fix: symmetric N/* handling across strands in compute_pwm - #40

Problem

Reported by a user via Tamar: how does compute_pwm treat N, and why does the forward branch use get_avg_log_prob() while the reverse branch uses log(0.25)?

Confirmed - it's a real strand asymmetry. In src/DnaPSSM.cpp, the two functions compute_pwm uses treat N/* differently on the two strands:

• Forward: p->get_avg_log_prob() - the column's mean log-probability, (log pA + log pC + log pG + log pT)/4.
• Reverse: c_log_quarter = log(0.25), a flat constant.

Since a normalized column sums to 1, these are equal only at a uniform column; at informative columns mean(log p) < log(0.25), so the forward strand penalizes an N more than the reverse strand.

Impact

For a bidirectional motif, a window and its reverse-complement must score identically. They do - until an N lands on an informative position:

no-N:   |seq - rc| = 0.0
with-N: seq = -8.24,  rc = -6.64,   |diff| = 1.60 nats

Only affects sequences containing N/* with bidirect = TRUE.

Fix

Use get_avg_log_prob() on both strands in integrate_energy and integrate_energy_max. That restores strand symmetry and matches the convention the symmetric likelihood routines (max_like_match, integrate_like_seg) already use on both strands - so avg is the codebase's intended convention and the reverse c_log_quarter was the outlier.

Out of scope (left as-is): integrate_like / update_like_vec carry a different asymmetry (forward log(0.25), reverse avg) but are learning/likelihood internals, not the compute_pwm energy path. like_thresh_match is intentionally symmetric (log(0.25) on both). The identical fix is being applied to misha's DnaPSSM.cpp (same shared code) in a companion PR.

Test

Added a strand-symmetry test: a motif-length window with an N on an informative column must score the same as its reverse-complement, for both max and logSumExp. Fails before, passes after. Full test suite green.

https://claude.ai/code/session_01PK3qefBGDoBwd9w5226FEq