openqarp 0.1.0__tar.gz

openqarp-0.1.0/.claude/skills/plan/SKILL.md

@@ -0,0 +1,49 @@
+---
+name: plan
+description: Scaffold a OpenQARP contribution plan in the house format under docs/contributions/. Use at the start of any standard- or structural-tier contribution, ideally during the contributor+reviewer sit-down.
+---
+
+# /plan — scaffold a contribution plan
+
+You are helping draft a plan that a reviewer will green-light in a Draft PR
+before implementation starts. Plan and implementation share one PR, so the
+green-lit plan is the contract the rest of that same PR gets diffed against
+— precision here saves review time later.
+
+## Before writing anything
+
+1. Read `docs/contributions/README.md` (tiers, loop, oracle rule).
+2. Read `docs/contracts/qarp_conventions.md` — at minimum
+   §13 (blocks), §14 (engines/devices), §15 (repo), §17 (symbols), plus any
+   section the topic touches. The plan must FIT these conventions; if it
+   needs to change one, that is a deliberate edit to the conventions doc,
+   named explicitly in the plan's Design section.
+3. Skim the open rows (🚧, 📋) of the index in `docs/contributions/README.md`
+   for adjacent work the plan should not collide with.
+
+## Then
+
+4. Copy `docs/contributions/_template.md` to
+   `docs/contributions/<topic>_plan.md` (short snake_case topic).
+5. Fill every section with what was agreed in the discussion — do not
+   invent scope. Rules that matter most:
+   - **Scope** is an explicit file list. The conformance review checks the
+     implementation commits touch nothing outside it without a declared
+     deviation.
+   - **API sketch** concrete enough to diff an implementation against.
+   - **Test plan**: one row per behaviour, and the oracle column is
+     mandatory — an independent reference (analytic value, published
+     number, openfermion/scipy), never the implementation's own output.
+     Naming oracles now, before code exists, is the point.
+   - **Phases** individually landable; date items as they complete.
+6. Add the plan's row to the index table in `docs/contributions/README.md`
+   (status 📋 Planned).
+7. Remind the author: branch, then open a **Draft** PR whose first commit is
+   the plan, using the default PR template (fill the *Plan* section, leave
+   the *Implementation* section for later). The reviewer's green-light in
+   that Draft PR is the design approval — do not start implementing before
+   it, and do not open a second PR.
+
+Ask the user for any decision the discussion has not settled rather than
+choosing silently — unresolved decisions belong in the plan as explicit
+open items, not as your guesses.

openqarp-0.1.0/.claude/skills/plan-review/SKILL.md

@@ -0,0 +1,61 @@
+---
+name: plan-review
+description: Conformance-review a OpenQARP implementation against the green-lit plan in docs/contributions/ (same PR — the plan is its first commit). Reviewer tool — produces a report for the human reviewer, never a merge decision.
+---
+
+# /plan-review — diff an implementation against its plan
+
+Input: a plan file in `docs/contributions/` and the implementation branch
+(or PR diff). If either is not given, ask. Your output is INPUT to the
+human reviewer — you never approve or block; write findings, not verdicts.
+
+## Procedure
+
+1. Read the plan fully: Scope, Design, API sketch, Test plan (oracle
+   column), Phases, Deviations log.
+2. Get the change set: `git diff <target>...<branch> --stat` and then the
+   full diff of the files that matter. Target is `develop` unless told
+   otherwise. The plan file is part of this diff by design (it is the
+   branch's first commit) — never report it as out-of-scope.
+3. **Check the plan was not rewritten to match the code.** Plan and
+   implementation share a branch, so the plan file is editable after the
+   green-light — an easy way to make deviations disappear. Diff the plan
+   file across its own history on the branch
+   (`git log -p <target>..<branch> -- <plan file>`). Post-green-light edits
+   to Scope, Design, API sketch or the Test plan's oracle column are a
+   finding unless the PR's "Deviations from plan" section declares them.
+   Later phase-box dates, folded-back deviations and status flips are
+   expected and fine.
+4. Check, in order:
+   - **Scope**: every touched file is in the plan's Scope list, or the
+     deviation is declared in the PR's "Deviations from plan" section.
+     Undeclared out-of-scope changes are the highest-priority finding.
+   - **Design & API**: implementation matches the sketch — names,
+     signatures, ownership/layering. Note drift even when it is an
+     improvement; declared-and-folded-back is the rule, not silence.
+   - **Test plan**: each row of the plan's test table exists in the diff,
+     and each named oracle actually appears in the assertions. Flag any
+     test that asserts the implementation against itself (round-trip
+     posing as an oracle) — per conventions, round-trips are additional,
+     never the oracle. Ask of each test: would it fail if the feature were
+     wrong?
+   - **Conventions**: `qarp_conventions.md` §13–§17 compliance for the
+     touched areas — LSB endianness, radians, SimpleBlock/CompositeBlock
+     inheritance (never `Block` internally), sorted-tuple `.symbols` and
+     `parameter_map` usage, SDK imports only in adapters, `zip(strict=)` justified.
+   - **Plan hygiene**: deviations folded back into the plan file; phase
+     checkboxes dated; status/index row updated if this PR completes the
+     plan.
+5. Do not relitigate the design itself — the Draft-PR green-light approved
+   it. If the design looks wrong in hindsight, flag it as exactly that: a
+   question for the humans, separate from conformance findings.
+
+## Report format
+
+- **Verdict-free summary** (2–3 sentences: what the PR does, overall
+  conformance impression).
+- **Findings**, most severe first, each with file:line, what the plan says,
+  what the code does.
+- **Unverifiable items** — anything you could not check (e.g. behaviour
+  needing a run) listed explicitly, never silently skipped.
+- **Suggested reviewer focus** — the 2–3 places human judgment matters most.

