fix: error on unequal-length sequences in calc_seq_pwm / extract_pwm

[aviezerl]

Problem

Follow-up to the compute_pwm variable-length fix (#38). Auditing the rest of the PWM family for the same class of bug turned up a second one in the MotifDB scoring path.

calc_seq_pwm() (and extract_pwm(), which wraps it) build a single rectangular one-hot matrix via do.call(rbind, ...) in seqs_to_onehot. With sequences of different lengths, rbind recycles the shorter rows up to the longest length, so the shorter sequences get wrong, batch-dependent scores - and it surfaces only as an easy-to-miss rbind warning ("number of columns of result is not a multiple of vector length"), not an error:

# shorter sequence scored in a batch with a longer one vs. alone
SHORT seq in batch: -11.16   |   alone: -14.59   ->  wrong

Fix

These functions are inherently equal-length (one rectangular matrix multiply against all motifs), and compute_pwm() already handles variable lengths correctly. So fail loudly instead of returning garbage:

if (length(unique(nchar(sequences))) > 1) {
    cli_abort("All sequences must have the same length for extract_pwm(). For sequences of different lengths, use compute_pwm().")
}

Mirrors the existing guard in compute_local_pwm() (which already errors when return_list = FALSE and lengths differ).

Not changed

compute_pwm (fixed in #38), compute_local_pwm, screen_local_pwm, mask_sequences_by_pwm, and gextract.local_pwm were all verified to be batch-independent on variable-length input. The misha package was also audited and is immune to this bug class (its scan range is always a per-PSSM constant clamped to each target's own length).

Test

Added a test asserting calc_seq_pwm errors on unequal lengths and still works on equal lengths.

https://claude.ai/code/session_01PK3qefBGDoBwd9w5226FEq