Add array storage to ActivationTracker scalar_history

[adamimos]

Background

Currently ActivationTracker._scalar_history stores individual scalar values:

{
    "pca/layer_0_cumvar_1": [(step, value), ...],
    "pca/layer_0_cumvar_2": [(step, value), ...],
    ...
}

For visualization purposes (e.g., CEV over training plots), it's more convenient to store the full cumulative variance array per step rather than individual components.

Proposed Enhancement

Add a parallel _array_history storage:

{
    "pca/layer_0_cumvar": [(step, np.array([...])), ...],
}

Or extend the existing API:

tracker.log_array(key, step, array)
tracker.get_array_history(pattern) -> dict[str, list[tuple[int, np.ndarray]]]

Use Case

This would allow visualization functions to directly receive the cumulative variance curve without needing to reassemble it from individual scalar components.

Workaround

For now, we're maintaining a separate in-memory dict in the training loop:

cve_history[layer].append({"step": step, "cumvar": cumvar_array})

---

🤖 Generated with Claude Code