The line-count computation used an exact LCS with a 1M-comparison
cap, so any full-content diff over ~1000 lines silently dropped the
+N/-M stats and large edited files showed no change counts. Replaced
the exact LCS with a linear-time multiset comparison — stats are now
always computed; only identical lines moved between positions
miscount, which is acceptable for display stats.
Co-authored-by: 钉萁 <dingqi.jww@alibaba-inc.com>