Add end-to-end tests for the encoding pipeline

[aadhavthaniga]

End-to-end tests for the encoding pipeline

Synthetic (tests/test_pipeline_synthetic.py, offline, ~1 min): plants a signal at FIR lags 2-3 in 10 vertices and checks recovery, the lag-specific kernel, FDR selection, FIR causality, no held-out leakage, determinism, and downsampler/FIR outputs against independent reference implementations.

Golden (tests/test_golden_lebel.py, opt-in): pins train_simple.py's wordrate results on the LeBel UTS03 assembly. Skips unless LITCODER_LEBEL_ASSEMBLY points at the file (download link in the README). Tolerances come from measured CPU/MPS/cross-library variance; see tests/fixtures/golden_lebel_uts03.json.

Code changes

• encoding/models/folding.py: fold shuffling used the unseeded global RNG, so every run fit a different model. Now seeded; the determinism and golden tests depend on this.
• train_simple.py: reads the assembly path from LITCODER_LEBEL_ASSEMBLY, falling back to the existing path.

Found along the way

• Non-legacy downsamplers (average, sum, last) misread split_indices and return ~n_words rows. Marked strict xfail.
• NaN features are silently zeroed while NaN targets pass through. Documented in a test.

Scope: the golden test uses wordrate (no model download, no downsampling); downsampling is covered by the synthetic tests.

Results: with the assembly, 29 passed, 3 xfailed. Without it, the 4 golden tests that need it skip.