feat(input): XML calculation-input schema + extracted dataclasses

[IanMayo]

Summary

Makes the input side of the pipeline symmetric with the output. The calculation input was an unvalidated JSON dict (json.loads → dict, shape enforced only implicitly by the builder). It's now an XML document held to its own contract, parsed into typed dataclasses generated from that schema — mirroring how Platform is generated from the output schema.

One file format (XML) throughout: mock input schema → extracted dataclasses → typed, validated parse → compute → validated output XML.

What changed

• schema/calculation_input.xsd — small XSD in the same salami-slice idiom (no-namespace, XSD 1.1, global elements, xs:all of refs, docs on declarations). Models the parameter vocabulary, which is deliberately a different shape from the output: a handful of parameters (base frequency, band ratio/count, bearing step, roll-off, directivity lobe) are expanded by the acoustic seams into the output's bands/sectors, and some inputs (e.g. the active sonar's detection threshold) are consumed to derive an output figure and never appear verbatim.
• src/acoustic_dataset/input_models/ — dataclasses generated from the input schema via xsdata. A separate package because the input and output schemas share element names (Characteristics, Sensors, RadiatedNoise, …), which would collide in a shared package's __init__.
• generate.py — regenerates both schemas (make generate does both; --schema <xsd> still does one); import-checks each module.
• build.py — load_input() runs the input XSD gate before parsing into a typed CalculationInput; the _build_* helpers read typed attributes instead of dict keys. acoustics.load_input removed so acoustics stays pure.
• examples/calculation_input.xml replaces the .json (same values).
• .gitattributes — pins LF so the schema-docs drift gate (a raw read_bytes() byte-compare) passes on Windows checkouts with core.autocrlf=true, not only on LF-based CI.
• CLI, pyproject.toml (ruff/mypy excludes for input_models/), tests, docs, and the CLI-commands contract updated to match.

Verification

• Pipeline output is byte-identical to the golden file — the input format changed without changing behaviour.
• ruff, mypy, and all 41 tests pass (including the schema-docs drift gate, which previously failed locally on the CRLF artifact).
• Output models show no drift after regeneration.

Note

Historical planning artifacts under specs/ (data-model.md, research.md, plan.md, tasks.md) still reference the old .json — left as point-in-time records. Happy to update if preferred.

🤖 Generated with Claude Code

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-06-19 12:03 UTC