#560: port pec_csv to the two-stage design (canonical rename in parse)

[jepegit]

Part of #560. Stacked on #602 — review that first; this diff is against its branch.

What

pec_csv gets parse() and declarations() beside its existing loader(), which is untouched. It was the one loader that couldn't be handled mechanically: a BaseLoader (no config_params), columns identified by alias set rather than fixed strings, and a different unit per column embedded in the header (Voltage (V) vs (mV)).

Design decisions and their rationale are in the plan (§2.6a); the pivotal one — canonical rename in parse() → static declarations — was a maintainer call.

How

• parse() reads → matches alias sets → renames matched columns to canonical vendor-neutral names (voltage, current, …) → scales each to its canonical unit → shifts 0-based cycles. The unit scaling is not optional: harmonize() has no per-column unit scaling, so an unscaled mV column would be tagged as V. Silent corruption — and the parity oracle catches it (a mutation removing the scaling shows a 3795 mV diff on potential).
• declarations() does not hand-transcribe the canonical→native map. It reuses derive_column_maps by inverting _COLUMN_KEY_TO_CELLPY_HEADER and composing with cellpy-core's legacy→native map — so pec inherits the provenance rule (vendor test dropped, not mapped onto the framework test_id) and the 0.2.3 energy mapping from the same code the config loaders use. Only the passthrough set is per file: the export's other columns (temperatures, OCV, peak power) are preserved under sanitised names, matching loader().

A false negative it fixed in the oracle

pec carries sparse per-measurement columns that are all-null in the fixture (cell_id, station_temperature_degc, the internal resistances). The numeric diff of two all-null columns is None, which the oracle had been counting as a mismatch — and which would equally have hidden a real values-vs-null difference. The comparison now checks null masks first, then non-null values. Mutation-checked both ways: the neware PER_STEP mutant still fails; an all-null-vs-all-null column agrees.

Result

• pec reaches value parity on every mapped and passthrough column; only exclusion is the shared date_time representation gap.
• Five parity cases now (maccor ×2, neware, custom, pec).
• tests/test_pec_two_stage.py pins the decisions directly.
• Full suite under PYTHONUTF8=1: 1205 passed, 0 failed.

Next on #560

arbin_res (ODBC — CI can only check the committed fixture) and the tier-2 arbin_sql family, then the switchover. Config-driven maccor_txt_zero and batmo_bdf_bdf remain without parity coverage.

🤖 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.