[DrunkOnJava]

Summary

Unit tests are banned in dwg-rs, on the model rvt-rs already runs (its tools/ci/test-targets.txt and verify-real-files.sh). This PR classes every tests/*.rs file, switches CI's Test job to run every class but unit, and says in the docs that unit tests are banned. The inline #[cfg(test)] suites in src/ stop running with it and are deleted area by area in the PRs that follow. Git history is the archive.

• tools/ci/test-targets.txt classes the 17 files as real, contract, hardening or unit, one reason per line. No file is cli: nothing under tests/ drives a built binary. hardening is a fifth class that rvt-rs keeps for its fuzz-finding replays; here it holds integration_fuzz_corpus_regression alone.
• tools/ci/run_test_targets.sh runs the real, contract and hardening targets in one --no-fail-fast invocation, each target with a skip= list in its own (libtest's --skip applies to a whole invocation), then --doc on its own (cargo refuses --doc with --test). It fails on a tests/*.rs file the manifest does not class, on a stale or duplicate row and on an unknown class. It is bash with shell: bash in CI, so it runs on the Windows leg without a Python install.
• CI: the Test job keeps its name and its matrix, so the main-protection ruleset is unchanged. cargo test --release --all-features becomes tools/ci/run_test_targets.sh --release --all-features (toolchain through RUSTUP_TOOLCHAIN, so the 1.85 leg still runs 1.85), and a --check-only step before the build fails a bad manifest in seconds. cargo build --all-targets stays, which still compiles the unit files under -D warnings.
• Docs: CLAUDE.md, CONTRIBUTING.md (a new "Unit tests are banned" section), the PR template, AGENTS.md, ARCHITECTURE.md section 10, docs/EXTENDING_DECODERS.md and RELEASE.md.

Classing

Class	Targets
real	block_names, canonical_corpus, class_census, corpus_roundtrip, integration_dxf_roundtrip, integration_gltf, integration_legacy_container (without its 7 byte-built tests), r2013_entity_values, samples
contract	code_table_agreement
hardening	integration_fuzz_corpus_regression
unit	dispatch_roundtrip, integration_entity_regression, integration_svg_goldens, integration_write_roundtrip, mutation_failure_modes, proptest_roundtrip

The unit files stay as unrun source, as in rvt-rs. They still build under --all-targets. integration_svg_goldens is unit because two of its three tests pin text from synthetic documents and the third renders a hard-coded SVG tagged with a drawing's version; none reads a drawing's content.

Invariants carried over before anything is deleted

• entities::dispatch fixed_entity_label_covers_every_body_arm (the one CLAUDE.md names) becomes tools/ci/dispatch_arms.py --check, a step of CI's Spec coverage job. The inline test checked a hand-written list of codes against the label function, one direction only. The script reads fixed_entity_label and decode_fixed_entity_body from the source and compares them both ways, so a new decoder that adds its arm to one list and not the other fails. A body arm without a label (records come back Unhandled, counted as skipped) fails; so does a label without a body arm (records come back as errors). I mutation-tested it on copies of dispatch.rs (a dropped label, a dropped body arm, a narrowed | group) and ran it on the dispatch.rs of the open PRs feat(entities): decode SHAPE from §20.4.37 (#140) #310, feat(objects): decode VIEWPORT ENTITY HEADER from §20.4.70 (#146) #311, feat(corpus): pin the 122 US Forest Service AC1012 drawings by Wayback URL (#345) #412, feat(entities): route LIGHT by class name and read it with the layout its R14 record closes on (#152) #426 and feat(objects): decode RAPIDRTRENDERENVIRONMENT to its record boundary (#152) #427: all pass.
• A finding. OBJECT_TYPE_GEODATA has a label arm and no body arm. It cannot be reached: GEODATA is a non-entity object, so decode_in_page sends it to dispatch_object first. The script lists it as the one tolerated label-only code, and fails the day it stops being one. Removing the arm needs dispatch.rs, which ten open PRs touch, so it waits.
• Already external and kept: the object-type and DIMENSION table agreement against the specification's Table 4 is tests/code_table_agreement.rs (class contract), which no open PR touches. On real data, the Corpus job's tools/corpus/baseline.tsv turns a lost label into a changed skipped count.
• Considered and not ported, because they check a constant against itself or are covered on real files: limits.rs profile ordering and default == safe; known_section_names_are_unique; version_magic_matches_spec_table; the CRC, cipher, bit-cursor and LZ77 known answers (every R2004-family drawing in the corpus opens through them, and examples/spec_examples runs the specification's own worked examples in CI's Test job); the Reed-Solomon and R2007 LZ tests (probe_r2007_rs on every pinned R2007 page); the class-map and handle-map writer round trips (probe_class_roundtrip --strict on every pinned drawing, and the canonical fixtures regenerated and diffed in coverage-smoke). No pure-maths module is kept: a kept suite that no job runs checks nothing.

