Every assembly below was produced by the native-Windows spades-*.exe (fully static, no
WSL/Docker/VM). Accuracy is alignment-free: 31-mer identity (every assembled 31-mer must
appear in the reference — catches any single-base, port-induced drift) and genome fraction
(fraction of reference 31-mers recovered). Reproduce with
../scripts/validation/validate_ecoli.py.
ENA ERR1473771 (Illumina HiSeq 2500, 2×100 bp, ~85×; a real K-12 MG1655-background
isolate), run through the default pipeline (BayesHammer error correction → multi-k assembly),
contigs compared to NC_000913.3 — ecoli_realreads_metrics.json:
| Metric | Value |
|---|---|
| Contigs (≥ 1 kb) | 185 (77) |
| Total assembly | 4,564,471 bp |
| Largest contig | 356,696 bp |
| N50 / L50 | 117,598 bp / 12 |
| Genome fraction | 99.21 % |
| 31-mer identity | 99.73 % |
| Wall-clock | ~3.6 min (-t 8 -m 24, incl. error correction) |
The residual ~0.27 % off-reference is real: a laboratory-evolved MG1655 derivative carries genuine strain variation vs the canonical reference, plus real sequencing error — not port drift.
The first SRA accession we tried was mislabeled/contaminated (0 % 31-mer match to MG1655), so the validation now verifies a candidate's reads against the reference before assembling.
145 contigs, N50 175.9 kb, largest 327.1 kb, 99.997 % identity / 99.71 % genome fraction. The whole 4.64 Mb chromosome at ~100 % identity; fragmentation is the expected short-read break at the seven ~5 kb rrn operons / IS elements, not a port defect.
The full chromosome assembles into a single 580,060 bp contig at 100.0 % 31-mer identity /
99.99 % genome fraction. This is the case that used to crash spades-core at K21 on Windows — a
single LLP64 bug (a 32-bit -1ul sentinel), now fixed.
Single 48,457 bp contig, 99.967 % identity / 99.885 % genome fraction, GC 49.8 %.
A deliberately punishing 2×300 MiSeq set (89 % singleton k-mers) drives k-mer counting to
k=127 over a >2 GB on-disk k-mer file — the exact path the LFS fix repairs (previously a hard
stat(2) … value too large abort). It now runs through k=127 at normal memory (~5.5 GB):
All 13 executables are fully static — objdump -p shows only Windows system DLLs (KERNEL32,
ADVAPI32, UCRT api-ms-win-crt-*), no MinGW DLLs. See
../scripts/spades-patch/README.md for the port internals.