feat(reads): raw FASTQ quality control

[BenjaminDEMAILLE]

Part of #163.

Stacked on #159.

Adds rustqc reads: quality control on raw FASTQ, before alignment. It sits apart from rna, dna and protein because the question is different. Those ask what the alignment says; this asks whether the reads are usable at all, whatever the library was for.

Writes a seqkit stats-compatible table and a FastQC-compatible fastqc_data.txt, so existing parsers keep working. Records stream through the accumulator and are never all held at once.

Parity

Full parity with seqkit 2.13.0 on all nineteen columns, and with FastQC 0.12.1 on four modules: per base sequence quality, per sequence quality scores, per base sequence content, and per base N content.

Four conventions that would have passed a less careful test

N50_num counts distinct lengths, not sequences. Three reads of 10, 10 and 3 give 1, not 2, because the two tens are one length. A file whose lengths are all distinct hides this completely, which is exactly how it slipped past the protein fixture in #158 where every length happened to be unique. That implementation is corrected here too, and the discriminating case now has its own test.

AvgQual averages error probabilities, not Phred scores. On this fixture the arithmetic mean is 34.00 while seqkit reports 25.60, because averaging in probability space is dominated by the worst bases. Reporting the arithmetic mean would flatter every run.

FastQC's binned rows average their positions rather than pooling their bases, for the quantiles as well as the mean. That is why a bin's tenth percentile can read 35.2 when every individual position's is a whole number. The two definitions agree until reads start running out, so the difference first appears at the 40 to 44 bin, in the seventh decimal for the mean and outright for the percentiles.

FastQC excludes N from the per-base composition denominator, so the four percentages sum to 100 even where the instrument called nothing. Including them shifts every percentage by the N rate, which on this fixture moves the fourth decimal.

One module written but not asserted

Per sequence GC content. FastQC does not report the plain distribution of per-read GC percentage: it spreads each read's contribution across neighbouring bins so a coarse discrete distribution plots smoothly, which makes its counts fractional and puts reads in bins none actually occupies. RustQC reports the plain rounded distribution, which is a different and defensible figure, so asserting equality would be asserting the wrong thing. It is called out in the code, the test and the report.

Module verdicts are written as pass uniformly rather than reproducing FastQC's pass/warn/fail thresholds, which are judgement rather than data. That is the one place the file is not a drop-in replacement.

Fixture

test.umi_1.fastq.gz from nf-core/test-datasets, 302 kB, 7890 reads of 41 to 151 bases with a few N calls, which is enough to exercise the ragged-length and N-content paths that a uniform file would not.

Tests

416 green, up from 397.

🤖 Generated with Claude Code