openqarp-0.1.0/.gitattributes

@@ -0,0 +1,3 @@
+# Vendored upstream sources: keep them out of language stats and diffs.
+cpp/libqarpx/third_party/csim/** linguist-vendored
+cpp/libqarpx/third_party/eigen/** linguist-vendored

openqarp-0.1.0/.github/ISSUE_TEMPLATE/bug.md

@@ -0,0 +1,23 @@
+---
+name: Bug
+about: Something is wrong, including a plausible-but-wrong number
+---
+
+## What is the bug about?
+
+<!-- Observed vs expected behaviour. Plausible-but-wrong numbers count as
+     bugs — include the reference value you expected and its source. -->
+
+## OpenQARP version / commit
+
+<!-- `qarp.__version__` or commit hash; note if qarpx was rebuilt locally. -->
+
+## Minimal example to reproduce
+
+```python
+```
+
+## Stacktrace / output
+
+```
+```

openqarp-0.1.0/.github/ISSUE_TEMPLATE/feature.md

@@ -0,0 +1,21 @@
+---
+name: Feature
+about: Propose a new block, algorithm, engine or capability
+---
+
+## What does it do?
+
+<!-- Describe the feature. -->
+
+## Why do we need it in OpenQARP?
+
+<!-- Use case, physics/algorithmic motivation. -->
+
+## References
+
+<!-- Research articles, other packages, benchmarks. -->
+
+## Expected tier
+
+<!-- trivial / standard / structural — see docs/contributions/README.md.
+     Standard/structural features start with a plan before implementation. -->

openqarp-0.1.0/.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,59 @@
+<!-- One PR per contribution, two gates inside it: green-light the plan while
+     this PR is a Draft, then mark it ready for implementation review.
+     The loop: docs/contributions/README.md -->
+
+## Summary
+
+<!-- What and why, 2–4 sentences. -->
+
+## Plan
+
+<!-- One of the two lines below; delete the other.
+     Tier definitions: docs/contributions/README.md -->
+Plan: docs/contributions/<topic>_plan.md
+Tier: trivial — <!-- one-line justification -->
+
+Reviewer: @
+Sit-down held: <!-- yes / no — structural tier requires yes -->
+
+### Design summary
+
+<!-- 3–5 sentences: the problem, the chosen approach, what it deliberately
+     does NOT do. Details live in the plan file — don't duplicate. -->
+
+### Gate 1 — green-light checklist (while Draft)
+
+<!-- Trivial tier: delete this whole section. -->
+
+- [ ] House-format header complete (Status / Author / Reviewer / Date /
+      Tier / Scope / Branch)
+- [ ] Scope is an explicit file list
+- [ ] Test plan table has the oracle column filled (independent references
+      named BEFORE implementation exists)
+- [ ] Conventions checked (`qarp_conventions.md` §13–§17); any convention
+      change named explicitly in the Design section
+- [ ] Index row added in `docs/contributions/README.md` (📋 Planned)
+- [ ] **Green-lit by the reviewer** — implementation starts only after this;
+      flip the index row to 🚧 In progress
+
+## Deviations from plan
+
+<!-- Measured against the plan AS GREEN-LIT, not as later edited.
+     Leave EMPTY to claim "implemented exactly as green-lit" (reviewers check
+     this claim). Otherwise list each deviation; fold them back into the
+     plan file before merge — declared drift is fine, silent drift is not. -->
+
+## Gate 2 — implementation checklist (before marking ready)
+
+- [ ] Tests added/updated; every numerical assertion has an independent
+      oracle (analytic / published / openfermion / scipy — round-trips are
+      additional, never the oracle)
+- [ ] Example in `examples/` (standard/structural features)
+- [ ] Plan file updated: deviations folded back, phase boxes dated,
+      status + index row in `docs/contributions/README.md` (standard/structural)
+- [ ] `ruff check .`, `ruff format --check .`, mypy, `pytest` green locally
+
+<!-- Default expectation is untouched: a convention change is a deliberate act,
+     green-lit in the plan — never a side effect of implementation. -->
+- [ ] `qarp_conventions.md` **has not been edited** — or, if it has, the change
+      was named in the plan and green-lit

