# Tooling Alignment
This note records the issue #112 comparison between `causal-triangulations` and the sibling `markov-chain-monte-carlo` repository. Keep it current when changing
repository tooling so future updates can be deliberate rather than copied wholesale.
## Current Parity
Both repositories now have explicit local configuration for the core Rust and Python support-tooling loop:
- `rust-toolchain.toml` pins the Rust toolchain and development components.
- `rustfmt.toml` records stable Rust formatting settings used by `cargo fmt`.
- `.taplo.toml` keeps TOML formatting conservative and Cargo-like.
- `ty.toml` scopes Ty to `scripts/` and pins the supported Python version.
- `pyproject.toml` owns Ruff, Ty, pytest, Semgrep, actionlint, yamllint, shellcheck, shfmt, uv packaging, and uv-managed development dependencies.
- `rumdl.toml`, `dprint.json`, `.yamllint`, `typos.toml`, and `clippy.toml` define documentation and configuration checks.
- `justfile` is the single local entry point for formatting, linting, tests, coverage, Semgrep, changelog, and setup commands.
Non-Rust repository tooling is normalized to a 160-column width: Ruff for Python support scripts, `rumdl` plus the Markdown raw line-length guard, `yamllint`,
`dprint`/`pretty_yaml`, and the generated changelog postprocessor all use 160 columns. Rust remains on the narrower `rustfmt` `max_width = 100` setting because
wide Rust signatures, trait bounds, and method chains become harder to scan at 160 columns.
The Markdown raw line-length guard counts bytes with `LC_ALL=C` so macOS, Linux, and Windows agree on UTF-8 punctuation near the 160-column boundary.
## Python Tooling
CDT's Python tooling is broader than MCMC's. MCMC currently has changelog and tag helpers; CDT also has benchmark, hardware, coverage, archive, and
performance-analysis scripts. The shared script patterns have already been ported here:
- secure subprocess wrappers in `scripts/subprocess_utils.py`;
- typed `subprocess.CompletedProcess[str]` helpers in tests;
- Ruff, Ty, pytest, and uv-managed development dependencies in `pyproject.toml`;
- actionlint, yamllint, shellcheck, and shfmt are installed through the uv development environment so the same `just ci` path can run on Linux, macOS, and
Windows without Homebrew or runner-global tool assumptions;
- Python Semgrep rules and fixtures for broad exception catches across all `scripts/**/*.py`, raw `Exception` in tests, ad hoc subprocess mocks, and missing
return annotations.
The broad-exception Semgrep rule now covers the full Python support-script tree. `scripts/benchmark_utils.py`, `scripts/hardware_utils.py`, and
`scripts/performance_analysis.py` use typed recoverable exception boundaries, so new broad `except Exception` recovery paths are treated as tooling regressions
rather than accepted legacy cleanup.
## Intentional Differences
Some differences remain because CDT has different workflows and project invariants:
- CDT runs examples through `scripts/run_all_examples.sh`, which discovers current examples dynamically, builds release examples once, applies a timeout while
running the compiled binaries, and checks stable semantic output markers in `--validate` mode without requiring exact numeric output.
- CDT keeps `archive-changelog` so completed release series move under `docs/archive/changelog/`; MCMC does not yet archive old changelog sections.
- CDT keeps a dedicated `performance.yml` workflow and local `perf-*` recipes. MCMC does not have matching CDT benchmark-baseline tooling.
- CDT exposes feature-gated long-running Rust checks through the `slow-tests` Cargo feature and the `just test-slow` recipe, keeping normal CI fast while giving
stabilization work a named path for heavier integration coverage. Issue #140 release validation runs those slow tests through CDT's local `perf` profile,
matching the dedicated large-scale debug recipes instead of measuring default test-profile overhead.
- CDT has a repository rule SARIF workflow for the local Semgrep rules. A Codacy workflow was not ported because it depends on project-specific
`CODACY_PROJECT_TOKEN` setup and would duplicate the existing repository-rule SARIF signal until Codacy is configured for this repository.
- CDT Semgrep rules include geometry-backend isolation, foliation/topology validation, focused prelude imports, doctest assertion-idiom enforcement,
Python support-script discipline, and typed error policies. These are repository-specific and should not be weakened while porting generic rules. The
`prefer-assert-matches-in-doctests` rule keeps public `///` examples on `std::assert_matches` (Rust 1.96.0) so documentation teaches the diagnostic-friendly
idiom rather than `assert!(matches!(...))`.
- CDT and MCMC both require Python `>=3.12` for repository-managed support tooling.
