fix: WFPTNormalDrift normalisation when t0 > 0

[cs-hall]

Summary

• WFPTNormalDrift.log_prob used v_loc**2 * RT where Navarro-Fuss / Blurton want decision time RT - t0, so the density integrated to less than 1 whenever t0 > 0 (e.g. 0.30 at v=2, sv=1, t0=0.8). Same bug WFPT had before b13135d.
• Test grid helper get_choice_RTs starts at dt instead of 0, so the integration tests no longer pass RT = 0 (out of support) and stop raising out-of-support warnings.
• Now computed in log space (no log(exp(...)) round trip), and -inf for RT <= t0 instead of NaN when sv² (RT - t0) + 1 < 0.

Tests

New (committed failing first, dc72d7c):

• test_wfpt_normal_drift_integrate_to_one: integrates to 1 for t0 in {0, 0.5, 0.8}, v_scale up to 1.5, negative drift.
• test_wfpt_normal_drift_matches_wfpt_as_v_scale_vanishes: reduces to WFPT as v_scale -> 0.

Before the fix: 4 failed (all t0 > 0 cases), 2 passed (t0 = 0).

Full suite after:

79 passed, 1 warning in 52.25s

The remaining warning is pytest declining to collect the TestDist namedtuple as a test class.

🤖 Generated with Claude Code