v0.13.0 -- the TimeSeries container

The heavy TimeSeries container -- the last unported portable slice of Numerics, and the one
four other members were waiting on. Numerics/Data/Time Series/TimeSeries.cs (2,334 lines) and
its Series base are ported, along with a DateTime value type that stands in for
System.DateTime (a corehydro addition: the calendar is what the container is built on, and the
core takes no external dependency). This closes branch port-time-series.

Added

• A time-series object in both packages. R gets time_series() and 28 ts_*() verbs over a
  corehydro_ts classed list; Python gets a TimeSeries class with the same operations as
  methods -- the shape distribution() / Distribution already established. Between them they
  reach the whole container: moving windows, cumulative sums and differences, the value
  transformations, missing-value replacement and interpolation, date filling, clipping, shifting,
  interval conversion, the summary and monthly statistics, the duration curve, the five block
  series (with ts_water_year() and ts_calendar_year() by name), peaks over threshold, seasonal
  decomposition, and two resamplers.
• Dates are native in both languages. R takes and returns POSIXct in UTC and also accepts
  Date, ISO 8601 strings and epoch seconds; Python takes and returns numpy.datetime64 and also
  accepts datetime, a pandas DatetimeIndex, ISO strings and floats. Epoch seconds travel on
  the wire, never .NET ticks -- a tick count near 6.3e17 is not exactly representable in a double,
  while epoch seconds is exact for every interval the container supports.
• A timeseries toolbox group, the nineteenth, plus four fixture files. One of them,
  fixtures/timeseries/date_time.json, is what puts the ported DateTime -- which has no
  upstream source file to be checked against -- permanently behind the dotnet oracle gate:
  AddMonths day-clamping, leap years, the fractional-AddDays split, DayOfYear, DayOfWeek
  and ToOADate.
• Four members un-gated, each deferred with the container it needed: the seasonal
  PointProcessModel data path and GeneratePOTTimeSeries, DataFrame.CreateBlockSeries and
  CreatePeaksOverThresholdSeries (with ExactData.DateTime), and the ARIMAX covariate
  forecast-tail extension (CovariateExtension = "block_bootstrap" / "knn"). The seasonal point
  process is reachable from both packages: analysis_data() takes dated observations and
  model_point_process() takes seasonal / time_block / start_month.
• autocorrelation() accepts a time series in both packages.

Fixed

• The phase itself broke nothing and found four upstream behaviours instead (see below). The
  documentation sweep that closed it did find one thing: worked example 04 pinned map_fitness
  at 473.558409617748, the value from before the v0.2.0 upstream sync moved MAP fitness onto
  the log-likelihood scale as -473.558409617748. Both language pages were wrong and both
  pipelines were green, because Quarto's freeze keys on the source file rather than on the
  package the source calls, and notebooks are never re-executed. Every other page was audited by
  forced re-execution and reproduces.
• The documentation had accumulated scope claims that stopped being true as v0.9.0 through
  v0.13.0 landed: five upstream example notebooks were still listed out of scope after being
  ported, Data.Statistics was marked partial when all thirteen of its classes are ported, and
  seven example pairs said capabilities were unreachable that mcmc_posterior(),
  autocorrelation(), ml_jenks_breaks(), summary_statistics() and correlation() have
  since made public.

Changed

• HEC-FDA is no longer listed as a porting target. It is being ported separately in
  hecfda; the status page carries it as a related
  effort instead of a planned upstream source.
• The rendered site output (219 files: the .html pages sitting beside their sources, their
  *_files/ figures, and site_libs/) is no longer tracked. site/_freeze still is.

