at something fundamental about how recursive self-observation behaves in practice.

I'm curious about the K sweep and stability analysis. What range of K (number of self-feedback passes) did you explore? Was there a regime where additional passes consistently degraded prediction quality (i.e., self-feedback became self-noise past some threshold)? Or did the CfC substrate tend to converge to a stable attractor regardless of K?

Also related: in Experiment 6, Probe 4 was the indirect measure of this effect. Does the standalone Experiment 3 replicate the direction of Probe 4's signal, or did isolating it reveal something different about the dynamics of first-vs-iterated prediction?