Align with CTBase 0.29 / CTModels 0.18 / ExaModels 0.12

.github/workflows/CI.yml

@@ -10,16 +10,16 @@ on:
 
 jobs:
   # Job for the GitHub hosted runners (ubuntu, macos, windows)
-  test-github-runner:
+  test-cpu-github:
     # A 'labeled' event fires once per label added, and re-evaluates against the PR's
     # *current* full label set — so adding several labels in a row would otherwise
     # re-run this job on every subsequent label add. Only react to 'labeled' when it's
     # this specific label; for other trigger types (synchronize/reopened), fall back to
     # checking the current label set as before.
     if: >
       github.event_name != 'pull_request' ||
-      (github.event.action == 'labeled' && github.event.label.name == 'github-runner') ||
-      (github.event.action != 'labeled' && contains(github.event.pull_request.labels.*.name, 'github-runner'))
+      (github.event.action == 'labeled' && github.event.label.name == 'run ci github-runner') ||
+      (github.event.action != 'labeled' && contains(github.event.pull_request.labels.*.name, 'run ci github-runner'))
     uses: control-toolbox/CTActions/.github/workflows/ci.yml@main
     with:
       runs_on: '["ubuntu-latest", "macos-latest", "windows-latest"]'
@@ -28,17 +28,17 @@ jobs:
     secrets:
       SSH_KEY: ${{ secrets.SSH_KEY }}
 
-  # Job for the self-hosted runner kkt (GPU/CUDA)
-  test-kkt-runner:
-    # See the comment on test-github-runner above.
+  # Job for the self-hosted runner occidata (GPU/CUDA)
+  test-gpu-occidata:
+    # See the comment on test-cpu-github above.
     if: >
       github.event_name != 'pull_request' ||
-      (github.event.action == 'labeled' && github.event.label.name == 'kkt-runner') ||
-      (github.event.action != 'labeled' && contains(github.event.pull_request.labels.*.name, 'kkt-runner'))
+      (github.event.action == 'labeled' && github.event.label.name == 'run ci occidata-runner') ||
+      (github.event.action != 'labeled' && contains(github.event.pull_request.labels.*.name, 'run ci occidata-runner'))
     uses: control-toolbox/CTActions/.github/workflows/ci.yml@main
     with:
       versions: '["1"]'
-      runs_on: '[["kkt"]]'
+      runs_on: '[["occidata"]]'
       runner_type: 'self-hosted'
       use_ct_registry: true
     secrets:

AGENTS.md

@@ -6,14 +6,23 @@ Quick-reference for any agent working on this repository.
 
 ## Repository Layout
 
-CTParser is a single flat module — no submodule split, no `ext/`.
+CTParser is a single flat module — no submodule split.
 
 ```text
 src/        # onepass.jl (the `@def` parser), defaults.jl, initial_guess.jl, utils.jl
+ext/        # CTParserExaModels.jl only — see below
 test/       # Test suite: flat files (not test/suite/)
 docs/       # Documentation site (DocumenterVitepress)
 ```
 
+`ext/` holds exactly one extension, and it is **temporary**. ExaModels ships the
+linear-algebra glue its expression nodes need (`dot`, `*` on node vectors/matrices,
+`det`, `norm`, `Null` zero elimination) but never declares that extension in its
+`[extensions]`, so Julia never loads it. Until upstream wires it up, CTParser carries
+a port of it, triggered by `ExaModels` + `LinearAlgebra` weak dependencies. Delete
+the file and the two weakdeps when upstream fixes it — see
+[#325](https://github.com/control-toolbox/CTParser.jl/issues/325).
+
 ---
 