Not in this PR

• integration_dxf_roundtrip, integration_gltf and integration_legacy_container's four corpus tests read ../../samples but are not among the targets the Corpus job lists, so they skip on every job today. That is its own finding and its own PR (chore(ci): run every corpus-reading real target in the Corpus job #435).
• README.md and STATUS.md still quote "921 tests" from 2026-10-27. Open PRs touch both files, so the counts are replaced in a docs PR that lands after the deletions (docs(readme): stop quoting a unit-suite pass count #436).
• dispatch.rs's doc comment still names the deleted-to-be inline test; it is fixed with that file's deletion, since ten open PRs touch the file.

Spec reference

Not a decoder change. Issue: none.

Test plan

• tools/ci/run_test_targets.sh --check-only passes, and fails on an unlisted file, a stale row, a duplicate and an unknown class (checked on a scratch copy).
• python3 tools/ci/dispatch_arms.py --check passes, and fails on each of three mutations.
• python3 tools/ci/spec_coverage.py --check, python3 tools/maintainer/train/arch.py check and python3 tools/maintainer/readme_status.py --check pass.
• CI evidence, run 40674359229, all 24 checks green:
  • each of the six Test legs (ubuntu, macOS and Windows, stable and 1.85) logs == 10 targets: block_names canonical_corpus class_census corpus_roundtrip integration_dxf_roundtrip integration_gltf r2013_entity_values samples code_table_agreement integration_fuzz_corpus_regression, then == integration_legacy_container, without 7 synthetic-fixture tests (4 passed, 7 filtered out), then the documentation examples (10 passed). 93 tests per leg, against the 921 of a bare cargo test; the corpus-gated ones skip, as before. Windows ran it under shell: bash with no Python.
  • Spec coverage: 41 decoded codes all labelled; 1 labelled without a body (OBJECT_TYPE_GEODATA).
  • Corpus (public DWG): its five corpus targets pass with the corpus present and no skipping in the log.

Source-provenance declaration

By submitting this PR I certify that my contribution does not incorporate executable code from any source whose license is incompatible with Apache-2.0. Specifically:

• No Autodesk DWG SDK source (RealDWG, ObjectARX, ObjectDBX) or NDA-protected Autodesk documentation was consulted or imported.
• No Open Design Alliance Teigha / Drawings SDK source was consulted or imported. The ODA's public Open Design Specification for .dwg files PDF is explicitly allowed (see CLEANROOM.md).
• No GPL-licensed DWG implementation source (LibreDWG, etc.) was consulted or imported. This matters because dwg-rs is Apache-2.0 and must remain compatible.

My contribution is based on:

• The ODA's freely-redistributable Open Design Specification for .dwg files (v5.4.1);
• Public reverse-engineering notes, research papers, or blog posts cited inline;
• Reading algorithm-description comments (not executable code) in permissively-licensed projects (MIT / Apache / BSD) to resolve spec ambiguities: none were read for this change.

See CLEANROOM.md for the full allowed/forbidden source list and the honest scope of what "clean-room" means for this project.

Developer Certificate of Origin

By submitting this PR, I certify that the contribution complies with the Developer Certificate of Origin 1.1. Specifically, I affirm that the contribution is my own work (or rightfully licensed to me) and that it can be distributed under the project's Apache-2.0 license.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-28T11:37:49.023438Z	059d21c	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.