[survey2ddi 5/6] Adopt the qwacback equivalence test + add a Python↔TS parity gate

Step 5 of 6 for the survey2ddi retirement. Plan: HANDOVER_SURVEY2DDI.md. Independent of the CLI/adapter issues — can run in parallel. Blocks the emit-core deletion in CorrelAid/survey2ddi.

Two pieces of coverage that must not die with survey2ddi's emitter

1. The qwacback equivalence test (permanent)

tests/integration/test_conversion_equivalence.py in CorrelAid/survey2ddi posts the same XLSForm through survey2ddi and qwacback and compares the DDI shape, for every answer type qwacback supports (qwacback/internal/examples/examples.go). It is marked integration (docker + the schematron-worker image, opt-in via -m integration).

qwacback is a listed consumer of this library (see the README's ecosystem table and HANDOVER_QWAC.md), so this comparison is worth more here than it ever was in survey2ddi. Port it to compare buildDdiXml against qwacback, reusing this repo's existing pytest validation setup (tests/validation/, which already runs XSD and Schematron checks) or the TS contract project — whichever fits the docker orchestration better. Note the shape normalization it needs: qwacback returns a bare <var> / <varGrp> when there is only one, so the test synthesizes a <dataDscr> container.

2. The Python↔TS byte-parity script (temporary)

A script that runs both emitters over the fixture corpus and byte-compares XML and CSV. It exists to gate the deletion of survey2ddi_core's emit core, and it gets deleted with it — do not wire it into CI as a permanent job.

Known-good starting point, already verified by hand: on a flat survey (no groups, no select_multiple, no _other), build_data_csv and buildDataCsv are byte-equal — quoting, CRLF, None → empty cell, space-joined multi expansion all agree.

Known divergence, expected and correct: get_canonical_columns returns input order while ddi_xml.py emits bucketed order, so any survey with a grid, a select_multiple or an _other pattern gets a Python CSV whose header does not match its own XML. This library derives columns from the emitter's own bucket walk, so it disagrees with Python and agrees with the XML. The script must classify that difference rather than fail on it — and any other difference is a finding to triage, since a Python behaviour missing here is a bug here.

Acceptance criteria

• qwacback equivalence test runs against buildDdiXml in this repo, same type coverage, still opt-in
• Parity script covers XML and CSV over the fixture corpus
• Column-order divergence is classified and reported, not silently tolerated and not fatal
• Any other divergence is filed as an issue here before survey2ddi deletes anything

Start with HANDOVER_SURVEY2DDI.md, the existing tests/validation setup, and the survey2ddi equivalence test; use buildDdiXml and the opt-in integration marker as entry points. Run the validation and integration coverage against qwacback, then compare XML and CSV across the fixture corpus, reporting the known column-order divergence while surfacing any other mismatch before survey2ddi deletion.