 ## Developer Resources

BREAKING.md

@@ -0,0 +1,153 @@
+<!-- markdownlint-disable MD024 -->
+# Breaking Changes
+
+Breaking changes in CTParser releases, and how to migrate. Tracked from the 0.8.15
+baseline onward; see [CHANGELOG.md](CHANGELOG.md) for the full record.
+
+## [0.9.5-beta] - 2026-08-31
+
+No breaking changes in this release.
+
+- Constraint bounds that depend on `v`, the state, the control or the time
+  ([#343](https://github.com/control-toolbox/CTParser.jl/issues/343)) are now rejected
+  with a `CTBase.ParsingError` naming the cause. Such definitions already failed to
+  build — previously with an internal `UndefVarError` leaked from generated code — so no
+  working `@def` changes. Rewrite the relation as a functional constraint, e.g.
+  `x₂(0) - v == 0` instead of `x₂(0) == v`.
+- Trace mode (`@def name … end true`) prints the parsed model once instead of twice
+  ([#344](https://github.com/control-toolbox/CTParser.jl/issues/344)). Trace output only;
+  the returned model and the public API are unchanged.
+
+## [0.9.3-beta] - 2026-08-30
+
+No breaking changes in this release.
+
+- Documentation-only fix ([#341](https://github.com/control-toolbox/CTParser.jl/issues/341)):
+  examples in docstrings now use static `julia` fences so they render correctly when
+  transcluded into consumer documentation. Runtime behavior and the public API are unchanged.
+
+## [0.9.2-beta] - 2026-08-28
+
+No breaking changes in this release.
+
+- Test-suite and release metadata only ([#339](https://github.com/control-toolbox/CTParser.jl/issues/339)):
+  GPU runner capability detection now recognises both the `kkt` and `occidata`
+  self-hosted runners, the GPU test tiers skip visibly with `Test.@test_skip` on
+  CPU runners, and a new `test/test_environment_contract.jl` enforces both. `src/`,
+  `ext/` and `.github/workflows/CI.yml` are unchanged.
+
+## [0.9.1-beta] - 2026-08-26
+
+No breaking changes in this release.
+
+- Structured CTBase errors no longer print an unrelated `Line n: ...` message on
+  `stdout`; their existing structured diagnostics and exception types are unchanged.
+
+## [0.9.0] - 2026-08-25
+
+### The `:exa` emission requires ExaModels ≥ 0.12
+
+ExaModels 0.12 deleted `src/deprecated.jl` and with it the mutable builder API —
+`variable`, `parameter`, `subexpr`, `constraint!`, `LegacyExaCore`. The code `@def`
+generates for `backend = :exa` now uses the functional API instead, where `add_var` /
+`add_con` / `add_obj` return `(new_core, result)` and the core is threaded through the
+build.
+
+**Who is affected**: anyone pinning ExaModels below 0.12 while using the `:exa` backend.
+Nothing else — `@def` problem definitions, the `:fun` backend, and the shape of the
+`(model, getter)` pair `build_examodel` returns are all unchanged.
+
+There is no compatibility shim, and adding one is not practical: CTParser does not
+depend on ExaModels (the module is reached through `prefix_exa()`), so it cannot detect
+the version at macro-expansion time and pick an emission.
+
+**Migration**: raise your ExaModels bound.
+
+```toml
+# Before
+ExaModels = "0.9"
+
+# After
+ExaModels = "0.12"
+```
+
+For reference, what changed in the generated code:
+
+```julia
+# Before — emitted against ExaModels 0.9–0.11
+c = ExaModels.ExaCore(base_type; backend, minimize)
+x = ExaModels.variable(c, n, 0:grid_size; lvar, uvar, start)
+ExaModels.constraint(c, expr for j in 0:grid_size-1; lcon, ucon)
+ExaModels.objective(c, expr for j in 0:grid_size-1)
+
+# After — emitted against ExaModels 0.12
+c = ExaModels.ExaCore(base_type; backend, minimize)   # unchanged
+c, x = ExaModels.add_var(c, n, 0:grid_size; lvar, uvar, start)
+c, _ = ExaModels.add_con(c, expr for j in 0:grid_size-1; lcon, ucon)
+c, _ = ExaModels.add_obj(c, expr for j in 0:grid_size-1)
+```
+
+### Untyped `String` errors are now `CTException` subtypes
+
+Seventeen sites in `src/onepass.jl` threw a bare `String`. Callers could not dispatch on
+them, and `showerror` fell back to `show`, rendering the message wrapped in quotes rather
+than the formatted block every other error in the ecosystem produces.
+
+**Who is affected**: code that catches these errors and compares the caught value to a
+string, or that matches on `String`.
+
+| what | type now |
+| --- | --- |
+| unknown numerical scheme | `CTBase.IncorrectArgument` |
+| lower/upper bound length mismatch | `CTBase.PreconditionError` |
+| bound lengths vs. the constrained range | `CTBase.PreconditionError` |
+| unknown value for the getter's `val` keyword | `CTBase.IncorrectArgument` |
+| unknown parsing backend | `CTBase.IncorrectArgument` |
+| `:fun` cannot be activated or deactivated | `CTBase.PreconditionError` |
+
+The split follows the Handbook's rule: a single argument's value out of domain is
+`IncorrectArgument`; a relational, state or timing contract is `PreconditionError`. A
+bound-length mismatch relates two things — the bounds to each other, or to the
+constrained range — hence `PreconditionError`. `:fun` is a perfectly valid backend name;
+what is forbidden is toggling it, which is a state contract, not a bad value.
+
+**Migration**:
+
+```julia
+# Before
+try
+    solve(ocp, :exa; scheme=:gauss_legendre_2)
+catch e
+    e == "unknown numerical scheme: gauss_legendre_2 (possible choices are ...)"
+end
+
+# After
+try
+    solve(ocp, :exa; scheme=:gauss_legendre_2)
+catch e
+    e isa CTBase.IncorrectArgument
+end
+```
+
+`@test_throws String` in a downstream test suite becomes `@test_throws
+CTBase.IncorrectArgument` or `@test_throws CTBase.PreconditionError`, per the table.
+
+## Non-breaking note (0.9.0)
+
+- **New `CTParserExaModels` extension.** Additive, and triggered only when both
+  `ExaModels` and `LinearAlgebra` are loaded — the main module gains no dependency. It
+  restores the linear algebra on ExaModels expression nodes that a dynamics like
+  `∂(x)(t) == A * x(t) + B * u(t)` needs, which ExaModels 0.12 ships but never registers
+  in its own `[extensions]`. **No migration required.** It is temporary: when upstream
+  wires its extension up ([madsuite-org/ExaModels.jl#323](https://github.com/madsuite-org/ExaModels.jl/issues/323)),
+  the file and the two weak dependencies go away, with no user-visible change either way.
+
+- **Compat bounds** raised for CTBase (`0.29`), CTModels (`0.18`), OrderedCollections
+  (`1, 2`) and Parameters (`0.13`), and widened for CUDA (`5, 6`), MadNLP (`0.9, 0.10`)
+  and MadNLPGPU (`0.8, 0.10`). **No breaking change**: none of these required a source
+  change, and the `:fun` test groups stayed green at 1054/1054 across the bump.
+
+## [0.8.15] - 2026-04-21 — baseline
+
+Reference version. Breaking changes are tracked from here onward; use `git log` for
+earlier history.