feat: live-formula Excel export (toolkit + canonical translator + verification)

AGENTS.md

@@ -98,6 +98,14 @@ small, trustworthy accounting kernel.
 9. Present only promotion-eligible challengers for human ratification. Never
    replace the champion without explicit approval.
 
+## Excel Output
+
+When the user wants Excel output with working formulas, use `model_to_excel` for
+the standard monthly model. For any other cadence or layout, generate a
+company-specific exporter from `pyfpa.excel.toolkit` in the generated namespace
+and register it as a report entrypoint. Install `formulas` and run
+`verify_workbook` before delivering. No workbook ships unverified.
+
 ## Kernel Guardrails
 
 - Never hide reconciliation differences, missing actuals, duplicate names, or

CLAUDE.md

@@ -57,3 +57,7 @@ After the initial architecture is approved, use `fpa-research-loop` for bounded
 autonomous improvement. You may generate, evaluate, and discard challengers
 without asking for approval each time. Human approval is required for promotion
 before a challenger replaces the champion.
+
+For Excel output with working formulas, use `openfpa model-export` or
+`pyfpa.model_to_excel` for the standard monthly model. Run `verify_workbook`
+before delivering. No workbook ships unverified.

README.md

@@ -209,6 +209,7 @@ openfpa scorecard-render
 openfpa experiment-list
 openfpa context-pack
 openfpa onboarding-render
+openfpa model-export
 ```
 
 In a source checkout without the `openfpa` console script installed, every
@@ -394,7 +395,9 @@ The base kernel includes:
 - 13-week direct-method cash forecasting;
 - revenue, COGS, operating-expense, debt, tax, and working-capital primitives;
 - reconciliation, segment, SKU, and divestiture analysis helpers;
-- CSV ingestion and Markdown or Excel reporting;
+- CSV ingestion and Markdown or Excel reporting: static value export via
+  `forecast_to_excel`, and a live-formula model workbook via `model_to_excel`
+  (named assumption cells, real formulas, verified against the engine in CI);
 - forecast snapshots, scoring, and holdout backtests;
 - workspace, intake, correction, experiment, retrieval, and research records;
 - experimental cross-company prior and skill mining.
@@ -403,6 +406,14 @@ The base kernel includes:
 with cohorts, projects, contracts, fleets, stores, or complex inventory may need
 a different generated model.
 
+The `model_to_excel` function compiles an `EntityConfig` into a two-sheet workbook:
+an Assumptions sheet of named, editable driver cells and a Model sheet where every
+P&L and cash-flow line is a real formula referencing those names. The `verify_workbook`
+function evaluates the workbook with a Python formula engine and compares every line
+and month to `cashflow_from_config`, so the workbook is verified against the engine
+before it is used. For cadences or layouts the kernel does not cover, compose
+`pyfpa.excel.toolkit` in a company-specific exporter and verify it the same way.
+
 ## Skills
 
 The repository includes skills for: