Make the environment resolvable, and run the tests in CI

.github/workflows/CI.yml

@@ -0,0 +1,61 @@
+name: CI
+
+on:
+  push:
+    branches:
+      - main
+    tags:
+      - "*"
+  pull_request:
+
+env:
+  JULIA_PKG_SERVER: internal.juliahub.com
+  # The MPC stack is not all registered: MPCComponents, the LinearMPC fork and the acados JLLs
+  # are cloned from Git, which needs the CLI client rather than libgit2 for the credentials the
+  # registry step installs.
+  JULIA_PKG_USE_CLI_GIT: true
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  test:
+    name: Julia ${{ matrix.version }} - ${{ matrix.os }} - ${{ matrix.arch }} - ${{ github.event_name }}
+
+    # This repository is public, and the org's `Default` runner group has
+    # `allows_public_repositories: false`, so a bare `runs-on: self-hosted` job never resolves and
+    # queues forever. Public repositories are served by the `public` runner group, whose scale sets
+    # advertise no bare `self-hosted` label and have to be named: the tiers are
+    # `self-hosted-{2vcpu-4gb,4vcpu-8gb,8vcpu-16gb,8vcpu-32gb}-public`. This suite compiles the
+    # multibody models and builds acados, so it takes the heavy-compute tier.
+    runs-on: self-hosted-8vcpu-16gb-public
+
+    strategy:
+      fail-fast: false
+      matrix:
+        version:
+          - "1.12"
+        os:
+          - self-hosted-8vcpu-16gb-public
+        arch:
+          - x64
+
+    steps:
+      - uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6
+
+      - uses: julia-actions/setup-julia@5c9647d97b78a5debe5164e9eec09d653d29bd71 # v2.6.1
+        with:
+          version: ${{ matrix.version }}
+          arch: ${{ matrix.arch }}
+
+      - name: "Set up private JuliaHub registries"
+        uses: "JuliaComputing/set-up-private-juliahub-registries@2f465506027adbe8086afc41dd3514e9020d1577" # v2.0.0
+        with:
+          encoded-juliahub-token: "${{ secrets.JULIAHUB_TOKEN_ENCODED }}"
+          ssh-key: "${{ secrets.JULIASIM_REGISTRY_SSH_KEY }}"
+          enable-JuliaHubRegistry: "true"
+
+      - uses: julia-actions/julia-buildpkg@e3eb439fad4f9aba7da2667e7510e4a46ebc46e1 # v1
+
+      - uses: julia-actions/julia-runtest@b67f25cc97aa716a662fc9586ea6d5d7bf6c2da3 # v1

Project.toml

@@ -13,6 +13,7 @@ DyadInterface = "99806f68-afab-45ca-9d8c-ceff6bc61f54"
 ElectricalComponents = "3fc4c519-469d-4ff5-98c9-d4d06b1514a2"
 HydraulicComponents = "b6ec4bdb-baef-4449-8ab3-f7a9cc7f1342"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
+LinearMPC = "82e1c212-e1a2-49d2-b26a-a31d6968e3bd"
 MPCComponents = "aba2bcdf-b216-4bf2-af71-aceb9999ebd9"
 Markdown = "d6f4376e-aef5-505a-96c1-9c027394607a"
 ModelingToolkit = "961ee093-0014-501f-94e3-6117800e7a78"
@@ -30,13 +31,26 @@ SynchJulia = "a1b2c3d4-5e6f-7a8b-9c0d-e1f2a3b4c5d6"
 SynchToolkit = "f500ffa8-9682-42ac-8d34-0ca7926c2e94"
 ThermalComponents = "084a29d9-4f48-4925-95b4-b7d4c79725f4"
 TranslationalComponents = "42acc370-819b-4877-bdd6-eb3294461f5d"
+acados_jll = "49ddb18e-ca18-5f65-a4dd-7588daaac186"
+tera_renderer_jll = "73a839a6-9370-58f6-a727-029f0c62a9a5"
 
 # MPCComponents is not registered, and the MPC stack it needs is not all released. What each pin
 # is for:
 #
 #   MPCComponents  its `main`, which has to include JuliaComputing/MPCComponents.jl#31: without it
 #                 the multibody prediction model is rejected as time-varying.
-#   LinearMPC, acados JLLs  unregistered forks MPCComponents builds on.
+#   LinearMPC, acados JLLs  unregistered forks MPCComponents builds on. They are listed under
+#                 `[deps]` rather than `[extras]` even though nothing here calls them: a `[sources]`
+#                 entry binds only for a package the active project depends on, so with them in
+#                 `[extras]` the resolver had no version to place for `acados_jll`, which
+#                 MPCComponents requires, and every `Pkg.resolve` failed with `has no known
+#                 versions`. LinearMPC failed more quietly still -- it is registered, so the
+#                 resolver silently took the release in place of the fork with the disturbance
+#                 cross term.
+#
+# MultibodyComponents is no longer taken from a local checkout: the registered 0.2.4 resolves
+# against the rest of this stack, so a sibling `../MultibodyComponents` is no longer needed and
+# this environment resolves from nothing but the registries and the sources below.
 #
 # DiscreteComponents is no longer pinned either: 0.4.0 has both `AlphaBetaGammaFilter`, the velocity
 # estimator (JuliaComputing/DiscreteComponents.jl#119), and `Latest`, the clock transition (#121),
@@ -55,7 +69,6 @@ TranslationalComponents = "42acc370-819b-4877-bdd6-eb3294461f5d"
 [sources]
 LinearMPC = {rev = "feat/disturbance-cross-term", url = "https://github.com/baggepinnen/LinearMPC.jl.git"}
 MPCComponents = {rev = "main", url = "https://github.com/JuliaComputing/MPCComponents.jl"}
-MultibodyComponents = {path = "../MultibodyComponents"}
 acados_jll = {rev = "main", url = "https://github.com/baggepinnen/acados_jll.jl"}
 tera_renderer_jll = {rev = "main", url = "https://github.com/baggepinnen/tera_renderer_jll.jl"}
 