openqarp-0.1.0/.github/actions/build-wheel/action.yml

@@ -0,0 +1,35 @@
+name: Build the qarpx wheel
+description: >-
+  The pipeline's only qarpx compile.  Every Python job installs this wheel
+  instead of rebuilding.  ccache is keyed per commit with a prefix fallback, so
+  a C++-free change restores the newest cache and collapses to packaging.
+runs:
+  using: composite
+  steps:
+    - name: System build deps
+      shell: bash
+      run: sudo apt-get update && sudo apt-get install -y --no-install-recommends cmake ccache
+    - name: Restore ccache
+      uses: actions/cache@v5
+      with:
+        path: .ccache
+        key: ccache-wheel-${{ runner.os }}-${{ github.sha }}
+        restore-keys: ccache-wheel-${{ runner.os }}-
+    - name: Build wheel
+      shell: bash
+      env:
+        CCACHE_DIR: ${{ github.workspace }}/.ccache
+        CCACHE_MAXSIZE: 1G
+      run: |
+        ccache --zero-stats
+        python -m pip wheel . --no-deps --wheel-dir dist \
+          --config-settings=cmake.define.CMAKE_C_COMPILER_LAUNCHER=ccache \
+          --config-settings=cmake.define.CMAKE_CXX_COMPILER_LAUNCHER=ccache
+        # 0% on a C++-free change means the cache is not surviving between runs.
+        ccache --show-stats
+    - name: Upload wheel
+      uses: actions/upload-artifact@v6
+      with:
+        name: wheel
+        path: dist/*.whl
+        retention-days: 1

