mlakolar · Oct 25, 2026 · Oct 25, 2026 · Oct 25, 2026 · Oct 25, 2026
diff --git a/.github/workflows/CI.yml b/.github/workflows/CI.yml
@@ -2,7 +2,7 @@ name: CI
 
 on:
     push:
-        branches: [main]
+        branches: [master]
         tags: ['*']
     pull_request:
     workflow_dispatch:
@@ -29,11 +29,15 @@ jobs:
               with:
                   version: ${{ matrix.julia-version }}
             - uses: julia-actions/cache@v3
-            - name: Install unregistered KernelSmoothing on Julia 1.10
-              if: matrix.julia-version == '1.10'
-              run: julia --project=. -e 'using Pkg; Pkg.add(url = "https://github.com/mlakolar/KernelSmoothing.jl")'
+            - name: Check naming conventions
+              run: ${{ runner.os == 'Windows' && 'python' || 'python3' }} scripts/check_naming.py
+            - name: Resolve dependencies from master checkouts
+              run: julia --color=yes scripts/instantiate.jl
             - uses: julia-actions/julia-buildpkg@v1
-            - uses: julia-actions/julia-runtest@v1
+            - name: Test against dependency master checkouts
+              run: julia --color=yes scripts/test_master.jl
+              env:
+                JULIA_TEST_COVERAGE: "true"
             - uses: julia-actions/julia-processcoverage@v1
             - uses: codecov/codecov-action@v7
               with:
@@ -55,7 +59,7 @@ jobs:
                   version: '1'
             - uses: julia-actions/cache@v3
             - name: Install dependencies
-              run: julia --project=docs -e 'using Pkg; Pkg.add(url = "https://github.com/mlakolar/KernelSmoothing.jl"); Pkg.develop(path = "."); Pkg.instantiate()'
+              run: julia scripts/instantiate.jl docs
             - name: Build and deploy
               run: julia --project=docs docs/make.jl
               env:

diff --git a/.gitignore b/.gitignore
@@ -5,3 +5,9 @@ Manifest.toml
 *.mem
 .DS_Store
 /data/build/
+
+# Generated master dependency test environments
+.master-test-env/
+
+# Local Julia-minor environments are regenerated from master checkouts
+/Manifest-v*.toml
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
@@ -0,0 +1,26 @@
+# Naming and development dependencies
+
+Types and modules use `UpperCamelCase`. Descriptive functions, fields, keyword
+arguments and private helpers use `snake_case`; mutating functions end in `!`.
+Keep established numerical abbreviations (`nobs`, `atol`, `rtol`), mathematical
+identifiers (`X`, `M`, `Σ`, `λ`) and names required by external interfaces.
+
+Develop and test against local `master` checkouts of workspace dependencies.
+Retain package UUIDs and version numbers during this migration. Use canonical
+names in new code; deprecated bindings remain available for downstream users.
+Frozen reference data and recorded numerical results are not renamed.
+
+From the package root, run the complete suite with:
+
+```sh
+julia scripts/test_master.jl
+```
+
+The runner prepares `.master-test-env/v<major>.<minor>` with this package,
+its test extras, and local master checkouts of workspace dependencies. Missing
+workspace checkouts are cloned from master; existing checkouts must already be
+on master. It supports Julia 1.10 and newer and keeps the package `Project.toml`
+unchanged. CI uses the same runner.
+
+Use `julia scripts/test_master.jl --setup-only` to prepare the environment for
+focused checks; the runner prints the environment path to use with `--project`.
diff --git a/NAMING_MIGRATION.md b/NAMING_MIGRATION.md
@@ -0,0 +1,19 @@
+# Naming migration
+
+Versions and UUIDs are unchanged. The development branch is `master`; the
+previous `main` branch remains available. Develop and test with the sibling
+`KernelSmoothing.jl` master checkout, as specified in the project sources.
+
+| Previous public name | Canonical name |
+| --- | --- |
+| `basismatrix` | `basis_matrix` |
+| `modesplot`, `screeplot` | `modes_plot`, `scree_plot` |
+| `covplot`, `designplot`, `fitplot`, `bandplot` | `cov_plot`, `design_plot`, `fit_plot`, `band_plot` |
+| FPCA keyword and stored key `maxK` | `max_k` |
+| Stored key `requested_maxK` | `requested_max_k` |
+
+Old function names remain deprecated bindings to the same generic functions.
+FPCA accepts the old `maxK` keyword and rejects simultaneous non-default old
+and new values. Stored selection/settings named tuples retain the old keys
+alongside canonical keys during the transition. Numerical defaults and
+external StatsAPI, LinearAlgebra and Tables interfaces are unchanged.
diff --git a/Project.toml b/Project.toml
@@ -15,7 +15,7 @@ StatsAPI = "82ae8749-77ed-4fe6-ae5f-f523153014b0"
 Tables = "bd369af6-aec1-5ad0-b16a-f7cc5008161c"
 