## Ported Updates
The useful updates ported from MCMC are:
- explicit `rustfmt.toml` formatting configuration;
- explicit uv package mode and pytest 9 minimum in `pyproject.toml`;
- docs.rs package metadata with `all-features = true`, matching MCMC so rendered API documentation includes every public feature-gated surface;
- CodeQL analysis for GitHub Actions and Rust, using `build-mode: none` for Rust;
- the MCMC-style `cliff.toml` template and `just changelog-unreleased <version>` flow, adapted to keep CDT's changelog archive step and avoid temporary local
release tags;
- a Semgrep rule that rejects `NaN` and infinity defaults after failed floating-point conversions, with a regression fixture under `tests/semgrep/`.
- production-only Rust Semgrep rules that reject bare `unwrap()` and explicit `panic!` in non-test `src/` code while preserving idiomatic fail-fast usage in
unit and integration tests.
- public-surface Rust Semgrep rules that reject `unwrap()` and `expect()` in public doctests, Cargo examples, and benchmark harnesses, with benchmarks using
explicit fixture helpers so failed setup still reports the operation that failed.
- The `causal-triangulations.rust.no-unwrap-expect-in-doctests` rule now distinguishes code-like `.unwrap()` and `.expect(...)` calls from prose mentions, and
`causal-triangulations.rust.no-unwrap-expect-in-benches-examples` keeps Semgrep fixture cross-matches scoped to the
`tests/semgrep/src/project_rules/rust_style.rs` fixture shape. Before this refinement, prose such as "do not use `.unwrap()`" could be reported and Semgrep
test mode needed broader name-based exclusions; after it, public-surface checks still reject panic-style calls while fixture-only exceptions are explicitly
anchored.
- low-risk MCMC Semgrep rules for public `*_unchecked` Rust APIs, dynamic error erasure in examples/benchmarks, and dynamic error erasure in doctests. The CDT
versions keep repository-owned rule IDs, explicit fixture paths, and typed-error guidance so user-facing examples continue to model `CdtResult` or concrete
errors.
- an `examples-validate` recipe that runs Cargo examples and verifies stable output markers for the user-facing example contracts.
The useful Semgrep updates ported from the sibling `delaunay` repository are:
- a silent numeric-conversion fallback rule, adapted to CDT `src/` paths and cleaned up in observable/result aggregation code so conversion failures use
explicit branches instead of `unwrap_or` sentinels;
- preventive Rust rules for `partial_cmp(...).unwrap_or(...)`, function-local imports in production source, and `#[allow(clippy::...)]` suppressions, each kept
repository-owned with CDT rule IDs and fixtures.
The useful `justfile` updates ported from `delaunay` are:
- `ci-slow`, a named CI-plus-slow-tests workflow using CDT's existing `slow-tests` feature gate and local `perf` profile so large-scale toroidal probes validate
release-relevant behavior within a bounded runtime;
- `cargo nextest` for runnable Rust unit, integration, CLI, slow, example, and release test recipes, while keeping rustdoc doctests on `cargo test --doc`
because nextest does not execute doctests;
- single-pass release example builds in `scripts/run_all_examples.sh`, which preserve semantic output validation while avoiding repeated `cargo run --example`
invocations;
- `bench-smoke`, adapted to CDT's `cdt_benchmarks` harness and Criterion's minimal sample settings so benchmark harnesses can be smoke-tested without producing
baseline-quality numbers;
- `bench-test-compile`, which layers release-profile integration-test compilation on top of the existing warning-denying benchmark compile check;
- opt-in release hygiene recipes `unused-deps` and `publish-check`, kept outside the default `lint` and `ci` paths so they are available before releases without
making routine validation slower or more tool-dependent;
- `citation-check`, using `cffconvert` through `uvx`, and inclusion of `CITATION.cff` in YAML/CFF formatting and linting so research-software citation metadata
is schema-validated before release;
- `rumdl` Markdown checking, `dprint`/`pretty_yaml` YAML formatting, check/fix aliases, and a corrected rename/copy note in `spell-check`.
- repository-owned Semgrep rules that keep user-facing `just check` examples before `just fix` examples and enforce SHA-pinned, allowlisted GitHub Actions with
readable version comments. Dependabot remains the update path for pinned external actions, with review focused on preserving both the SHA and readable version
comment.
- Delaunay's non-Rust 160-column tooling policy, adapted for CDT while intentionally leaving Rust formatting at `rustfmt`'s 100-column width.
- MCMC-boundary Semgrep rules for issue #155, which reject production CDT-local Metropolis-Hastings `exp(log_alpha)` acceptance draws and manual
accepted/rejected sampler counter increments. These are CDT-specific because this crate still owns proposal planning, proposal-site telemetry, and result
translation, while `markov-chain-monte-carlo` owns the reusable sampler mechanics.
- The planned-step sampler-state sync rule now anchors `record_planned_step(...)` to call statements so it continues to catch missing
`sampler.replace_state(...)` after recorded planned steps without false-positive matches on the helper function declaration when that helper contains fallible
telemetry construction.
