fix: score each sequence independently in compute_pwm

[aviezerl]

Problem

Reported by Tamar: compute_pwm() gives a sequence a different score depending on the other sequences in the batch when sequences have different lengths.

compute_pwm(c("AACCAGCT", "AACCAGCTTCATTCATGCAAATTAGGGGACCACAGGAGTCCC"),
            pssm = motif_pssms$oct4, func = "max", bidirect = TRUE)
# 2nd sequence -> -14.62

compute_pwm(c("AACCAGCTTCATTCATGCAAATTAGGGGACCACAGGAGTCCC"),
            pssm = motif_pssms$oct4, func = "max", bidirect = TRUE)
# same sequence alone -> -5.58

Root cause

The motif-scan window was capped to nchar(sequences[1]) for the whole batch (passed to the C++ spat_max, and the flat spatial bin was sized the same way). A sequence longer than the first was therefore only scanned over its first nchar(sequences[1]) starting positions, missing motif hits further along. A shorter first sequence (8 bp) truncated the 42 bp Oct4 sequence to its first 8 positions.

The C++ core (shared with misha's DnaPSSM) already clamps the scan per sequence to end - motif_len; the cap was purely an artifact of the R wrapper deriving one window from sequences[1].

Fix

Use max(nchar(sequences)) for both the scan range and the flat spatial bin. Each sequence is then scanned to its own end, independently of its batch companions. Equal-length batches (the regression / screen_kmers path) are byte-identical since max == first.

The sibling functions (compute_local_pwm, screen_local_pwm, mask_sequences_by_pwm) carry the same nchar(sequences[1]) line but pass motif-length windows to the scan, so the range is inert there - left unchanged.

Test

Added a regression test in test-compute_pwm.R asserting a sequence's score is batch-position-independent (both max and logSumExp). Fails before the fix, passes after; existing PWM tests unchanged.

https://claude.ai/code/session_01PK3qefBGDoBwd9w5226FEq