The comments in `_run_basic_gradient` and `test_adstack_basic_gradient_f64` incorrectly claim that a 1e-12 abs floor "would miss an f32-narrowing regression" — in reality, f32 vs f64 differences for `0.95**n_iter` are ~1e-7 to 1e-8 (4-5 orders of magnitude *larger* than 1e-12), so the floor would still catch f32-narrowing regressions. The correct justification for `abs_tol=0` is that without it the effective tolerance becomes ~1e-12 absolute (dominating the 1e-14 relative), which misses sub-f32-