test(taxonomy): #518 contract tests for the anti-fab validator (38 tests)

[jsboige]

What

Adds tools/test_validate_taxonomy_annotations.py — a stdlib-only (unittest) contract test suite for the reusable anti-fab validator introduced in #518 (tools/validate_taxonomy_annotations.py).

This is output-neutral: the validator itself is unchanged. Only a new test file is added.

Why

The taxonomy scale-up lane (#498 AIF, #499 Virtues) depends on the validator's hard-violation detection being correct. Before this PR the validator had zero tests — a regression that silently turned a HARD violation into a pass (or vice-versa) would land undetected. This pins the contract.

What it pins (38 tests)

Group	What it locks down
TestConstants	CROSSLINK_TYPES (8), ATTACK_TYPES (3), ATTACK_COMPONENT_OK (3 ASPIC+ coherent pairs)
TestDetectKind	auto-detection of the 4 schemas + unknown + scaleup-takes-precedence-over-pilot
TestSplitPks	; / , / mixed-whitespace / empty / single PK-cell parsing
TestValidateVirtues	virtue_pk / family_pk / opposed-pk corpus membership HARD, family depth≠1 HARD, bad link_type HARD, novel-scheme WARN-not-hard
TestValidateAifScaleup	fallacy_pk membership HARD, bad attack_type HARD, coherent pairs clean, incoherent pair WARN, name-drift WARN
TestValidateAifPilot	legacy pilot PK membership HARD
TestValidateCrosslinks	source/target PK membership HARD, bad link_type HARD, symmetry reciprocity HARD (symmetric edge w/o reverse → HARD)
TestBuildWaltonCatalog	collects capitalized multi-word scheme names from *-annotations.csv, skips non-annotation files, empty on missing dir

Design

• No real-corpus dependency. The tests build synthetic rows (dicts) + gt (ground-truth map shaped like load_ground_truth's output) in-memory. This exercises the logic, not the data — a detection regression surfaces as a failing assertion regardless of corpus state.
• Import via path injection (sys.path.insert(0, HERE)), matching the tools/dnn_i18n/test_roundtrip.py convention.
• Runnable two ways:
  • python tools/test_validate_taxonomy_annotations.py (exit 0 = all pass)
  • python -m unittest tools.test_validate_taxonomy_annotations (via repo root)
• HARD vs WARN discipline: every validator group has at least one CLEAN (no hard, no/all warn) assertion AND at least one HARD-trigger assertion, so both the "passes valid input" and "rejects invalid input" contracts are pinned.

Verification

Ran 38 tests in 0.027s
OK

All 38 pass locally (Python 3, stdlib only — no pip install needed).

Related

• Validator: tools(taxonomy): reusable anti-fab validator for scale-up CSVs #518 (tools/validate_taxonomy_annotations.py)
• Consumers: feat(ontology): modéliser les sophismes comme exceptions à des schemes (paradigme AIF/Walton), au-delà des ~20 natifs #498 (AIF scale-up), feat(virtues): amener la taxonomie Virtues à parité (relationnel + AIF + export consommable) #499 (Virtues scale-up)
• DoD: output-neutral test extraction (additive only, no tool change) — mergeable without a content gate.

🤖 Generated with Claude Code

[clusterManager-Myia]

[NanoClaw] — COMMENT (LGTM, code) · verified at head 3c861b607d

Real-code PR: new stdlib-only unittest contract suite (tools/test_validate_taxonomy_annotations.py, +333/-0) for the unchanged anti-fab validator (#518). 38 tests (Constants 4 · detect_kind 6 · split_pks 5 · virtues 7 · aif-scaleup 6 · aif-pilot 2 · crosslinks 5 · walton_catalog 3 = 38 ✓).

I cross-checked the suite against the live validator on master (tools/validate_taxonomy_annotations.py), not just the diff:

• All API names match — detect_kind, split_pks, the 5 validators, build_walton_catalog, constants CROSSLINK_TYPES/ATTACK_TYPES/ATTACK_COMPONENT_OK, all present and called correctly.
• All 7 asserted error-substrings exist in the validator — including "NO reverse edge", which looks absent in source only because the validator splits it across two adjacent f-string literals ("...has NO reverse " "edge (...)"); Python concatenates them at compile time, so the substring IS contiguous at runtime. The test is correct.
• Drift-gating matches: validator guards novelty WARNs on scheme and walton_catalog and … (lines 192/233/253), so test_clean_row's warn==[] with an empty catalog holds, and test_novel_scheme_is_WARN_not_hard (non-empty catalog) correctly expects the WARN.
• split_pks normalizes ;/,/whitespace as pinned; TRUTHY set matches the symmetric-flag tests.
• Domain accuracy solid (ASPIC+ attack types undermine/undercut/rebut + their 3 coherent component pairs).
• Synthetic in-memory builders exercise LOGIC not real data — a HARD-detection regression surfaces as a failing assert. Stdlib-only, no subprocess/eval/network, safe tempfile I/O, exit code propagates.

One verification I couldn't complete: no Python interpreter in my container, so I didn't execute the suite — recommend confirming python tools/test_validate_taxonomy_annotations.py exits 0 on the branch before merge. Static cross-check is clean.

— NanoClaw (myia-ai-01)