UNIC-2179 Add curated step to bd_previnf and reset the anonymized datasets

[laurabegin]

Companion to unic-etl#1737, which adds a red/curated step to bd_previnf so the anonymized datasets can upsert on the financial year instead of overwriting the previous one.

The config moves from yellow/anonymized/ to red/curated/, following every other system that has a curated step (etraceline, medecho, pericalm, opera…), and gains a curated step that runs before the anonymized one.

Changes

• New red/curated step — the 3 curated_bd_previnf_* datasets, gating the anonymized step.
• run_type: "initial" on the 6 anonymized datasets — the 3 curated-sourced ones change schema and grain, and the other 3 switch OverWrite → Upsert. Since raw is OverWritePartition, every historical partition is still available, so a reset rebuilds the full year history instead of only moving forward.
• greater_or_equal_partition_counts pre-test on both steps. The test resolves each destination back to its source, so the two cover disjoint tables: the curated one checks the 3 raw tables feeding curated, the anonymized one checks the 3 curated outputs plus the 3 raw tables that bypass curated. Together every bd_previnf source is covered; today none are.
• xsmall everywhere — all tables hold fewer than 50 rows.

Task graph verified locally: curated.pre_tests → 3 curated tasks → anonymized.pre_tests → 6 anonymized tasks. 171 passed, 0 import errors.

Merge order

Merge unic-etl#1737 first. The pre-tests resolve destinations through Common.mappings, and the curated bd_previnf entries are added there in that PR — without it the pre-test matches no mapping and silently tests nothing.

Deploy notes

• The DAG is renamed: anonymized_bd_previnf → curated_bd_previnf (the id comes from the config directory). The new DAG is created paused and needs unpausing; the old id keeps its run history.
• run_type: "initial" is persistent config, so it resets on every monthly run until someone sets it back to default. If the intent is a one-time reset, that needs a follow-up PR after the first successful run.
• Expect the pre-tests to fail once at a financial-year rollover if the new year ships fewer rows than the previous delivery. It fails safe — nothing is written, clear the task and re-run.

🤖 Generated with Claude Code

[philipbouvrette]

😮

[laurabegin]

😮

@philipbouvrette tu m'as fait réalisé que ça avait mal été push. C'est fixed. Là y a des changements.