## Issue #162 CI And Security Alignment
Issue #162 refreshed the CI and security baseline against `markov-chain-monte-carlo` after acgetchell/markov-chain-monte-carlo#68 and #57:
- The main CI matrix now installs Python tooling with `uv sync --locked --group dev` and runs `just ci` on Ubuntu, macOS, and Windows.
- Rust CLI tools used in PR-running workflows are installed through `taiki-e/cache-cargo-install-action@417450f3c33ee20393705369577571770643d4c7`
(`v3.0.7`) instead of `taiki-e/install-action` or ad hoc `cargo install` cache scripts.
- The GitHub Actions Semgrep allowlist now permits `taiki-e/cache-cargo-install-action` and `zizmorcore/zizmor-action`, and no longer permits
`taiki-e/install-action`.
- Repository-owned Semgrep rules now also guard checkout credential persistence, `pull_request_target`, direct `github-script` expression interpolation,
unlocked workflow `uv sync`, direct Python `subprocess.run` bypasses, and direct MCMC imports outside the CDT Metropolis adapter boundary.
- Tool versions are pinned in workflow `env` blocks and mirrored in `justfile` constants for `cargo-nextest`, `dprint`, `rumdl`, `taplo`, `typos`, `zizmor`,
`cargo-llvm-cov`, and `git-cliff`.
- Local validation includes `just zizmor` through `lint-config`, while `.github/workflows/zizmor.yml` uploads the GitHub Actions security signal in CI.
- `SECURITY.md` documents private vulnerability reporting and the repository security check set.
Cold-cache PR runs after tool-version changes are expected to be slower while `taiki-e/cache-cargo-install-action` seeds tool caches. Warm-cache timing should
be measured after the first cache-warming run before making additional CI-shape changes.
## CI Shape Evaluation
Issue #131 evaluated larger CI-shape changes after the lower-risk speedups from
issue #130. Recent successful PR CI runs on May 22, 2026 showed the macOS and
Windows matrix legs often determining wall-clock duration:
- `perf/130-ci-speedups` (#137): Ubuntu completed in about 3m31s, macOS in
about 3m52s, and Windows in about 3m40s.
- `feat/checked-delaunay-api` (#128): Ubuntu completed in about 4m28s, macOS
in about 4m56s, and Windows in about 6m10s.
The repository intentionally keeps the existing `build (ubuntu-latest)`,
`build (macos-latest)`, and `build (windows-latest)` required check contexts,
and keeps Rust target coverage on all three platforms. A broader split that
would run the comprehensive repository/tooling path only on Linux was rejected
because it would weaken the current macOS build signal.
The adopted change is smaller: the Windows leg now uses
`cargo nextest run --workspace --all-targets --no-run` instead of an initial
`cargo build --all-targets`, then runs lib, binary, example, and integration
test targets through one nextest invocation. That keeps the full Rust target
compile surface on Windows, including benchmark compilation, while avoiding
accidental Criterion benchmark execution in the test job. Rustdoc doctests
remain on `cargo test --doc` because nextest does not execute doctests.
The May 2026 Delaunay tooling refresh was reviewed for portable pieces. CDT
adopted `cargo-llvm-cov` 0.8.7 in both local setup and Codecov, added `rumdl`
formatting to generated changelog recipes, and ported archive heading
normalization so historical changelog archives are cleaned even when a release
only rewrites the active minor series. Delaunay's profiling, Codacy SARIF
filtering, and benchmark-comparison machinery remained deferred from that pass
because CDT needed a different benchmark contract and has no configured Codacy
workflow.
The follow-up benchmark alignment introduced a CDT-specific
`ci_performance_suite` rather than copying Delaunay's dimensional construction
contract. CDT's suite covers generated open-boundary and toroidal
triangulations, validation, individual ergodic move attempts, ten-sweep
random-move workloads, and ten-sweep Metropolis runs. The suite uses a local
`perf` profile and feeds the existing performance-analysis scripts, while the
larger Delaunay profiling and same-machine comparison helpers remain deferred.
## Deferred Updates
These were evaluated but not ported in this pass:
- `codacy.yml`: defer until the repository has an intentional Codacy project token and a decision about whether Codacy should upload repository-owned
OpenGrep/Semgrep findings in addition to `.github/workflows/semgrep-sarif.yml`.
- Delaunay's hot-path `FastHashMap`/`FastHashSet` rule: defer because CDT does not currently define equivalent hash aliases or the same `src/core` hot-path
layout.
- Delaunay's `profiling_suite` and same-machine baseline recipes: defer because CDT's `ci_performance_suite` now provides the portable regression contract,
while deeper profiling still needs CDT-specific benchmark interpretation.