feat: publish verified market comparison with robust accounting

[dev-belly]

AlphaForge's real-data script could overwrite the published synthetic case and attach assurance to data loaded from a different working directory. Keep real runs separate, require the repository root for the real-research entrypoint, and record source hashes plus explicit historical-membership, filing-date, and capitalization limitations. Incomplete source coverage remains exploratory and fails the strict completeness gate.

Publish a frozen comparison of equal weight, fixed momentum, Ridge, LightGBM, and one explicit size-feature ablation using shared forecasts, calendars, outcomes, declared seeds, costs, and execution assumptions. All 22 cases are retained. The independent verifier recomputes net/gross headlines and seed aggregates, rejects nonfinite or missing metrics and incomplete/duplicate case inventories, and checks dataset-identity consistency. Preserve the original byte-hashed artifacts across Git LF/CRLF checkouts rather than changing their manifest hashes.

Harden accounting boundaries: representable fees avoid intermediate overflow; nonfinite/unpayable costs, invalid NAV, and unrepresentable share quantities or daily returns raise clear errors. Actual bankruptcy is a legal absorbing zero-capital state. Source data, frozen returns, six evidence files, model choices, and measured results remain unchanged. The publication discloses its source and diversification limits and does not claim tradable alpha.

Validation: the final combined revision passed 1,350 tests in CI with 94% coverage, all three Linux Python versions, Windows Python 3.12, and the documentation build. Locally, the accounting version passed 1,317 complete-suite tests; the later verifier and entrypoint additions passed 77 focused regressions plus lint, formatting, type checks, navigation checks, and strict documentation build. Independent direct-formula checks match 22 cases × ten portfolio headlines and ten aggregate groups; all six frozen artifacts remain byte-identical. Four Git checkout combinations preserve their bytes, and altered/rehashed evidence is rejected. The existing single test warning is non-blocking.