easy-multimercluster crashes with `munmap_chunk(): invalid pointer` on mixed-chain-count inputs (regression after v10)

[bviggiano]

Summary

The Linux x86_64 AVX2 build of foldseek from https://mmseqs.com/foldseek/foldseek-linux-avx2.tar.gz (continuous master HEAD, BuildID 8a31a73a..., dated 2026-03-10, commit message "Set chain tm default paramter to 0") heap-corrupts and aborts inside the internal clust step of easy-multimercluster when given certain mixed-chain-count input sets:

[================================================================ munmap_chunk(): invalid pointer
Aborted (core dumped)
Error: Clustering died
Error: Multimercluster died

The same inputs cluster cleanly with the v10 tagged release (10-941cd33, Jan 2025), so this is a regression introduced somewhere in the master commits between Jan 2025 and March 2026.

Trigger

Any input set with 3 or more multimers that contains a 3-chain assembly mixed with non-3-chain assemblies. 2-input sets always work; 3+-sets with all-same chain counts work; 3+-sets with a 3-chain assembly mixed with 2-chain or 4-chain crash.

Input set	Chain counts	Result
1TIM + 8TIM	2, 2	pass
3BIK + 1TIM	3, 2	pass (2-set)
4HHB + 2HHB	4, 4	pass
3BIK + 1TIM + 8TIM	3, 2, 2	CRASH
3BIK + 4HHB + 2HHB	3, 4, 4	CRASH
4HHB + 2HHB + 1TIM	4, 4, 2	pass
4HHB + 1TIM + 8TIM	4, 2, 2	pass
3BIK + 3CMP + 3SE8	3, 3, 3	pass (homogeneous)
3BIK (chain C stripped) + 1TIM + 8TIM	2, 2, 2	pass
3BIK (HETATMs stripped) + 1TIM + 8TIM	3, 2, 2	crash

Stripping HETATMs from 3BIK doesn't help; stripping its 3rd chain (making it 2-chain) does. So the trigger is the chain-count combination, not the file content.

Reproduction

mkdir -p /tmp/fs_repro/structures /tmp/fs_repro/fs_tmp
for pdb in 3BIK 1TIM 8TIM; do
  curl -sS "https://files.rcsb.org/download/${pdb}.pdb" \
    -o "/tmp/fs_repro/structures/${pdb}.pdb"
done

foldseek easy-multimercluster \
  /tmp/fs_repro/structures \
  /tmp/fs_repro/multimercluster \
  /tmp/fs_repro/fs_tmp \
  --multimer-tm-threshold 0.65 \
  --chain-tm-threshold 0.001 \
  --interface-lddt-threshold 0.5 \
  --alignment-type 2 \
  --tmscore-threshold 0.0 \
  --lddt-threshold 0.0 \
  --threads 2

End of stderr:

Add missing connections
[================================================================ munmap_chunk(): invalid pointer
Aborted (core dumped)
Error: Clustering died
Error: Multimercluster died

Bisection

• v10 (10-941cd33, Jan 2025) — https://github.com/steineggerlab/foldseek/releases/download/10-941cd33/foldseek-linux-avx2.tar.gz — handles all the above input sets cleanly.
• master HEAD (BuildID 8a31a73a..., 2026-03-10) — https://mmseqs.com/foldseek/foldseek-linux-avx2.tar.gz — crashes as shown.

The regression was introduced somewhere in master between the v10 tag and 2026-03-10. The crash signature points at the clust "Add missing connections" phase.

Environment

• OS: Linux 5.15.0-176-generic, x86_64
• Inputs: vanilla RCSB PDB downloads, no preprocessing
• Reproduction is deterministic (fires on every invocation with the same inputs)