CI on every push and pull request, the corpus as a job, templates, and 0.0.1

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,8 @@
+blank_issues_enabled: true
+contact_links:
+  - name: Security vulnerability
+    url: https://github.com/a85tract/RecastEngine/security/advisories/new
+    about: Report privately, never as a public issue; see SECURITY.md.
+  - name: Question or idea
+    url: https://github.com/a85tract/RecastEngine/discussions
+    about: Anything that is not a defect or a refused translation.

.github/ISSUE_TEMPLATE/defect.md

@@ -0,0 +1,19 @@
+---
+name: A defect
+about: A translation the gate passed that is wrong, a crash, or a plugin that misbehaves
+labels: bug
+---
+
+**What happened.** The command, its output, and what should have happened
+instead. If a gate passed something it should not have, this is the most
+serious kind of report there is: say which gate and attach the evidence
+manifest.
+
+```
+```
+
+**Reproduction.** The smallest input that shows it. A corpus case can be
+named instead of attached.
+
+**Version.** `recast version`, the Python version, and `gfortran --version`
+if an oracle was involved.

.github/ISSUE_TEMPLATE/refused-translation.md

@@ -0,0 +1,22 @@
+---
+name: A unit the engine refused or deferred
+about: The engine said what it would not translate; you think it should have
+labels: enhancement, area:frontend
+---
+
+**The Fortran.** The smallest module that reproduces it, inline or attached.
+A unit the corpus already carries can be named instead.
+
+**What the engine said.** The refusal or the deferred block, copied from
+`recast run` or from the evidence manifest -- the exact text, since a
+refusal names its reason and the reason is what is being discussed.
+
+```
+```
+
+**What you expected.** The NumPy the rule should have produced, if you have
+an opinion, and why it is bit-exact with the Fortran. A rule that is added
+here has to hold against f2py on the corpus, so the argument is the
+contribution.
+
+**Version.** `recast version` and the Python version.

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,10 @@
+**What this changes, and the evidence.** A new verifier shows the numbers
+it produces; a new transform shows what lands in `deferred` as well as what
+translates; a rule shows the corpus unit it holds on. See CONTRIBUTING.md,
+"What a good PR looks like".
+
+**Checks.**
+
+- [ ] `tools/ci_local.sh` passes, or the CI on this pull request does
+- [ ] every commit is signed off (`git commit -s`), and I am in `CLA-SIGNATORIES.md`
+- [ ] no name or path from CONTRIBUTING.md's "already taken" list is added

.github/workflows/ci.yml

@@ -1,14 +1,9 @@
 name: ci
 
-# Automatic runs are off while the Actions allowance is exhausted, and as of
-# 2026-08-21 the workflow itself is disabled on GitHub (`gh workflow disable`),
-# so even a manual dispatch is refused. The allowance returns in September.
-#
-# To restore, both halves are needed and only one of them is visible here:
-#   gh workflow enable ci.yml
-#   ...then put back the push (main) and pull_request triggers below.
-# Until then, tools/ci_local.sh runs these same jobs from the same file.
 on:
+  push:
+    branches: [main]
+  pull_request:
   workflow_dispatch:
 
 permissions:
@@ -78,6 +73,23 @@ jobs:
       # rather than skip for want of fparser.
       - run: uv run --python ${{ matrix.python }} --extra dev --extra fortran --extra translate pytest --cov=recast
 
+  # The twelve corpus cases, run rather than trusted: corpus/baseline.json is
+  # what the engine says about them, and a change in what it says is a change
+  # in the engine, which this job makes visible as a diff rather than letting
+  # the file lag behind main. One minute; the cases are small on purpose.
+  corpus:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          submodules: true
+      - uses: astral-sh/setup-uv@v3
+      - run: sudo apt-get update && sudo apt-get install -y gfortran
+      - run: uv python install 3.11
+      - run: uv run --extra dev --extra fortran --extra translate --extra verify python tools/corpus.py run
+      - name: The baseline is what main says
+        run: git diff --exit-code -- corpus/baseline.json
+
   # The contract, run rather than read. This is the suite an out-of-tree
   # extension has to pass to be an extension and not a fork, so the engine
   # running it against its own plugins is the minimum: a check nobody's plugins
@@ -99,7 +111,7 @@ jobs:
       # transforms also have to be *constructible*: determinism is an attribute
       # of an instance, and a factory that cannot run leaves the agentic-gate
       # rule unable to judge anything.
-      - run: uv run --extra dev --extra fortran --extra translate --extra verify --extra numba --extra jax pytest conformance/ -rs
+      - run: uv run --extra dev --extra fortran --extra translate --extra verify pytest conformance/ -rs
 
   spine:
     runs-on: ubuntu-latest
@@ -132,29 +144,3 @@ jobs:
       # the missing extra is reported when it is asked to parse, not before.
       - run: uv run --no-dev recast plugins | grep -qx "  fortran"
 
