[BaruqueRodrigues]

Execute the pinned principal consumer's state execution, municipal recovery and publication contracts against controlled inputs. Cover failure isolation, all-city failure, dengue/chikungunya/Zika, synthetic RData normalization, numeric SQL precision and idempotent PostgreSQL persistence. Preserve immutable consumer/reference sources and compare persisted content by natural keys.

Reapply on the corrected stack, removing the duplicate parameter helper already moved to #28. Add real Bayesian consumer recovery: the mandatory INLA runner now requires all four scenarios and fails on skips, missing scenarios, warnings or errors. Upload comparison outputs, source manifests and environment evidence.

Validation already performed locally on the consolidated candidate: PostgreSQL suite passed 474 expectations with zero failures/warnings (four INLA-only skips executed separately). The mandatory real-INLA suite passed all four scenarios and 14 expectations on both macOS and Docker, without skips/failures/warnings. Docker R CMD check with built vignettes passed with zero errors/warnings and one documented global-connection NOTE; static lint passed. Evidence and reproduction details are committed under docs/refactoring/evidence/review-2026-10-29/.

Operational gates remain open: real staging Zika coverage requires read-only access and an operational municipality list, neither present in the repository. No consumer pin, migration registry or production database was changed. National capacity and posterior intervals across different fitting-thread budgets require operational/scientific assessment. This PR does not declare production migration complete.

Stacked on #36; #20 remains the integration PR pending review and explicit merges.

[esloch]

@BaruqueRodrigues.

A correção da perda dos parâmetros na recuperação municipal resolve o problema que identificamos anteriormente. A preservação da precisão numérica na geração SQL também é importante.

A ressalva continua sendo a validação com INLA real, principalmente quanto a concorrência, memória e falhas operacionais, conforme solicitado no PR #34.

Lembrando que ele depende do #36.

[esloch]

Thanks, @BaruqueRodrigues.

I reviewed the latest changes.

The remaining validation should focus on the full production workload, using our current baseline of approximately 2h40 on 8 cores, along with the staging Zika audit and scientific review.