Plan, apply and roll back changes to a production ontology, with a blast radius report and a proof an auditor can re-check.
Rust ai-nativedescription-logicsformal-verificationinfrastructure-as-codeisabelle-languageknowledge-graphleanlean4mcpmcp-serverontologyowlowl2reasoningrustshaclsparqltheorem-proving
Merge branch 'fix/319-320-shacl-harnesses'
FABIOTESSa2d623efe62a
diff --cc CHANGELOG.mdindex cafac7fc,72d7c636..82d86e32--- a/CHANGELOG.md+++ b/CHANGELOG.md@@@ -6,70 -6,40 +6,104 @@@ All notable changes to Open Ontologies### Fixed+- **`diff` and `drift` no longer hang on a symmetric blank-node graph** (#293).+ Both canonicalised blank nodes through oxrdf's RDFC-1.0, which sets no limit+ on `hash_n_degree_quads` calls, so the cost was factorial: a 7-node+ blank-node clique took 17 to 23 seconds to diff against itself and an 8-node+ clique was still running when killed at 120 seconds. They now use the+ repository's own `canon` module, the same Recommendation gated on the W3C+ suite and bounded by its 4000-call budget; Pizza gets the same canonical+ labels as before. When either snapshot exhausts the budget, both are cut to+ the triples that mention no blank node and the report says so: `exact:+ false`, `approximate_reason`, and the omitted counts (`omitted_old`/+ `omitted_new` for diff, `omitted_v1`/`omitted_v2` for drift). `kgcl_json`+ carries the flag; the `kgcl` CNL format cannot, so it refuses an+ approximate drift and names the formats that can. Every report now carries+ `exact`, true on ordinary input. `canon` itself built all k! neighbour+ permutations before its first budget check, so two identical 11-leaf+ blank-node stars cost 10 s and 1.78 GB to refuse; permutations are now+ generated lazily in the same order, and the same input takes 0.04 s and+ 17 MB.+- **Lint and enforce feedback can no longer waive an error, and never drop an+ issue without a record** (#325, reported by @noahgift). After+ `suppress_threshold` dismissals (default 3) an issue was removed whatever its+ severity, leaving only `suppressed_count`, so the orchestrator that loops+ until enforce is clean could dismiss an error three times and read+ "0 violations, 1 suppressed" with no rule, entity or severity anywhere. An+ `error` is now never suppressed or downgraded (a dismissed one stays in+ `issues`/`violations` at `error` and carries `dismissals`); a warning or info+ that reaches the threshold moves to a new `suppressed[]` list with `rule`,+ `entity`, `original_severity`, `dismissals` and `message`. `issue_count`,+ `compliance` and `suppressed_count` are computed as before.+ `tests/feedback_suppression_test.rs` holds it for both tools.+- **A malformed SHACL count bound no longer turns a shape into a no-op** (#321,+ reported by @noahgift). `sh:minCount` fell back to 0 and `sh:maxCount` to+ `u64::MAX` when the bound did not parse, and a bad `sh:qualifiedMinCount` /+ `sh:qualifiedMaxCount` skipped the row in silence, so `sh:minCount "one"` or+ `sh:maxCount -1` over a node with no value returned `conforms: true`. All+ four now go to `skipped_constraints` with the reason, which makes the verdict+ `conforms: null`, as `sh:minLength` already did. The bound must be a+ non-negative `xsd:integer` literal: the datatype is checked because the+ store renders Turtle `1.0` as `"1"^^xsd:decimal`. The verified path+ (`verified: true`) refuses the same bounds before running `oo-shacl`, whose+ `readNat` declined "one" and -1 but evaluated `1.0` as 1. `onto_shacl_check`+ reports every ill-formed count or length literal as `ill_formed_literal`+ with its reason. `tests/shacl_count_bound_test.rs` holds it.