-  # The port spine: frontend -> port.jax -> numpy-anchor -> differential.tolerance.
-  # This is the only job that installs jax, and until it existed
-  # tests/test_port_spine.py skipped in CI and had run on one laptop only --
-  # the same gap the spine job above was created to close for the f2py chain.
-  #
-  # It needs no compiler, and that is a property of the oracle rather than a
-  # convenience: numpy-anchor re-derives the reference from the same Facts, so
-  # nothing here builds Fortran. The link this job cannot check is the one that
-  # does -- the NumPy translation being bit-exact against the Fortran, which is
-  # the spine job's claim above.
-  port-spine:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v5
-      - uses: astral-sh/setup-uv@v3
-      - run: uv python install 3.11
-      - run: uv run --extra dev --extra fortran --extra translate --extra jax pytest tests/test_port_spine.py -v
-      # The public form of the same claim, on sources anyone can read.
-      - run: uv run --extra fortran --extra translate --extra jax recast run port corpus/toy_physics --config corpus/toy_physics/port.json --summary corpus/toy_physics/port-verification.json
-      # No `git diff --exit-code` here, deliberately, and the asymmetry with the
-      # spine job is the point. Bit-exactness is a device-independent claim, so
-      # its summary is committed *and* checked byte for byte. A ULP count is
-      # not: XLA's CPU backend does not promise the same last bit on x86 as on
-      # arm64, so the committed port-verification.json is a record for a reader
-      # without JAX, not a regression signal. What this job gates on is the
-      # verdict -- the run exits non-zero unless every unit passes its gate.

.github/workflows/hygiene.yml

@@ -4,15 +4,12 @@ name: hygiene
 # after. Migrating 10k lines out of two NCAR-bound private repositories is
 # exactly the situation where a site path or an allocation account slips in.
 
-# Automatic runs are off while the Actions allowance is exhausted, and as of
-# 2026-08-21 the workflow itself is disabled on GitHub (`gh workflow disable`),
-# so even a manual dispatch is refused. The allowance returns in September.
-#
-# To restore, both halves are needed and only one of them is visible here:
-#   gh workflow enable hygiene.yml
-#   ...then put back the push, pull_request, and Monday 06:00 UTC schedule triggers below.
-# Until then, tools/ci_local.sh runs these same jobs from the same file.
 on:
+  push:
+    branches: [main]
+  pull_request:
+  schedule:
+    - cron: "0 6 * * 1"
   workflow_dispatch:
 
 permissions:

.github/workflows/release.yml

@@ -0,0 +1,42 @@
+name: release
+
+# A published GitHub release of a `v*` tag builds the sdist and wheel from
+# that tag and uploads them to PyPI through trusted publishing (OIDC): no
+# token lives anywhere. PyPI has to know this repository and this workflow
+# as the publisher of `recast-engine` first; until it does, the publish step
+# says so and the built files are still attached to the run.
+on:
+  release:
+    types: [published]
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v5
+      - uses: astral-sh/setup-uv@v3
+      - run: uv python install 3.11
+      - run: uv build
+      - run: uv run --with twine --no-project twine check dist/*
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist
+          path: dist/
+
+  publish:
+    needs: build
+    if: github.event_name == 'release'
+    runs-on: ubuntu-latest
+    environment: pypi
+    permissions:
+      id-token: write
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: dist
+          path: dist/
+      - uses: pypa/gh-action-pypi-publish@release/v1

README.md

@@ -27,7 +27,7 @@ the corpus of numerical libraries under `corpus/` is what it is proven on.
 
 ```console
 $ recast doctor
-recast 0.0.1.dev0  python 3.11.16
+recast 0.0.1  python 3.11.16
 ```
 
 The remaining `doctor` output is a live plugin count and registry inventory;

docs/cli.md

@@ -2,7 +2,7 @@
 
 ```console
 $ recast doctor
-recast 0.0.1.dev0  python 3.11.16
+recast 0.0.1  python 3.11.16
 ```
 