@@ -69,11 +82,12 @@ DyadInterface = "7.1.0"
 ElectricalComponents = "2.2.1"
 HydraulicComponents = "2.0.3"
 LinearAlgebra = "1.12.0"
+LinearMPC = "0.10"
 MPCComponents = "0.1.0"
 Markdown = "1"
 ModelingToolkit = "11"
 Moshi = "0.3"
-MultibodyComponents = "0.2.2"
+MultibodyComponents = "0.2.4"
 OrderedCollections = "1.8"
 OrdinaryDiffEqDefault = "2"
 PrecompileTools = "1"
@@ -86,18 +100,17 @@ SynchJulia = "0.8"
 SynchToolkit = "0.5.1"
 ThermalComponents = "2.0.5"
 TranslationalComponents = "2.4.2"
+acados_jll = "0.6.0"
+tera_renderer_jll = "0.2"
 
 [dyad]
 kernel = "2.3.0"
 
 [extras]
 CSV = "336ed68f-0bac-5ca0-87d4-7b16caf5d00b"
 DataFrames = "a93c6f00-e57d-5684-b7b6-d8193f3e46c0"
-LinearMPC = "82e1c212-e1a2-49d2-b26a-a31d6968e3bd"
 Plots = "91a5bcdd-55d7-5caf-9e0b-520d859cae80"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
-acados_jll = "49ddb18e-ca18-5f65-a4dd-7588daaac186"
-tera_renderer_jll = "73a839a6-9370-58f6-a727-029f0c62a9a5"
 
 [targets]
 test = ["CSV", "DataFrames", "Plots", "Test"]

README.md

@@ -30,9 +30,9 @@ pkg> add ModelingToolkit MultibodyComponents SynchToolkit OrdinaryDiffEqDefault
 ### The simulation-only branch
 
 Those commands resolve `main` only for someone who can reach the repositories the MPC
-controller is pinned to: MPCComponents is not registered, and `[sources]` names branch
-a branch build of LinearMPC, two unregistered acados JLLs and one local checkout
-(see [Environment](#environment) below). Everyone
+controller is pinned to: MPCComponents is not registered, and `[sources]` names a branch
+build of LinearMPC and two unregistered acados JLLs (see [Environment](#environment) below).
+Everyone
 else installs the branch `simulation-only`, which is `main` as of the commit before the
 MPC work and resolves entirely to registered versions:
 
@@ -272,22 +272,28 @@ JuliaComputing/SynchToolkit.jl#186, the array clocked variables the MPC's output
 KeyError: key (control_system₊mpc₊u(t))[1] not found
 ```
 
-This branch checks in `Manifest.toml` and `test/Manifest.toml`, resolved against the pinned stack.
-MPCComponents comes from its GitHub `main` (which has to include
+MPCComponents comes from its GitHub `main`, which has to include
 [JuliaComputing/MPCComponents.jl#31](https://github.com/JuliaComputing/MPCComponents.jl/pull/31):
-without it the multibody prediction model is rejected as time-varying), and one package has to
-come from a local checkout recorded relative to this repository: `../MultibodyComponents` (the
-`~/.julia/dev` checkout; the registered release does not resolve against these pins). With both
-next to the repo,
+without it the multibody prediction model is rejected as time-varying. Nothing comes from a local
+checkout any more -- the registered MultibodyComponents 0.2.4 resolves against this stack -- so
 
 ```
 julia --project=path/to/QuanserComponents -e 'using Pkg; Pkg.instantiate()'   # the package
 julia --project=path/to/QuanserComponents/test test/hardware_mpc.jl           # the scripts
 ```
 
-is all it takes. For an environment of your own, copy the `[sources]` block (and the `[extras]`
-entries they refer to) from Project.toml into it. The first `using` precompiles the multibody and
-acados trees, a few minutes.
+is all it takes, from a clone and nothing else. `Manifest.toml` and `test/Manifest.toml` are still
+checked in, so an instantiate reproduces the stack the tests ran against, but both environments now
+also resolve from scratch.
+
+The unregistered packages are listed under `[deps]` rather than `[extras]`, which is what makes
+that true. A `[sources]` entry binds only for a package the active project depends on, so while
+`acados_jll` sat in `[extras]` the resolver had no version to place for it -- MPCComponents
+requires it -- and `Pkg.resolve` failed with `acados_jll has no known versions`. LinearMPC failed
+less visibly: it is registered, so the resolver quietly took the release instead of the fork
+carrying the disturbance cross term. For an environment of your own, copy the `[sources]` block
+and the `[deps]` entries it names. The first `using` precompiles the multibody and acados trees,
+a few minutes.
 
 ## Layout