fix: Cut OpenLab CDS peak memory 4x and handle optional ACAML metadata

[nathan-stender]

Motivation

A 745 MB Agilent OpenLab CDS .rslt bundle (10 injections) could not be converted at all. Two independent problems:

1. Memory. The read path held the entire file in memory twice, and the nested-zip access pattern thrashed.
2. Parsability. Four ACAML elements that are absent from some exports were accessed unconditionally.

Memory

• Added NamedFileContents.get_seekable_bytes_stream(). The existing get_bytes_stream() does self.contents.read() and wraps the result in a BytesIO, holding the file twice. The new method returns the underlying handle when it's already a seekable binary stream (as it is for allotrope_from_file), falling back to the old behaviour otherwise. Safe because to_allotrope.py and discover_vendor() both seek(0) before handing the handle to a parser, and the new method seeks to 0 itself. Only the OpenLab CDS parser opts in; the ~30 other callers of get_bytes_stream() are untouched.
• decode_data now spools each .dx member to a NamedTemporaryFile in 1 MB chunks before opening it as a nested zip. Previously it opened a ZipFile over a ZipExtFile, which isn't seekable, so zipfile emulated every backward seek by re-reading and discarding in 16 MB chunks — and the .CH members sit at byte ~74.1 M of a 74.5 MB member.

Measured with an RSS sampler on the 745 MB bundle: peak RSS for the decode stage ~1.00 GB → ~235 MB; full end-to-end conversion peaks at ~218 MB. Decoding is also faster (1.7 s → 1.3 s) now that the seek emulation is gone.

Optional ACAML metadata

Each of these previously raised on the affected file:

Location	Problem	Fix
decode_acaml_data	Resources/SeparationMedium absent → KeyError	.get(), plus new get_column_info() in the structure layer, which resolves to {} so the column fields are omitted
decode_data	no .sqx in the bundle → StopIteration	.sqx is now optional; the analysis method falls back to the ACAML Content/Method names via new get_acaml_analysis_method(), used only when they agree on a single name
extract_rx_file	Result/InjectionCompound absent → KeyError	hoisted to result_data.get("InjectionCompound"); peaks are emitted without compound metadata
create_metadata	single-token instrument name → .split()[1] IndexError	new get_brand_name(), returns the second token only when present

Also guarded the unconditional peak["Area"]["@unit"], peak["Height"]["@unit"] and peak["Peak Metadata"] reads in create_peak, whose sibling expressions already treated those keys as optional.

The bundle now converts to 30 measurements (10 injections × 2 chromatograms + 1 pressure trace) in 2.0 s.

Test data

New fixture Sirius-2023-09-01 07-52-44-04-00.rslt, 33 KB, containing no customer data. Cut down from the existing anonymised Luxo HPLC fixture: 2 injections instead of 10, and each .dx keeps only the three members the parser reads (2 .CH + the pressure .IT), dropping ~1.1 MB of unread instrument traces per injection. Then edited to reproduce all four conditions above.

Confirmed to be real regression coverage: with the source changes stashed, the new fixture fails (KeyError: 'SeparationMedium') while the existing fixture still passes.

The pre-existing Luxo HPLC expected JSON is byte-identical — no changes to existing test data.

Out of scope

• 3D UV data cubes. The .UV member is 95.5% of each .dx and is not read, so the output has no 3D spectra. ASM does define three-dimensional ultraviolet spectrum data cube (AFR_0002551) in the LC schema and generated model, but it isn't exposed by the schema mapper and no parser emits it. Deliberately deferred.
• Calibration sample role. constants.SAMPLE_ROLE_TYPE maps only Blank and Sample, so calibration standards get no sample role type. Adding it would change output for existing files containing calibration samples, so it's left for a separate change.
• Remaining memory scaling. Retained data grows ~4.8 MB per injection because decode_data_cubes converts numpy float64 arrays into Python float lists (~4× blowup). Peak is now dominated by retained data rather than file I/O. Keeping the cubes as numpy arrays until serialization is the next lever.

Testing

• hatch run lint — clean (ruff, black, mypy over 788 files)
• hatch run test_all.py3.10:pytest tests/ — 1488 passed

🤖 Generated with Claude Code