 `doctor` then prints the live plugin count and registry inventory. The count is

docs/getting-started.md

@@ -81,7 +81,7 @@ Now check the installation:
 
 ```console
 $ recast doctor
-recast 0.0.1.dev0  python 3.11.16
+recast 0.0.1  python 3.11.16
 24 plugin(s) registered across 10 kinds
 ```
 

docs/reading-the-evidence.md

@@ -223,7 +223,7 @@ deliberately leaves out. The differential one for `satvap`:
     "files": ["satvap_constants.py", "satvap_numpy.py"]
   },
   "reference": {"oracle": "f2py-golden", "key": "f2py:satvap:8c6c49ccd27d2248"},
-  "environment": {"engine": "recast 0.0.1.dev0", "platform": "macOS-26.5.2-arm64-arm-64bit", "python": "3.11.16"},
+  "environment": {"engine": "recast 0.0.1", "platform": "macOS-26.5.2-arm64-arm-64bit", "python": "3.11.16"},
   "cc_test": {"commit": "unknown", "version": "unknown"},
   "cases": [],
   "result": {

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "recast-engine"
-version = "0.0.1.dev0"
+version = "0.0.1"
 description = "An agentic engine for modernizing legacy scientific software"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -134,7 +134,10 @@ line-length = 100
 src = ["src", "tests"]
 # corpus/ is twelve upstream Fortran libraries vendored as submodules, to be
 # translated, not styled. A few carry Python of their own; it is theirs.
-extend-exclude = ["corpus"]
+# docs/ quotes code the engine emitted, parentheses and all; the formatter
+# rewriting a fenced block there would misquote it. So the docs are not
+# formatted, though their prose is linted like anything else.
+extend-exclude = ["corpus", "docs/*.md", "*.md"]
 
 [tool.ruff.lint]
 select = ["E", "F", "I", "N", "UP", "B", "S", "RUF"]
@@ -155,6 +158,17 @@ packages = ["recast"]
 strict = true
 warn_unreachable = true
 
+# Modules an installation may lack -- the optional recipes, the optional
+# conformance cases, the flat adapters -- are imported behind a guard; where
+# they are absent the import is not an error.
+[[tool.mypy.overrides]]
+module = [
+    "recast.recipes.optional",
+    "recast.conformance.builtin_optional",
+    "recast.fortran.flatten",
+]
+ignore_missing_imports = true
+
 # fparser2 ships no py.typed marker. It is confined to recast.fortran, and the
 # node types it returns are untyped trees anyway -- the frontend's own boundary
 # (Unit, Facts) is where the types have to hold.

src/recast/__init__.py

@@ -11,7 +11,7 @@
 
 from __future__ import annotations
 
-__version__ = "0.0.1.dev0"
+__version__ = "0.0.1"
 
 from recast.engines import (
     ArtifactContract,

tests/test_f2py_oracle.py

@@ -1982,7 +1982,7 @@ def test_a_scalar_logical_inout_goes_through_the_wrapper_as_an_integer() -> None
 """
 
 
-@pytest.mark.skipif(GFORTRAN is None, reason="needs gfortran")
+@pytest.mark.skipif(GFORTRAN is None or not MESON, reason="needs gfortran and meson")
 def test_a_complex_valued_subprogram_is_compared_on_both_parts(tmp_path: Path) -> None:
     """A complex result or argument was ``unsupported declared dtype(s)`` and
     left uncompared, while the emitter built it as float64 (#20). Complex
@@ -2031,7 +2031,7 @@ def test_a_complex_valued_subprogram_is_compared_on_both_parts(tmp_path: Path) -
 """
 
 
-@pytest.mark.skipif(GFORTRAN is None, reason="needs gfortran")
+@pytest.mark.skipif(GFORTRAN is None or not MESON, reason="needs gfortran and meson")
 def test_a_reference_reading_outside_its_array_declines_the_draw_by_name(tmp_path: Path) -> None:
     """A subscript outside the array is not a value the source computes: the
     reference reads whatever memory sits beside the buffer in its process,
@@ -2076,7 +2076,7 @@ def test_a_reference_reading_outside_its_array_declines_the_draw_by_name(tmp_pat
 """
 
 
-@pytest.mark.skipif(GFORTRAN is None, reason="needs gfortran")
+@pytest.mark.skipif(GFORTRAN is None or not MESON, reason="needs gfortran and meson")
 def test_an_error_stop_in_the_reference_is_a_report_not_a_dead_run(tmp_path: Path) -> None:
     """``error stop`` in the compiled reference is ``exit()`` in whatever
     process imported it: no report, no summary, every other unit's verdict