[traviswheeler]

Rebuilt on dev (2026-10-06). The three earlier commits were based on
main; this branch is now one commit on top of dev 503e3ca, and the PR
base is dev.

Two changes to how sufr create writes its output.

Positioned parallel writes. Write the header, text, suffix array and
LCP array into the output file at precomputed offsets from several
threads (64 MiB chunks) instead of streaming them through one BufWriter.
write() drains each partition's scratch buffer in parallel and still
patches the LCP at each partition boundary. The write helper sits next
to the positional read helper in file_access.rs with the same
cfg(unix)/cfg(windows) split. The whole workspace type-checks for
x86_64-pc-windows-msvc, aarch64-apple-darwin and x86_64-apple-darwin
(with needletail's compression features off, which is otherwise what
stops a cross-compile here); nobody has run it on Windows.

LCP array optional, off by default. Adds write_lcp to
SufrBuilderArgs and sufr create --write-lcp. Searching never reads
the LCP array. list --show-lcp fails on a file without one and says to
rebuild with --write-lcp. For a query-time max query length shorter
than the build-time value, the subsampling step read the LCP array; on a
file without one it now compares each suffix with its predecessor in the
text, streaming the suffix array in 4M-element blocks. A file without an
LCP array stores 0 as the LCP offset; older files, which always have one,
open as before. summarize reports whether a file has an LCP array.
FileAccessIter refills 16 MiB at a time instead of 1 GiB.

Timing on 300 MB random DNA, mask 11011, 16 threads, 16 partitions, two
runs each: dev 8.1 s; this branch 6.5 s with the LCP array, 6.0 s
without. Peak memory 4.5-4.7 GB in all three. The file is 4 bytes per
suffix smaller without the LCP array.

Tests: dev's tests pass with write_lcp: true. A new test checks that
subsampling with and without the LCP array gives the same result in
seed-mask and MQL mode and matches an independent definition. With
--write-lcp the output is byte-identical to dev's on data/inputs and
the 300 MB file across default, -n 3, --seed-mask 11011,
--max-query-len 20 and --dna builds, except --max-query-len 20
without --dna on the 300 MB file, where two runs of unmodified dev
also differ from each other at the same byte.

Found while testing, present on unmodified dev, not addressed here:
locate -m 3 and extract -m 3 panic at an assertion in
file_access.rs when a query is longer than the max query length.

Implemented with assistance from Claude (Fable 5.1)

[jebrosen]

This looks generally good, with two caveats here I noticed that we might still deem acceptable. Both are adjacent to some future cleanups I have in mind anyway.

[jebrosen]

LGTM! I validated identical output for the --write-lcp enabled path on a few sample files.