nanoql 2.2

Added

• Experimental: a tranche's expected loss from the saddle-point model's own tail.
  SaddlePointTailLossModel<TCopula, TInitTraits> integrates the model's saddle-point tail over the tranche, where
  SaddlePointLossModel reads a large-pool loss. On a 100-name pool (hazard 1%, recovery 40%), where that reading puts
  the whole loss in the first tranche at correlation 0, it is within 2% of the recursive model on the same latent model
  on the 3% to 22% tranches at correlations from 0.3, at one and five years. At high correlation it refuses a tranche
  whose tail it can't read: one attaching close above zero, [0.1%, 3%] from 0.6, or a senior one crowding the pool's
  largest loss, [22%, 60%] at 0.9 (#13).
• Experimental: the analytic GJR-GARCH engine within the no-arbitrage bounds.
  AnalyticGJRGARCHEngine.CalculateWithinBounds, on an engine or static, returns Calculate's results but refuses a
  price outside a call's or a put's no-arbitrage bounds beyond rounding: the Edgeworth price leaves them in both wings
  at every λ, and everywhere once the variance's moments grow. Calculate is unchanged (#14).
• Where the closed-form barrier and lookback engines are exact: each engine's remarks state its condition under term
  structures, on a volatility of time alone: (r − q)/σ² constant over the option's life; for a rebate paid at the hit,
  r/σ² and q/σ² each constant; for the partial-time windows, σ and r − q flat. Each names the engine to use outside it,
  where the library has one (#15).

Fixed

• Black's formula: a price below zero within rounding, which its arithmetic can reach at tiny standard deviations,
  is 0, where QuantLib's check refused it ("negative value"): within 16 ε of the discounted larger of the forward and
  the strike, a few ulps of the forward. A larger negative and NaN still refuse. The Gaussian copula spread engine then
  prices volatilities within rounding of zero, from about 1e-14 to 1e-8, but for most from 4e-14 to 1.3e-9, which its
  density calculator refuses (#12).
• Kahale's interpolated smile section: a segment whose fit missed its right end's price, a jump in the section and
  an arbitrage, is refitted to the same ends and slopes, so nothing else in the section moves; on 163 SABR smiles of a
  sweep the fit missed on 60. The route is opt-in: KahaleSmileSection's interpolate, and the Markov-functional
  model's KahaleInterpolation adjustment (#16).

Known limits

• The Excel add-ins are unsigned. A firm whose Excel requires add-ins signed by a trusted publisher signs the .xll
  with its own certificate.
• The add-ins run on the .NET 10 Desktop Runtime, Excel-DNA's just-in-time path, not compiled to native code.
• The bindings, the C++ layer among them, cover Windows x64 and Linux x64 (glibc 2.28 or later); Windows Arm64,
  Linux Arm64 and macOS on Apple silicon follow in later releases, and Intel Macs have none, though nanoql's .NET
  library runs there.
• Python 3.11 or later.
• The Java layer is built and tested in the repository and not released.

The platforms in this release's run

• Windows x64: 8,728 of 8,728 tests pass, the reference platform.
• Linux x64: 8,655 of 8,728 tests pass; 73 fail, reported and not blocking: NanoQL.Tests.Conventions.HestonSlvMcQualityMaximumTests.TheCalibrationsLargestErrorIsWithinFifteenBasisPoints, NanoQL.Tests.Divergences.AbcdCalibrationGuardTests.TheFitsQuantLibCompletesKeepItsBits(set: 5, a: 0.45336334464724903, b: -1.4254877636528123, c: 1.5935876342523034, d: 0.26597942707915356), NanoQL.Tests.Divergences.CalibrationBasketNominalTests.AFlippedSolutionKeepsItsNominal, NanoQL.Tests.Divergences.CirFellerConstraintTests.ACalibrationWithTheSpeedAndLevelFixedIsQuantLibs, NanoQL.Tests.Divergences.CmsCalibrationGuessTests.TheFixedPathStartsAtTheGuess, NanoQL.Tests.Divergences.CmsCalibrationReversionTransformTests.TheLevenbergMarquardtFitsKeepQuantLibsValues, NanoQL.Tests.Divergences.CmsCalibrationReversionTransformTests.TheSimplexFitsHaveAFiniteReversion, NanoQL.Tests.Divergences.CrossCurrencySwapCompoundedSpreadFairSpreadTests.TheBasisSwapsCompoundingLegsFairSpreadIsSolved, NanoQL.Tests.Divergences.CrossCurrencySwapCompoundedSpreadFairSpreadTests.TheFixedVsFloatingSwapsCompoundingLegsFairSpreadIsSolved, NanoQL.Tests.Divergences.CrossCurrencySwapCompoundedSpreadFairSpreadTests.TheMtMSwapsCompoundingLegsFairSpreadIsSolved and 63 more.
• macOS Arm64: 8,652 of 8,728 tests pass; 76 fail, reported and not blocking: NanoQL.Tests.Conventions.BoostSignTests.TheSignOfANaNIsItsSignBit, NanoQL.Tests.Divergences.AbcdCalibrationGuardTests.TheFitsQuantLibCompletesKeepItsBits(set: 5, a: 0.45336334464724903, b: -1.4254877636528123, c: 1.5935876342523034, d: 0.26597942707915356), NanoQL.Tests.Divergences.BetaRiskTests.AMultiEventPathMovesItsLaterDatesOnly, NanoQL.Tests.Divergences.CTSMMCapletAlphaFormCalibrationTests.ACalibrationWhoseReturnedPassFitsKeepsQuantLibsNumbers, NanoQL.Tests.Divergences.CalibrationBasketNominalTests.AFlippedSolutionKeepsItsNominal, NanoQL.Tests.Divergences.CirFellerConstraintTests.ACalibrationWithTheSpeedAndLevelFixedIsQuantLibs, NanoQL.Tests.Divergences.CmsCalibrationGuessTests.TheFixedPathStartsAtTheGuess, NanoQL.Tests.Divergences.CmsCalibrationReversionTransformTests.TheLevenbergMarquardtFitsKeepQuantLibsValues, NanoQL.Tests.Divergences.CmsCalibrationReversionTransformTests.TheSimplexFitsHaveAFiniteReversion, NanoQL.Tests.Divergences.CrossCurrencySwapCompoundedSpreadFairSpreadTests.TheBasisSwapsCompoundingLegsFairSpreadIsSolved and 66 more.
• The linux-x64 bindings: every check green.
• The win-x64 bindings: every check green.