openqarp-0.1.0/.github/workflows/ci.yml

@@ -0,0 +1,294 @@
+# The per-change gate.  Order mirrors the legacy pipeline: lint first so a ruff
+# failure never costs a 17-minute compile; one wheel build feeds mypy, pytest
+# and docs; ctest builds its own tree (tests ON, which the wheel omits).  The
+# Pages deploy sits behind every test job, so docs never publish for a commit
+# whose tests did not pass; it also carries the coverage badge endpoint, so the
+# README badge shows the number this gate measured.  Tags run the full battery:
+# a release tag revalidates the exact tagged commit.
+name: ci
+
+on:
+  push:
+    branches: [develop, main, "release/**"]
+    tags: ["v*"]
+  pull_request:
+  workflow_dispatch:
+
+concurrency:
+  group: ci-${{ github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+
+env:
+  PYTHON_VERSION: "3.12"
+
+jobs:
+  lint:
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          fetch-depth: 0   # the large-files check diffs against the base commit
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+      - name: ruff (check + format)
+        run: |
+          # Pin must match pyproject [dev] and .pre-commit-config.yaml.
+          python -m pip install ruff==0.15.22
+          ruff --version
+          ruff check .
+          ruff format --check .
+      - name: README mirror is byte-identical
+        run: |
+          diff -u README.rst docs/source/getting_started.rst || {
+            echo "docs/source/getting_started.rst has drifted from README.rst — edit README, then copy it over the mirror."
+            exit 1
+          }
+      - name: Narrative-docs static gate
+        run: python scripts/ci/check_docs_code.py --static docs/source
+      - name: Notebook outputs are stripped
+        run: |
+          failed=""
+          while IFS= read -r nb; do
+            python3 -c "import json,sys;d=json.load(open('$nb'));sys.exit(1 if any(c.get('outputs') or c.get('execution_count') for c in d.get('cells',[])) else 0)" || failed="$failed\n  $nb"
+          done < <(git ls-files '*.ipynb')
+          if [ -n "$failed" ]; then
+            echo "ERROR: The following notebooks have outputs. Clear them before pushing:"
+            echo -e "$failed"
+            echo "Run: jupyter nbconvert --clear-output --inplace <notebook>"
+            exit 1
+          fi
+      - name: No large or forbidden files in the change
+        env:
+          # PR: the merge base GitHub computed.  Push: the previous tip, all
+          # zeros on a new branch or tag — then there is nothing to diff.
+          BASE: ${{ github.event_name == 'pull_request' && github.event.pull_request.base.sha || github.event.before }}
+        run: |
+          FORBIDDEN_EXTENSIONS="pkl npy npz data csv json log"
+          MAX_SIZE_KB=500
+          if [ -z "$BASE" ] || echo "$BASE" | grep -qE '^0+$'; then
+            echo "No base commit (new branch, tag or manual run) — nothing to diff, skipping."
+            exit 0
+          fi
+          git cat-file -e "${BASE}^{commit}" 2>/dev/null || git fetch --depth=50 origin "$BASE" 2>/dev/null || true
+          git cat-file -e "${BASE}^{commit}" 2>/dev/null || { echo "ERROR: base commit $BASE is not present and could not be fetched."; exit 1; }
+          failed=""
+          for file in $(git diff --name-only --diff-filter=AM "$BASE" HEAD); do
+            [ ! -f "$file" ] && continue
+            # tests/assets/ holds provenance-scripted test data: the extension
+            # ban is waived there, the size cap is not.
+            case "$file" in
+              tests/assets/*) ;;
+              *)
+                ext="${file##*.}"
+                for forbidden in $FORBIDDEN_EXTENSIONS; do
+                  [ "$ext" = "$forbidden" ] && { failed="$failed\n  $file (forbidden type: .$ext)"; break; }
+                done
+                ;;
+            esac
+            size_kb=$(du -k "$file" | cut -f1)
+            [ "$size_kb" -gt "$MAX_SIZE_KB" ] && failed="$failed\n  $file (${size_kb}KB exceeds ${MAX_SIZE_KB}KB limit)"
+          done
+          if [ -n "$failed" ]; then
+            echo "ERROR: The following files violate repository policies:"
+            echo -e "$failed"
+            echo "Forbidden types: ${FORBIDDEN_EXTENSIONS}; max file size: ${MAX_SIZE_KB}KB"
+            exit 1
+          fi
+
+  wheel:
+    needs: [lint]
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+      - uses: ./.github/actions/build-wheel
+
+  mypy:
+    needs: [wheel]
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+      - uses: actions/download-artifact@v7
+        with:
+          name: wheel
+          path: dist
+      - run: python -m pip install "$(ls dist/*.whl)[full-dev]"
+      - run: |
+          mypy --version
+          mypy qarp/ tests/ --disable-error-code=import-untyped --disable-error-code=method-assign
+
+  pytest:
+    needs: [wheel]
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+      - uses: actions/download-artifact@v7
+        with:
+          name: wheel
+          path: dist
+      - run: python -m pip install "$(ls dist/*.whl)[full-dev]"
+      # No --cov-fail-under here: this job answers "did the tests pass"; the
+      # coverage job reads the artefacts and answers "did coverage regress".
+      - run: pytest tests/ --cov --cov-report=term --cov-report=xml:coverage.xml
+      - uses: actions/upload-artifact@v6
+        if: always()
+        with:
+          name: coverage
+          path: |