 [sources]
-KernelSmoothing = {url = "https://github.com/mlakolar/KernelSmoothing.jl", rev = "master"}
+KernelSmoothing = {path = "../KernelSmoothing.jl"}
 
 [compat]
 Artifacts = "1"

diff --git a/README.md b/README.md
@@ -1,8 +1,8 @@
 # FunctionalDataAnalysis.jl
 
-[![CI](https://github.com/mlakolar/FunctionalDataAnalysis.jl/actions/workflows/CI.yml/badge.svg?branch=main)](https://github.com/mlakolar/FunctionalDataAnalysis.jl/actions/workflows/CI.yml?query=branch%3Amain)
+[![CI](https://github.com/mlakolar/FunctionalDataAnalysis.jl/actions/workflows/CI.yml/badge.svg?branch=master)](https://github.com/mlakolar/FunctionalDataAnalysis.jl/actions/workflows/CI.yml?query=branch%3Amaster)
 [![Docs](https://img.shields.io/badge/docs-dev-blue.svg)](https://mlakolar.github.io/FunctionalDataAnalysis.jl/dev/)
-[![codecov](https://codecov.io/gh/mlakolar/FunctionalDataAnalysis.jl/branch/main/graph/badge.svg)](https://codecov.io/gh/mlakolar/FunctionalDataAnalysis.jl)
+[![codecov](https://codecov.io/gh/mlakolar/FunctionalDataAnalysis.jl/branch/master/graph/badge.svg)](https://codecov.io/gh/mlakolar/FunctionalDataAnalysis.jl)
 [![Aqua QA](https://raw.githubusercontent.com/JuliaTesting/Aqua.jl/master/badge.svg)](https://github.com/JuliaTesting/Aqua.jl)
 
