[BenjaminDEMAILLE]

Part of #354 (test/parity_suite.sh on macOS).

Problem

On macOS, test/parity_suite.sh cannot produce one real comparison, and it does not say so.

• timed calls /usr/bin/time -f "%e %M". macOS's time has no -f, so every index build and alignment is reported as failed.
• The chr21 download checks "$(gunzip -c … | wc -l)" = 8000000. macOS's wc pads the count with spaces, so a complete 8,000,000-line file is reported as download incomplete and fetch stops.
• Every alignment passes --readFilesCommand zcat. On macOS, zcat is the compress one: it only reads .Z files. STAR and rustar-aligner both read 0 reads, exit 0, and parity_compare.py prints records 0/0, identical 0 … Log.final.out: identical for all 21 configurations.

The last one is the dangerous part. I ran the suite on macOS this way first: every configuration "matched".

Change

• timed runs the command through a few lines of Python (resource.getrusage(RUSAGE_CHILDREN)), and writes the same seconds KB file that parity_compare.py reads. ru_maxrss is in bytes on macOS and KB on Linux; the script converts.
• The wc -l count goes through tr -d " ".
• ZCAT=$(command -v gzcat || echo zcat): gzcat on macOS, zcat on Linux. It stays one word, because main still takes one word for --readFilesCommand (--readFilesCommand takes what STAR takes #353 changes that).
• parity_compare.py prints a WARNING when either tool's Log.final.out reports 0 input reads.
• The header says that Homebrew's STAR stops on geneInfo.tab with --quantMode or --soloType. From STAR's source, that comes from the build, not from the options: Transcriptome.cpp:13 reads trInfoDir from sjdbInsert.outDir, which is empty, while Genome_genomeLoad.cpp loads the index parameters into a separate P1. A Linux build of the same 2.7.11b source runs them. I did not dig further into the Homebrew build.

Measured

macOS (Apple silicon), Homebrew STAR 2.7.11b for alignment and a Linux (Docker, arm64) build of the 2.7.11b source for the quant and solo configurations, main at 32fbbc1:

	Before	After
fetch	stops at ERR188021_1: download incomplete	completes
y10k_se	0 / 0 records, "identical"	9,724 / 9,724 identical, 0 reads differ
peak RSS column	n/a	STAR 940 MB, rustar-aligner 168 MB
a run that read 0 reads	silent	WARNING: rustar-aligner read 0 reads; check --readFilesCommand (y10k_se)

With these changes the whole suite reproduced the README's comparison on macOS: yeast 0 differing reads in every configuration (1 in y100k_pe_attr), chr21 PE 3, fusion 24, solo default 36/36 files identical. The only new observation is in Log.final.out on yeast PE: 15,208 "too short" and 26 "other" for STAR against 15,209 and 25, which the README does not mention.

Tests

None added: this is test tooling. The timing helper was checked on a 50 MB allocation (61,792 KB reported) and on a failing command (exit code passed through), and bash -n passes.

Not checked

• Linux: the changes keep GNU time's output format and fall back to zcat, but I did not rerun the suite on Linux.
• Only the steps listed above were changed; macOS's bash 3.2 runs the script.

🤖 Generated with Claude Code