+- **`onto_classify_el` is an EL classifier now, and says when its table is not+ complete** (reported by @noahgift in #324). It ran the OWL-RL-ext rules+ under a header saying they were "sufficient for the EL fragment" and returned+ "the full classification table". RL derives through a left-hand conjunction+ or existential only for an individual, so `A ⊑ B, A ⊑ C, B ⊓ C ⊑ D` came back+ without `A ⊑ D`, as did an existential reached through a conjunctive filler,+ a domain, a range or a transitive role (MEASURED: all five missing before,+ all five returned after). The issue's second example,+ `A ⊑ ∃r.B, B ⊑ C, ∃r.C ⊑ D`, was already returned, through `scm-svf1`. The+ read-back also stopped at 100,000 rows with no field to say so: a 460-class+ chain (105,570 pairs) came back as 100,000 after about 420 s; it now+ returns all 105,570 in 5.6 s (both debug builds). The tool now normalises the axioms the tableau's+ parser reads and saturates them with the EL completion rules (CR1-CR4,+ ⊥-propagation, role hierarchy, transitive roles, ranges compiled into+ fillers). The report gains `complete` (true only inside EL with nothing set+ aside), `incompleteness`, `consistent`, `truncated` (no cap remains, always+ false) and `unsatisfiable`; outside EL the RL rules are still unioned in.+ `tests/classify_el_differential_test.rs` checks it against the `owl-dl`+ tableau on 120 generated EL TBoxes (no disagreement; dropping the range+ rule makes it report 6) and against HermiT's+ Pizza classification (every reported pair entailed).+ - **The W3C SHACL conformance harness can say no, field by field, and+ ratchets a second, full-tuple column** (reported by @noahgift in #319). It+ compared `sh:conforms` and the set of (focus, source shape, component) and+ collapsed every blank-node source shape to `_:`, so two different property+ shapes compared equal, and nothing showed the comparison could reject+ anything. Blank property shapes are now keyed by parent node shape plus+ `sh:path` (the gated column stays at 61 of 120 PASS). A second column,+ ratcheted on its own in `tests/w3c_shacl_baseline.json`, compares the+ multiset of (focus, path, value, source shape, component, severity) with+ paths compared structurally: 12 of 120 PASS, 48 of the gap being `sh:value`,+ which the engine does not emit (#318). Severity went into the new column and+ not the gated one because it drops the gated column from 61 to 60: the+ engine keeps only the local name of a custom `sh:severity`+ (`core/misc/severity-002`). The header's reason for ignoring severity, that+ the engine did not read `sh:severity`, had been false since #206 and is gone.+ A planted control mutates one PASS case's expected report in each of the six+ fields and requires each copy to be rejected by name.+ - **`tools/shacl_differential.py` can no longer go green on nothing**+ (reported by @noahgift in #320). It exited 0 on everything but a FALSE_CLEAN,+ including no cases, no pyshacl and no comparisons, and the CI job's+ case-studies corpus yields 4 pairs that all come back UNDETERMINED. It now+ exits 2 (NOT MEASURED) for a missing or unpinned oracle, a missing engine,+ empty discovery, fewer pairs run than discovered, or no pair compared; plants+ a must-agree control and four mutated copies of pyshacl's report on every run+ and exits 1 if any is misclassified; keys results on (focus, path, severity,+ source shape, component) instead of (focus, path); records pyshacl, rdflib,+ the engine binary's sha256 and every input's sha256; and with `--w3c`+ arbitrates each disagreement against the W3C expected report. CI pins+ pyshacl 0.40.1 and rdflib 7.6.0 and runs the corpus and the W3C suite+ together: 117 pairs, 73 compared, 0 blocking, with 12 W3C-confirmed+ FALSE_CLEANs reported as KNOWN because the conformance ratchet already+ holds each as FAIL. `tools/fol_differential.py`+ (every goal undetermined) and `tools/horn_differential.py` (every case not+ run) had the same exit-0-on-nothing and now exit 2.## [2.1.0] - 2026-10-23