 Functional data analysis in Julia, with planned support for dense, irregular
@@ -26,6 +26,20 @@ using Pkg
 Pkg.add(url = "https://github.com/mlakolar/FunctionalDataAnalysis.jl")
 ```
 
+## Development and testing
+
+Use a `master` checkout for this package and its workspace dependencies. From
+the package root, run the complete test suite with:
+
+```sh
+julia scripts/test_master.jl
+```
+
+The runner resolves sibling master checkouts and test extras in a dedicated
+`.master-test-env` environment, leaving the package dependency declarations
+unchanged. Missing workspace dependencies are cloned from their master branches.
+See [CONTRIBUTING.md](CONTRIBUTING.md) for the naming and development policy.
+
 ## Licence
 
 The package code is MIT-licensed (see `LICENSE`). The datasets loaded by

diff --git a/benchmark/sigma2/common.jl b/benchmark/sigma2/common.jl
@@ -28,7 +28,7 @@ end
 # Same whole-curve CurveCV loss as the package, using sufficient statistics
 # at exact duplicate coordinates within each training/held-out fold. On dense
 # data this avoids predicting the same pair once per subject. No smoother is
-# implemented here; every prediction uses KernelSmoothing.localpolynomial.
+# implemented here; every prediction uses KernelSmoothing.local_polynomial.
 function grouped_cv(selector::CurveCV, fd, pool, kernel, kind)
     assignments = collect(1:nobs(fd))
     if selector.folds != nobs(fd)
@@ -52,7 +52,7 @@ function grouped_cv(selector::CurveCV, fd, pool, kernel, kind)
 end
 
 cv_predictions(train, held, h, kernel, ::Val{:covariance}) =
-    KernelSmoothing.localpolynomial((train.x, train.y), train.z, hcat(held.x, held.y);
+    KernelSmoothing.local_polynomial((train.x, train.y), train.z, hcat(held.x, held.y);
         bandwidth=h, kernel, weights=train.weights)
 cv_predictions(train, held, h, kernel, ::Val{:mean}) =
     FDA._mean_predict(train.x, train.z, train.weights, held.x,
@@ -115,7 +115,7 @@ function estimate(prepared, estimator)
         # Benchmark-only FGM-inspired reference: use the package eigen step,
         # all positive components, and FGM's relative cutoff, not an FPCA K cap.
         system = FDA._fpca_eigen(surface, quadrature_weights(grid), mean.values;
-            maxK=length(grid), eigen_tol=100eps(Float64), sign_tol=1e-12)
+            max_k=length(grid), eigen_tol=100eps(Float64), sign_tol=1e-12)
         surface = (system.eigenfunctions .* system.eigenvalues') * system.eigenfunctions'
         surface = (surface + surface')/2
         # FGM bsplinekit 6f0eac3, src/doubly.jl:223-226: arithmetic mean on

diff --git a/docs/Project.toml b/docs/Project.toml
@@ -8,7 +8,7 @@ Plots = "91a5bcdd-55d7-5caf-9e0b-520d859cae80"
 Tables = "bd369af6-aec1-5ad0-b16a-f7cc5008161c"
 
 [sources]
-KernelSmoothing = {url = "https://github.com/mlakolar/KernelSmoothing.jl", rev = "master"}
+KernelSmoothing = {path = "../../KernelSmoothing.jl"}
 FunctionalDataAnalysis = {path = ".."}
 
 [compat]

diff --git a/docs/make.jl b/docs/make.jl
@@ -7,7 +7,7 @@ import StatsAPI
 makedocs(;
     sitename = "FunctionalDataAnalysis.jl",
     modules = [FunctionalDataAnalysis],
-    format = Documenter.HTML(; edit_link = "main"),
+    format = Documenter.HTML(; edit_link = "master"),
     pages = [
         "Home" => "index.md",
         "Interface" => "interface.md",
@@ -32,5 +32,5 @@ makedocs(;
 
 deploydocs(;
     repo = "github.com/mlakolar/FunctionalDataAnalysis.jl.git",
-    devbranch = "main",
+    devbranch = "master",
 )
diff --git a/docs/src/bases.md b/docs/src/bases.md
@@ -8,7 +8,7 @@ julia> using FunctionalDataAnalysis
 
 julia> b = MonomialBasis((0, 1), 2);
 
-julia> basismatrix(b, [0.0, 0.5, 1.0])
+julia> basis_matrix(b, [0.0, 0.5, 1.0])
 3×3 Matrix{Float64}:
  1.0  0.0  0.0
  1.0  0.5  0.25
@@ -34,7 +34,7 @@ BSplineBasis
 FourierBasis
 MonomialBasis
 nbasis
-basismatrix
+basis_matrix
 gram
 penalty
 ```
diff --git a/docs/src/datasets.md b/docs/src/datasets.md
@@ -43,7 +43,7 @@ growth.covariates
 ```
 
 ```@example datasets
-model = StatsAPI.fit(FPCA, growth.data; ncomponents=2, maxK=2)
+model = StatsAPI.fit(FPCA, growth.data; ncomponents=2, max_k=2)
 mu = mean_data(model)
 plot(mu.grid, mu.values; title="Mean height: four Berkeley sample curves",
     xlabel="Age (years)", ylabel="Height (cm)")
@@ -53,7 +53,7 @@ savefig("datasets-growth-mean.svg"); nothing # hide
 ![](datasets-growth-mean.svg)
 
 ```@example datasets
-modesplot(model, 1; xlabel="Age (years)", ylabel="Height (cm)")
+modes_plot(model, 1; xlabel="Age (years)", ylabel="Height (cm)")
 savefig("datasets-growth-modes.svg"); nothing # hide
 ```
 

