#560: port arbin_res to the two-stage design

[jepegit]

Part of #560. Off master, not stacked. Overlaps slightly with #607 (pec) — see the note at the end.

What

arbin_res was the tier-1 loader I flagged as awkward: Access .res via ODBC on Windows, mdbtools on posix, multi-test files, aux-table joins. Two facts made it tractable:

• CI can check it. The loader_arbin_res golden runs green on Linux CI — mdbtools is installed there — so the parity case runs in CI, not only locally. Where no backend exists, the parity test skips gracefully, mirroring the golden's skip_reason.
• The vendor stage is thin. get_headers_normal() is already a {cellpy attr → Arbin column} map — the same shape the configuration loaders carry — so declarations() reuses derive_column_maps and inherits the provenance rule (Test_ID dropped, not mapped onto the framework test_id) for free. parse() is just the database read via _loader_win/_loader_posix, stopping before the rename and the Excel-serial datetime conversion that harmonize() now owns. It shares loader()'s read path, so the two cannot drift.

Scope

Single test: the golden fixture is single-test with no aux columns, and loader() still handles the multi-test split and aux joins. A multi-test .res — where Test_ID is nonzero and harmonize()'s framework test_id differs from the vendor value — is the switchover's concern. In this fixture Test_ID is 0, so the agreement holds by the file's value, not by construction (noted for the switchover).

Verification

• Value parity on all 17 shared measurement columns; only date_time excused, as everywhere.
• Mutation-checked: swapping the charge/discharge vendor names is caught. A column-drop mutation is not caught by this oracle (shared-columns-only) — but the loader_arbin_res golden pins the full column set, so that's layered.
• tests/test_arbin_two_stage.py pins the decisions and skips without a backend.
• Full suite under PYTHONUTF8=1: 1241 passed, 0 failed.

Merge-order note

Two small changes here also appear in #607 (pec), because both are the first non-config loaders: the parity oracle reading loader.config_params via getattr, and the per-case backend skip. If #607 merges first, PARITY_CASES conflicts trivially (both append a case) — resolvable in seconds. The getattr line is identical in both and merges cleanly.

🤖 Generated with Claude Code

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.