Dependency survey (Aug 2026): umbrella for the crate evaluations

[BenjaminDEMAILLE]

Umbrella for the third-party crate survey run in August 2026. Each sub-issue is one dependency
question in the form CONTRIBUTING.md:58 asks for: raised in an issue before any PR, with the
reasoning written down.

Every sub-issue is expected to end in a written outcome, either "adopted, here is the
justification" or "declined, here is why", so that the same crate does not get re-surveyed in six
months. There is currently no place to write that down, which is itself one of the sub-issues.

Already covered elsewhere, deliberately not duplicated here

• Dependency discussion: libsais (vendored C via cc) for suffix-array construction #162 — libsais (vendored C via cc) for suffix-array construction
• add praseq as an option for input #95 — paraseq as an input option
• Parallel inputs ({C,V,,}BQ or parallel readers for fastq) #97 — parallel inputs / parallel FASTQ readers
• New Dependency: hyalite for CR4 #197 — hyalite for CR4

Already in the tree, so not re-proposed

The survey turned up several crates that this project already uses, which is a good sign about the
existing dependency choices: memmap2, rustc-hash, libdeflater, tempfile, assert_cmd,
mimalloc (with an explicit rationale in Cargo.toml), flate2 on the zlib-rs backend, bstr,
caps-sa.

Sub-issues

#	Subject	Blocked by
#207	noodles 0.113 → 0.115, noodles-bgzf 0.49 → 0.51, plus the multithreaded BGZF writer question	—
#202	sufr / libsufr vs the caps-sa incumbent (sibling of #162)	—
#203	The rust-seq crates: packed-seq, simd-minimizers, seq-hash	#204
#204	A benchmark harness (divan or criterion) — enabler for everything below	—
#205	Portable SIMD crate vs the hand-rolled intrinsics in simd_scan.rs	#204
#206	--limitBAMsortRAM aborts instead of spilling: external sort	—
#208	Interval-overlap crates	#204
#209	DEPENDENCIES.md recording accepted and declined crates	—
#210	Audit: does DashMap/HashMap iteration order reach any output?	—
#218	Read other compressed input formats (bz2 / zstd / xz): niffler	—

Standing constraints for all of them

1. Faithfulness to STAR is the product. A crate that changes which alignments are found, or the
   order in which anything observable is emitted, is out regardless of its benchmarks.
2. Determinism across thread counts. Output must not depend on --runThreadN, on hash seeds,
   or on an unspecified iteration order.
3. Five platforms including Windows. A non-Rust dependency is a supply-chain and maintenance
   decision, per CONTRIBUTING.md.
4. Measure before optimising. Several sub-issues are explicitly blocked on the benchmark
   harness, because there is no benches/ directory today.