diff --git a/docs/src/fdapace.md b/docs/src/fdapace.md
@@ -73,7 +73,7 @@ identifies the matching conventions and the differences retained by design.
 | Choice | fdapace 0.6.0: executable source | FunctionalDataAnalysis.jl and reason |
 |:--|:--|:--|
 | Work grid / `nRegGrid` | `R/SetOptions.R`, `SetOptions`: Dense and DenseWithMV use the number of unique pooled times after `signif(t,6)`; Sparse and `p>>n` use 51. `R/FPCA.R`, `FPCA` then creates an equally spaced grid over the observed range. | Dense uses the supplied data grid and its stored weights, without rounding or resampling. It preserves the data representation. PACE uses 51 equally spaced points over the observed range unless a work grid is supplied, including when PACE is explicitly requested for common-grid data. |
-| `maxK` | `SetOptions`: `min(nRegGrid-2,n-2)`, additionally capped at 20 **only for `methodMuCovEst="smooth"`**; values below 1 are reset to 1. An explicit `maxK` above `nRegGrid-2` is reduced with a message. | The dense default is `max(1,min(m-2,n-2))`; PACE uses `max(1,min(20,m-2,n-2))`, matching the respective audited routes. An explicit dense cap may use up to `min(m,n-1)` and an explicit PACE cap up to `m`; requested components above the cap or retained positive rank throw. |
+| `max_k` | `SetOptions`: `min(nRegGrid-2,n-2)`, additionally capped at 20 **only for `methodMuCovEst="smooth"`**; values below 1 are reset to 1. An explicit `max_k` above `nRegGrid-2` is reduced with a message. | The dense default is `max(1,min(m-2,n-2))`; PACE uses `max(1,min(20,m-2,n-2))`, matching the respective audited routes. An explicit dense cap may use up to `min(m,n-1)` and an explicit PACE cap up to `m`; requested components above the cap or retained positive rank throw. |
 | `methodSelectK` | `SetOptions`: `"FVE"`. `R/SelectK.R`, `SelectK` also accepts `"AIC"`, `"BIC"`, or a positive integer. | `ncomponents=FVE(0.99)` or a positive integer. PACE also accepts `AIC()` and `BIC()`. |
 | `FVEthreshold` | `SetOptions`: 0.99 for `FPCA`. The standalone `SelectK` function separately defaults to 0.95. | `FVE(0.99)` follows the `FPCA` entry point. |
 | Kernel | `SetOptions` initially chooses `"epan"` for Dense and `"gauss"` for every other data type, including DenseWithMV. A later invalid-kernel recovery branch chooses Epanechnikov for DenseWithMV. | The dense sample mean and covariance require no kernel. Optional presmoothing records the supplied smoother and kernel. PACE defaults to Gaussian, following the audited non-dense default. Direct pooled-smoother calls retain their Epanechnikov default; every kernel remains overridable. |
@@ -86,7 +86,7 @@ identifies the matching conventions and the differences retained by design.
 | Smoothed noise estimator | `PC_CovE` uses the rotated diagonal; `SetOptions` supplies the middle-half `rotationCut=c(.25,.75)`, while the internal formal default is `c(0,1)`. | `SmoothedCovariance` and PACE default to the rotated diagonal over the full observed domain, selected by the issue #17 Discovery benchmark. Explicit `noise_method=:surface_diagonal` and `noise_interval=:middle_half` retain the milestone 2 rule. |
 | `rho` | `SetOptions`: `methodRho="vanilla"`, `userRho=NULL`. This means no extra regularization; details are below. | `rho` regularization is not implemented. No ridge is silently added. |
 
-The remembered claim that fdapace always uses `maxK=min(20,m-2)` is therefore
+The remembered claim that fdapace always uses `max_k=min(20,m-2)` is therefore
 incorrect for 0.6.0. So is a description of GCV as its default bandwidth
 selector. The dense component cap follows the audited fdapace default, replacing
 the milestone plan's cap of 20.
@@ -215,7 +215,7 @@ conditioned system is therefore not automatically regularized. An unused
 solver.
 
 The covariance used in that `Sigma_Y` is the **full positive-semidefinite
-fitted covariance up to `maxK`**, not only the selected `K` components:
+fitted covariance up to `max_k`**, not only the selected `K` components:
 `GetEigenAnalysisResults` constructs `eigObj$fittedCov` before selecting K,
 and `FPCA` passes that object to `GetCEScores`. `FVEfittedCov` controls the
 separate user-returned `fittedCovUser`; it does not change this score
@@ -330,7 +330,7 @@ The Discovery benchmark led to `:rotated` with `:full` as the package defaults.
 `R/GetEigenAnalysisResults.R`, `GetEigenAnalysisResults` computes ordinary
 matrix eigenvectors of the covariance surface `G`. It keeps eigenvalues
 `d>=0`, including exact zeros, drops strictly negative ones, and truncates
-the remaining descending list to `maxK`. It does not record a dropped
+the remaining descending list to `max_k`. It does not record a dropped
 negative mass or apply a relative positive-eigenvalue cutoff.
 
