[jayhesselberth]

The reorder (--max / --order any) path called serialize_block(&all), re-serializing the whole input to FASTQ text only so the plain gate candidate could parse it back, which held a second input-sized buffer. This cuts the candidate directly from the buffered RawBlock + offsets. This is a memory win only. Refs #303.

Changes

• block.rs: new BlockSource enum (Parsed | Buffered) so write_plain_layout builds blocks from either source; the two compress.rs callers pass Parsed.
• parse.rs: new block_ranges_from_lens; block_ranges delegates to it.
• Platform is now resolved with resolve_platform_block on buffered reads (as the clustered path already did) instead of resolve_platform_buf on re-serialized text. A new test asserts they agree on a mixed-header fixture. An exotic header (e.g. trailing tab) could in theory detect differently; none found.
• No format change.

Evidence (SRR453566_1 GAIIx, 1.84 GB)

• Archives byte-identical to main (e8dd3d3): --max at 1/8/32 threads and --order any at 8/32 (180,584,067 B each).
• Peak RSS --max: 9.67 -> 8.18 GB (1t), 10.86 -> 9.28 (8t), 11.27 -> 9.73 (32t). --order any: 9.95 -> 8.18 (8t), 10.24 -> 8.55 (32t). About 15% saving.
• Wall time within noise (hyperfine, 5 runs, exclusive node): --max 8t 69.6+-2.0 vs 69.5+-2.8 s; 32t 62.3+-3.0 vs 61.7+-2.5 s.
• Roundtrip: --max decode equals input as a sorted set of name/seq/qual records (reorders reads and drops repeated +, as documented).
• Tests: plain_candidate_from_slices_matches_serialized; workspace nextest 623 passed, 1 skipped; doctests, clippy, fmt clean.

Part (b) of #303 (overlapping encode_reference_frame with pass 2) was measured and gave no wall or RSS gain, so it was dropped. Overlapping it with names/permutation or the plain candidate was not tried.

🤖 Generated with Claude Code