test(split): add edge-case tests for val_frac=0, n=1, n=2, and invalid fracs

[protosphinx]

Why

case_chrono_split supports val_frac=0 (documented as valid by the
[0, 1) bound check), a train_frac=0 guard, and a
train_frac + val_frac >= 1 guard. None of these paths had a test. The
function also has non-obvious integer-flooring behavior for small n
(n=1 always goes entirely to test; n=2 produces an empty val partition)
that is worth locking down explicitly so future refactors don't
silently regress the documented boundary.

What

• test_split_val_frac_zero_gives_two_way_partition: verifies that
  val_frac=0.0 is accepted and produces an empty val list.
• test_split_single_case_goes_entirely_to_test: n=1 floors both
  n_train and n_val to 0, so the sole case lands in test.
• test_split_two_cases_val_rounds_to_zero: n=2 with default fracs
  gives int(2*0.1)=0 val; confirms train/test boundary.
• test_split_train_frac_zero_raises: asserts train_frac=0 triggers
  the (0, 1) open-interval ValueError.
• test_split_fracs_sum_to_one_raises: asserts
  train_frac + val_frac == 1.0 triggers the >= 1 ValueError.

Tests

• All 5 additions are in tests/test_split.py.
• They exercise pm_bench/split.py directly with no mocking.
• Pure CPython, no network, no filesystem.

Self-merge gate

• all CI checks pass
• LOC delta < 250 (added + removed)
• no public-API surface change
• no runtime-dependency additions
• no workflow file changes
• tests added or extended

---

Generated by Claude Code