cadrumo.application.storage.calc_sheets._parity_comparison module¶

Pure per-casilla comparison across the calculation surfaces.

Extracted from the parity harness so the comparison can be reached without the harness’s write path. That separation is load-bearing rather than tidy: verify_modelo_parity() acquires its spreadsheet side by CREATING or updating the workbook, seeding operator inputs into Entradas and relations into Tarifas, and reading Cálculos back — so any caller that must not write cannot go through it. The export preview is exactly that caller, and the decision record requires it to reuse this comparison rather than grow a second differ.

The functions here take mappings and casilla definitions only. They deliberately do NOT take a snapshot or a scenario: the export preview holds neither, and a comparison that demanded them would have to be reimplemented for it, which is the duplication this module exists to prevent.

The three surfaces are named from the parity use — local, sheets, aeat — and the preview use reads them as “what the plan would write”, “what the spreadsheet currently holds”, and “absent”. A divergent row is then precisely a cell whose value would change.

COMPARISON IS EXACT, AND THAT IS THE CONTRACT RATHER THAN AN OVERSIGHT. The reconcile surfaces compare filed against computed at the tolerance the registry publishes per verification expectation, and it would be natural to assume this module simply forgot to. It did not, for three reasons, and the first is decisive: THE REGISTRY PUBLISHES NO TOLERANCE FOR THIS AXIS. Its tolerance is declared on verification expectations, which govern filed-versus-computed reconciliation — a legal question about whether a taxpayer’s return agrees with the authority. Engine-versus-spreadsheet is not that question, so adopting that tolerance here would apply a legal allowance to an axis no law governs.

Second, the preview consumer asks “would this cell’s value change”, and any difference is a write. A tolerance would make the preview UNDER-report cells the apply would overwrite, which is the one error a dry run must not make. Third, the parity consumer asks whether the spreadsheet reproduces the engine, and a cent of slack there would hide exactly the formula transcription error the harness exists to catch.

So a future reader reconciling this module against the reconcile surfaces should not fold it onto verification_policy().tolerance. If a real rounding difference is ever observed between the Decimal runtime and a spreadsheet formula, the fix is to quantize both sides to the modelo’s money scale before comparison, not to widen the comparison.