Make batch failure tracking fail-loud, not observability-only (C-21)

[Polichinel]

Context

enrich_dataframe_with_pg_info and enrich_dataframe_with_country_info track failed batches via failed_batches counter and summary ERROR log. But:

• failed_batches and failed_gids are NOT included in the return value
• The caller receives a partial DataFrame with NaN rows and no programmatic signal
• This does not satisfy ADR-003's fail-loud requirement

The downstream _validate() catches the NaN but the error message says "null values in country_iso_a3" — misleading the operator toward spatial mapping issues rather than batch computation failures.

Requirements

Either:

• Option A: Raise an exception when failed_batches > 0 (strict fail-loud)
• Option B: Return failed_gids alongside the DataFrame (e.g., as a tuple or DataFrame attribute)
• Option C: Add a strict=True parameter — when True, raise on batch failure; when False, log-only (current behavior)

Risk Register

C-21 (Tier 2). Part of Cluster B.

Note

If ADR-011 replaces the mapper with a Parquet lookup, batch failures become impossible (a merge can't fail per-batch). This concern becomes moot.