[DylanCosto]

Prebuilt indexes now supply the k-mer size for derived chaining penalties as well as seeding. A k=15 index produces the same output as mapping directly with -k15 -w10, rather than silently retaining the preset's k=19 penalties and changing alignments.

Split-index merging now reads every query file, preserves input read groups and filters copied comments consistently. The single-part FASTA fast path also honors --split-prefix. Multi-part SAM/BAM output automatically uses the existing merge path, producing a complete reference dictionary instead of turning later-contig alignments into unmapped records or failing BAM conversion. Automatically generated temporary files are cleaned up on success and failure; --split-prefix still controls their location. Merging requires regular query files because they are reread; stdin/pipes are rejected with an actionable message.

SAM/PAF output is explicitly flushed before success is reported. Writer failures also disconnect the input queue before joining the reader, preventing a full-disk/broken-output error from hanging the process.

Validation on x86-64 Linux:

• cargo test --release --locked --features jemalloc: 214 passed, 5 benchmark/stress tests ignored; existing golden output hashes unchanged. Re-ran all 13 CLI regression tests after final edits.
• cargo clippy --release --locked --features jemalloc --all-targets passed.
• All 13 Python validation tests and both archived-evidence integrity checks passed.
• Focused coverage only: extend the existing index-equivalence regression for a mismatched preset k; add one split-output regression covering multiple queries, reference formats and BAM headers; add one Linux output-error regression covering final flush errors and cancellation without a hang.

No full-genome variant-calling benchmark or native ARM execution was run locally. Default matching-preset golden outputs are unchanged; the alignment change restores equivalence to direct FASTA mapping for indexes with a different k.

Performance benchmark — 2026-10-29

Compared main b44551d with this PR's fd23a9d on an i9-12900 under x86-64 WSL/Linux, Rust 1.99.1, release + jemalloc, warm-cache Linux ext4 input/output. Runs were sequential with alternating version order and one warmup. ONT input: 40,960 reads / 254.4 Mb (256 renamed copies of the bundled synthetic fixture). HiFi input: 8,192 reads / 112.4 Mb (1,024 renamed copies of the eight real HG002 fixture reads). References are deliberately small (0.66 Mb / 84 kb).

Case, 8 threads	Main median	PR median	Measured wall-time change
Single-part ONT SAM, 3 paired trials	4.43 s	4.37 s	-1.1%
Single-part ONT sorted BAM, 5-pair follow-up	5.08 s	5.33 s	+5.0%
Single-part HiFi SAM, 5-pair follow-up	3.33 s	3.33 s	-0.05%
Automatic multi-part SAM, 3 paired trials	7.03 s (invalid header)	33.87 s (valid)	4.82× as long
Explicit --split-prefix SAM, 3 paired trials	34.09 s	34.45 s	+1.1%

Confirmed performance caveat: automatic multi-part SAM/BAM now selects the existing serial split-mapping/merge path. This fixes the invalid output but loses parallel mapping throughput. Peak RSS for multi-part SAM falls from about 1,100 MiB to 74 MiB. Existing explicit split mode has essentially the same cost before and after. Preserving multi-part parallel throughput needs a follow-up change to that mapping pass before merging if this regression is unacceptable.

The old automatic multi-part SAM has 25,088 records referencing contigs absent from its header; the PR has all three contigs declared and 40,960 distinct primary read names. Old automatic multi-part BAM fails with reference not in header; the PR succeeds in 34.78 s median, so there is no successful old BAM runtime to compare.

Ordinary SAM/HiFi and sorted-BAM logical outputs are byte-identical after removing @PG. The sorted-BAM timing is noisy: the initial three-pair pass showed +12.5%, while the five-pair follow-up showed +5.0% with overlapping ranges (main 4.83–5.33 s; PR 4.73–5.68 s). No precise ordinary-run speed guarantee follows from these small samples. Old/new raw index files are byte-identical; the follow-up uses the same index file for both executables.

This is a code-path throughput/output-equivalence benchmark, not a whole-genome throughput or variant-calling accuracy benchmark.