 It reports eigenvalues `lambda=h*d`, with the first regular-grid spacing
@@ -348,7 +348,7 @@ eigenvalues above `eigen_tol*maximum(abs,spectrum)` with default
 `eigen_tol=1e-10`, and records the number, absolute mass, and share of the
 strictly negative spectrum. The relative cutoff is an explicit numerical
 choice, rather than a statistical tuning rule from fdapace. Negative mass
-is measured before the `maxK` cap.
+is measured before the `max_k` cap.
 
 The sign agrees with fdapace when the mean overlap is nonzero. If its
 absolute value is at most
@@ -358,7 +358,7 @@ This fallback follows the convention used by FunctionalGraphicalModels.jl.
 Both numerical tolerances are stored on the fit. They are explicit package
 choices that the milestone leaves to the implementation.
 
-Both implementations compute FVE **after** the `maxK` truncation:
+Both implementations compute FVE **after** the `max_k` truncation:
 `cumsum(lambda)/sum(lambda)` over that retained positive spectrum. Thus the
 last retained value is one even if the cap discarded positive components.
 `GetEigenAnalysisResults` first uses `>=` to find its threshold crossing;
@@ -375,7 +375,7 @@ which is distinct from the full fitted covariance used by CE scoring.
 When `methodSelectK` is numeric, `SetOptions` first sets `FVEthreshold=1`,
 and `SelectK` then checks the fixed requested K against available components.
 
-PACE stores the criterion for every candidate from 1 through `maxK`, then
+PACE stores the criterion for every candidate from 1 through `max_k`, then
 applies the first-increase rule. Beyond the retained positive rank, extra
 zero-variance components leave the likelihood unchanged and increase only
 the penalty. It additionally records the
@@ -433,7 +433,7 @@ estimation. These fixtures predate the rotated default; fdapace's enclosing
 The fixed-bandwidth run estimates its mean and covariance from the sparse
 fixture. It uses `userBwMu=0.30`, `userBwCov=0.40`, `kernel="epan"`,
 `useBinnedData="OFF"`, `useBinnedCov=FALSE`, observation weighting,
-`methodXi="CE"`, `methodRho="vanilla"`, fixed `K=3`, and `maxK=5`.
+`methodXi="CE"`, `methodRho="vanilla"`, fixed `K=3`, and `max_k=5`.
 Its 41-point work grid equals the observed-time union. Thus mean evaluation
 before forming residual products agrees in both implementations. Public
 Julia fits reproduce those mean and covariance values at relative error
@@ -488,7 +488,7 @@ recreate `test/data/m3_dense_input.csv` and
 `test/data/m3_fdapace_dense.csv`. Each CSV begins with R and fdapace
 versions. The script has no random draws. It uses 18 rank-five curves on
 21 grid points, `dataType="Dense"`, cross-sectional estimation,
-`error=FALSE`, integration scores, `maxK=5`, no binning, and a fixed K of 3.
+`error=FALSE`, integration scores, `max_k=5`, no binning, and a fixed K of 3.
 A second `FVE(0.99)` run selects K=4; a full-K run records the full FVE curve.
 The first three reference eigenvalues are approximately
 `4.448802620311068`, `1.165542244194415`, and `0.3479570319234146`.
@@ -699,7 +699,7 @@ Two source issues require explicit corrections to exercise native
 `fitted.FPCA` on the milestone 3 supplied-input sparse fixture:
 
 1. `SubsetFPCA` truncates `phi`, `xiEst`, and `lambda`, but leaves `xiVar`
-   at the full retained rank. With selected K smaller than `maxK`, native
+   at the full retained rank. With selected K smaller than `max_k`, native
    fitted bands fail with nonconformable matrices. The reference script
    verifies that failure and subsets only `xiVar` to its selected upper-left
    block. The CE denominator still contains all five retained components.