+            coverage.xml
+            .coverage
+          include-hidden-files: true   # .coverage is the raw data file the ratchet needs
+          retention-days: 7
+
+  coverage:
+    needs: [pytest]
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    env:
+      # RATCHET FLOOR — only ever rises (qarp_conventions.md §15).  Lowering it
+      # to make a run pass is the violation the gate exists to prevent.
+      COVERAGE_FLOOR: "94.0"
+      # Patch gate: lines THIS change touched; PR-only, since off a PR there
+      # is no base and "the lines this change touched" has no meaning.
+      PATCH_COVERAGE_FLOOR: "80"
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          fetch-depth: 0   # diff-cover needs the merge base with the target branch
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+      - uses: actions/download-artifact@v7
+        with:
+          name: coverage
+          path: .
+      - run: python -m pip install "diff-cover==10.5.0" "coverage==7.15.4"
+      - name: Artefacts arrived
+        # A missing coverage.xml makes diff-cover report "no lines with coverage
+        # information" and exit 0 — indistinguishable from a non-Python change.
+        run: |
+          test -s coverage.xml || { echo "coverage.xml missing/empty — pytest artefact did not arrive"; exit 1; }
+          test -s .coverage   || { echo ".coverage missing/empty — cannot enforce the ratchet floor"; exit 1; }
+      - name: Ratchet floor
+        run: coverage report --fail-under="$COVERAGE_FLOOR"
+      - name: Badge endpoint (shields.io JSON, published with the docs)
+        # Colour follows the floor, so badge and gate can never disagree.
+        run: |
+          total=$(coverage report --format=total)
+          if python -c "import sys; sys.exit(0 if float('$total') >= float('$COVERAGE_FLOOR') else 1)"; then color=brightgreen; else color=red; fi
+          mkdir -p badges
+          printf '{"schemaVersion":1,"label":"coverage","message":"%s%%","color":"%s"}\n' "$total" "$color" > badges/coverage.json
+          cat badges/coverage.json
+      - uses: actions/upload-artifact@v6
+        with:
+          name: coverage-badge
+          path: badges/coverage.json
+          retention-days: 7
+      - name: Patch gate (PRs only)
+        if: github.event_name == 'pull_request'
+        run: |
+          git fetch origin "${{ github.base_ref }}"
+          diff-cover coverage.xml --compare-branch="origin/${{ github.base_ref }}" --fail-under="$PATCH_COVERAGE_FLOOR"
+
+  docs:
+    needs: [wheel]
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+      - uses: actions/download-artifact@v7
+        with:
+          name: wheel
+          path: dist
+      - run: python -m pip install "$(ls dist/*.whl)[docs]"
+      - run: |
+          chmod +x build_docs.sh
+          ./build_docs.sh
+      - name: Keep the site for the Pages deploy
+        if: github.event_name == 'push' && github.ref == 'refs/heads/develop'
+        uses: actions/upload-artifact@v6
+        with:
+          name: docs-html
+          path: docs/_build/html
+          retention-days: 1
+
+  ctest:
+    needs: [lint]
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    env:
+      CCACHE_DIR: ${{ github.workspace }}/.ccache
+      CCACHE_MAXSIZE: 1G
+    steps:
+      - uses: actions/checkout@v5
+      - run: sudo apt-get update && sudo apt-get install -y --no-install-recommends cmake ccache
+      - uses: actions/cache@v5
+        with:
+          path: .ccache
+          # Own key: tests ON shares few objects with the wheel build.
+          key: ccache-ctest-${{ runner.os }}-${{ github.sha }}
+          restore-keys: ccache-ctest-${{ runner.os }}-
+      - name: Build and run the C++ suite (shipping pure-Eigen CSD)
+        run: |
+          ccache --zero-stats
+          cmake -S cpp/libqarpx -B cpp/libqarpx/build-ci -DQARP_BUILD_TESTS=ON -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache
+          cmake --build cpp/libqarpx/build-ci --target qarpx_tests --parallel "$(nproc)"
+          ctest --test-dir cpp/libqarpx/build-ci --output-on-failure -j "$(nproc)"
+          ccache --show-stats
+
+  pages:
+    # Only develop publishes, and only once every test job is green.
+    if: github.event_name == 'push' && github.ref == 'refs/heads/develop'
+    needs: [docs, mypy, pytest, coverage, ctest]
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      pages: write
+      id-token: write
+    environment:
+      name: github-pages
+      url: ${{ steps.deployment.outputs.page_url }}
+    concurrency:
+      group: pages
+      cancel-in-progress: false
+    steps:
+      - uses: actions/download-artifact@v7
+        with:
+          name: docs-html
+          path: site
+      - uses: actions/download-artifact@v7
+        with:
+          name: coverage-badge
+          path: site/badges
+      - run: touch site/.nojekyll
+      # Inlined upload-pages-artifact: the composite still wraps a Node 20
+      # upload-artifact.  Same tar flags, same artifact name deploy-pages expects.
+      - name: Archive the site
+        run: tar --dereference --hard-dereference --directory site -cf "$RUNNER_TEMP/artifact.tar" --exclude=.git --exclude=.github .
+      - uses: actions/upload-artifact@v6
+        with:
+          name: github-pages
+          path: ${{ runner.temp }}/artifact.tar
+          retention-days: 1
+      - id: deployment
+        uses: actions/deploy-pages@v5

