perf(float32): make the data path float32 at its source

[daharoni]

What

Phase 1 of moving minian off float64: switch the data path to float32 at its source. Because dask/xarray/numpy infer output dtype from their inputs, flipping the source is enough for the bulk of the pipeline to follow.

Two changes:

1. io.load_videos default np.float64 -> np.float32. The footage is 8-bit, so every loaded value is represented exactly in float32 (no precision loss at load), while halving memory and on-disk size. Callers that pass dtype explicitly - including the demo notebook (uint8) and every test - are unaffected.
2. Pipeline notebook: Y.astype(float) -> Y.astype(np.float32) at the Y_fm_chk save. This was the single spot promoting the demo movie to float64 - the ~69 GB-vs-34 GB cube behind the slow save/rechunk in feat(pipeline-nb): expose rechunker mem_limit as a tunable parameter #342. Everything downstream (Y_fm_chk, Y_hw_chk, seed init, CNMF prep) is now float32; all other notebook casts were already float32.

Scope (deliberately limited)

This PR only changes the data source. The CNMF gufunc output_dtypes in cnmf.py still emit float64, and sklearn LassoLars / cvxpy solvers upcast internally, so numeric results shift only as much as a float32 input warrants. Tightening those (gufunc output_dtypes, internal np.zeros dtypes, double accumulators) is a validated follow-up (Phase 2).

Validation

test_pipeline_notebook runs this notebook end-to-end in CI and checks cell count (286 ±10) and the C/S/A sums (rel 5e-2 / 1e-1 / 5e-2) - tolerances built for exactly this kind of numeric drift. That's the gate for the notebook change. The io default change is covered by test_pre_processing (which passes dtype explicitly, so behavior is unchanged there).

Note

Targets master, which is the v2 development branch, so the load_videos default dtype change (a public-API default change suited to a major version) is appropriate here. It's also backward-compatible for 8-bit-sourced data, since those values are exact in float32.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!