Notes

• Four upstream findings, each measured against the real compiled library, mirrored rather
  than corrected, and written up in docs/upstream-csharp-issues.md: CumulativeSum silently
  drops the series' time interval; three indexed math overloads reach through an index their own
  guard rejected while their seven siblings skip it; MovingAverage / MovingSum swap their
  ArgumentException arguments; and -- the consequential one -- PointProcessModel's seasonal
  day-of-year list comes out date-sorted while the likelihood pairs it positionally with the
  caller's unsorted exact series, so a record not already in date order has each magnitude scored
  against another event's day.
• The Autocorrelation TimeSeries overloads are still not ported, and that is now a
  measurement rather than a severance. They are the same four bodies with a
  missing-value-skipping mean, and the difference is not observable, because the lag-0
  autocovariance sums over every observation. Driven against the real library, both overload
  families return identical values on a clean series and all-NaN on a gappy one.
• A seeded resampling run is bit-identical between R and Python, and
  fixtures/timeseries/timeseries_cross_language.json asserts it at zero tolerance. Unlike the
  callback layer's honest limit, every operation here happens inside the shared core, so this is a
  guarantee rather than a best effort.

---

Both packages install from source and need a C++17 compiler.

pak::pak("cameronbracken/corehydro/corehydror@v0.13.0")

pip install "git+https://github.com/cameronbracken/corehydro.git@v0.13.0#subdirectory=corehydropy"

Validated against USACE-RMC Numerics 2.1.4 and RMC-BestFit 2.0.0. Full notes in
CHANGELOG.md;
documentation at https://cameronbracken.github.io/corehydro/.

v0.1.0-alpha

Changelog

All notable changes to corehydro (the shared C++ core, the corehydror R package, and
the corehydropy Python package) are documented here. The format follows
Keep a Changelog, and the project uses
semantic versioning. The three components are versioned together.

Unreleased

0.1.0 - 2026-07-11

First tagged release. Everything below is new.

Added

• Shared C++17 core with full parity to the USACE-RMC Numerics and RMC-BestFit C#
  libraries (the probability, estimation, and analysis layers): all 42 univariate
  distributions, multivariate distributions and seven bivariate copulas, eight MCMC
  samplers (RWMH, ARWMH, DEMCz, DEMCzs, HMC, NUTS, Gibbs, SNIS) with Gelman-Rubin and
  ESS diagnostics, bootstrap uncertainty, the MLE/MAP/Bayesian/GMM estimators, all
  RMC-BestFit model families (flood frequency, Bulletin 17C, mixtures, competing risks,
  point process, time series, spatial GEV, rating curves, bivariate/copula), the full
  Analyses layer, and the Diagnostics layer (leverage, PSIS-LOO influence, prior
  influence, predictive checks). See the porting status page for the exact scope.
• R package corehydror and Python package corehydropy, thin bindings over
  the same compiled core. A bit-exact Mersenne Twister port means seeded results are
  identical across R, Python, and the upstream C# libraries.
• Public distribution API: construct any of 38 families by name; density, CDF,
  quantile, moments, L-moments, log-likelihood, seeded random generation, and fitting
  by MLE, L-moments, or product moments.
• Analysis functions (19 in each package): univariate_analysis,
  fit_distributions, bulletin17c_analysis, mixture/competing-risk/point-process/
  composite/spatial-GEV analyses, AR/MA/ARIMA/ARIMAX time-series analyses, bivariate
  and coincident-frequency analyses, rating curves, bootstrap uncertainty, predictive
  checks, and estimation diagnostics.
• mcmc_sample(): direct MCMC over any distribution family (7 samplers) with
  chains, acceptance rates, MAP, posterior summaries, R-hat, and ESS.
• Statistics utilities: Multiple Grubbs-Beck low-outlier test, Box-Cox and
  Yeo-Johnson transforms, plotting positions, Latin hypercube sampling.
• Validation: a language-neutral oracle fixture suite consumed by C++, R, and
  Python runners, plus a dev-only dotnet gate that replays every fixture against the
  real C# libraries (4,100+ values reproduced, 0 failures).
• Documentation site with worked examples in both languages ported from the
  official Numerics-Python-Examples repository (11 example pairs, each ending in an
  executable reproduction check against the upstream C# outputs), a Python API
  reference (quartodoc), an R API reference (pkgdown), and a porting status page.
• Developer tooling: pixi environment with tasks mirroring the Makefile targets,
  GitHub Actions CI (3 OS matrix) and Pages deployment.

Changed

• Renamed from bestfit (packages bestfitr/bestfitpy) to corehydro
  (corehydror/corehydropy), reflecting the goal of carrying code from both
  USACE-RMC and HEC libraries in one package family.