openqarp-0.1.0/.github/workflows/nightly.yml

@@ -0,0 +1,123 @@
+# The tiers the per-change gate deliberately excludes: slow-marked tests, the
+# randomised hypothesis profile, the executed notebooks, the benchmark smoke
+# run and the LAPACK build of the C++ suite.  Nightly matches what they catch
+# (upstream bumps, invariant drift), and workflow_dispatch is the on-demand
+# button for a branch that touches one of them.  Schedules run on the default
+# branch only.
+name: nightly
+
+on:
+  schedule:
+    - cron: "0 3 * * *"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+env:
+  PYTHON_VERSION: "3.12"
+
+jobs:
+  wheel:
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+      - uses: ./.github/actions/build-wheel
+
+  pytest-nightly:
+    needs: [wheel]
+    runs-on: ubuntu-latest
+    timeout-minutes: 120
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+      - uses: actions/download-artifact@v7
+        with:
+          name: wheel
+          path: dist
+      - run: python -m pip install "$(ls dist/*.whl)[full-dev]"
+      # CLI -m overrides pytest.ini's `-m "not slow and not bench"`.  Exit 5
+      # (nothing collected) is a real failure: the marker tier emptied out.
+      - run: pytest tests/ -m slow
+      - run: HYPOTHESIS_PROFILE=nightly pytest tests/ -m property
+
+  notebooks:
+    needs: [wheel]
+    runs-on: ubuntu-latest
+    timeout-minutes: 180
+    env:
+      TQDM_DISABLE: "1"          # a progress bar in a kernel is one output line per tick
+      OMP_NUM_THREADS: "2"       # pools sized from the host's core count oversubscribe a shared runner
+      OPENBLAS_NUM_THREADS: "2"
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+      - uses: actions/download-artifact@v7
+        with:
+          name: wheel
+          path: dist
+      # [full] not [full-dev]: [notebooks] stays out of the install the coverage
+      # floor is measured from; [integrations] carries the emit/absorb SDKs.
+      - run: python -m pip install "$(ls dist/*.whl)[full,notebooks,integrations]"
+      # The only job with the [integrations] SDKs installed, so the emit/absorb
+      # adapter suites run un-skipped here and nowhere else.
+      - run: pytest tests/test_emit tests/test_absorb -p no:cacheprovider
+      - run: pytest --nbmake --nbmake-timeout=1800 examples/ --junitxml=nbmake.xml
+      - uses: actions/upload-artifact@v6
+        if: always()
+        with:
+          name: nbmake-junit
+          path: nbmake.xml
+          retention-days: 7
+
+  bench-smoke:
+    needs: [wheel]
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    env:
+      OMP_NUM_THREADS: "2"
+      OPENBLAS_NUM_THREADS: "2"
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+      - uses: actions/download-artifact@v7
+        with:
+          name: wheel
+          path: dist
+      - run: python -m pip install "$(ls dist/*.whl)[full,integrations,bench]"
+      # Correctness only — timings on a shared runner are noise, never a gate.
+      - run: python -m benchmarks.smoke --require-all
+
+  ctest-lapack:
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    env:
+      CCACHE_DIR: ${{ github.workspace }}/.ccache
+      CCACHE_MAXSIZE: 1G
+    steps:
+      - uses: actions/checkout@v5
+      - run: sudo apt-get update && sudo apt-get install -y --no-install-recommends cmake ccache libopenblas-dev liblapack-dev
+      - uses: actions/cache@v5
+        with:
+          path: .ccache
+          # Own key: QARP_USE_LAPACK is a PUBLIC compile definition, no object is shared.
+          key: ccache-ctest-lapack-${{ runner.os }}-${{ github.sha }}
+          restore-keys: ccache-ctest-lapack-${{ runner.os }}-
+      - name: Build and run the C++ suite on the LAPACK zuncsd CSD
+        run: |
+          ccache --zero-stats
+          cmake -S cpp/libqarpx -B cpp/libqarpx/build-ci-lapack -DQARP_BUILD_TESTS=ON -DQARP_USE_LAPACK=ON -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache
+          cmake --build cpp/libqarpx/build-ci-lapack --target qarpx_tests --parallel "$(nproc)"
+          ctest --test-dir cpp/libqarpx/build-ci-lapack --output-on-failure -j "$(nproc)"
+          ccache --show-stats