[guijacquemet]

Summary

PR 1 of the Stage 1 design (section 5): the executable-conformance groundwork for the experiment-description reader. Tests and test data only. No change to the package, the CLI, the notebook, any output or any existing run. It waits for R7 / Gate G0 like the rest of Stage 1 and must not merge earlier.

The normative text and fixtures are in MorphoBricks-Spec at commit 8922a359ec261148a23a4b6e740a1bdb69e9e318 (docs/experiment-v1.md, docs/conformance-plan.md sections 1 and 2, fixtures/).

Stack: this is the base of a stack of three. PR 2 (reader phases 1-3) and PR 3 (phases 4-5 and warnings) are opened on top of it, against main; their diffs contain this one until it merges.

What changed

• tests/spec_fixtures/ is a partial copy of the Spec at the pinned commit, laid out as in the Spec: fixtures/ (179 files), docs/experiment-v1.md, tools/audit/refval.py (the reference validator, test oracle only, nothing in src/ imports it). PIN.toml holds the full commit SHA, git rev-parse <sha>:fixtures (6d114d80...) and the two blob IDs. .gitattributes sets -text so no checkout converts line endings.
• tests/test_spec_pin.py (pin integrity): the vendored fixtures/ hashes to the pin with git's tree algorithm (own implementation, checked against known git answers including the folder-sorts-as-name/ rule; git itself gives 6d114d80... for the committed folder). One changed, added, removed, renamed or CRLF-converted file fails, as do the two other files and a pin that is not a 40-digit SHA.
• tests/spec_gate.py (+ conftest.py): a skipped, skipif, importorskip or xfailed test in any tests/test_spec_*.py is reported as a failure. Tested in a child pytest.
• tests/test_spec_no_fallback.py ("no fallback rules"): AST scan for a second table of codes, a copy of the forbidden or reserved names (compared with the vendored text), code literals outside the one rules module, and lenient/legacy/fallback/default-version markers. The scan is tested against known-bad snippets. Until PR 2 adds description*.py the reader-module part has nothing to scan; PR 2 makes that non-vacuous.
• tests/test_spec_vendored_missing.py: with the vendored tree empty, absent or without fixtures/, the suite fails (collection error), never passes on defaults and never skips.
• tests/test_spec_catalogue.py: every fixture (discovered from the directory) has Fixture/Expected/Source headers matching its name and its catalogue row; the catalogue lists every file and nothing else; 36 valid, 130 invalid, 12 transitions.
• pyproject.toml / tools/check.py: the vendored copies are excluded from ruff and the code rules (byte-exact files, not our code). docs/development.md and CHANGELOG.md describe the pin and the gate.

Regression test: new tests, so there is no old-code run. Negative check done by hand: changing one byte of v01_single_control.toml makes tests/test_spec_pin.py fail (11 failed, 5 passed); restored, 16 passed.

Notes for review

• Pin updates are a reviewed change (tests/spec_fixtures/README.md lists the steps).
• Not verified: the Windows and macOS CI legs (the -text attribute is what protects the hashes there).
• No existing issue covers Stage 1, so this does not close one.

Tests

python tools/check.py: ruff check clean, ruff format clean, code rules clean, 2713 passed, 4 skipped (2316 passed, 4 skipped on main 3231966; the 397 new tests pass; the 4 skips are the same pre-existing ones in test_cli_log, test_discovery, test_folder_rules, none in a Spec test).

---

Generated by Claude Code