Reconciler

The bridge auto-decomposes why the end-of-period forecast changed between any two plan versions. Pure arithmetic, no LLM. The key primitive is cash_io, which sums legs on asset accounts only - structurally preventing the debt double-count that plagued earlier manual analysis. Components (per expense/income line, anchor restatement, transfers, fun changes) always sum to delta within a dollar. Shown as a horizontal waterfall chart with red/green bars.

The version viewer lets you click any row in the 28-version history table and see the full snapshot read-only: anchor, end-of-period, fun budget, income/expense line tables with per-month values, and the bridge waterfall for that version. The history table now shows five columns (When, Note, Anchor, Fun, End cash w/fun) so you can see the entire plan evolution at a glance - from $952 at v23 through the fake-anchor correction down to $155 at v30.

When the financial plan gets re-anchored to fresh data, the forecast changes - sometimes by hundreds of dollars - and the question is always "why did it move?" The bridge answers that automatically: pure arithmetic decomposition into per-line components (dining worsened by X, groceries improved by Y, income landed short by Z), with anchor restatement, debt-service transfers, and fun budget changes as separate buckets. Components always sum to the delta within a dollar.

The key design decision was the cash_io primitive: sum legs on asset accounts only, not credit cards or lines of credit. This structurally prevents the debt double-count that an earlier audit flagged ($285 per re-anchor cycle) - interest accrues on liability accounts with zero cash impact, while card payments show cleanly as transfers.

Three retroactive bridges computed and stored. The waterfall chart lives in the Plan tab with version selector buttons and an expandable detail table showing old forecast, window actual, and new forecast for each line.