fix(swrl): builtin atoms in inference/materialization — prefix lookup, datatype-property namespaces, WHERE-filter wiring, Postgres dialect

[rickynho-rccl]

Fixes #171.

Attribute-conditioned SWRL rules silently inferred zero triples. Four independent defects (detailed in the issue); this PR fixes all four with minimal diffs:

• SWRLBuiltinRegistry: _normalize() strips an optional namespace prefix so swrlb:greaterThanOrEqual, bare, and any casing all resolve. The editor emits the prefixed form.
• SWRLEngine._build_uri_map: class dataProperties are folded into the property normalisation so datatype predicates resolve to the data namespace exactly like object properties.
• SWRLSQLTranslator.build_inference_sql / build_materialization_sql: builtin atoms are partitioned out of class/property atoms and emitted as WHERE filters via the existing _build_builtin_filters (mirroring build_violation_sql).
• Dialect support: SWRLSQLTranslator(dialect=...); "postgres" maps TRY_CAST → CAST and DOUBLE → DOUBLE PRECISION. LakebaseBase.get_query_translator() now passes dialect="postgres".

Tests:

• tests/units/ontology/test_swrl_builtin_defects.py — one class per defect: prefix-tolerant registry lookup; datatype properties in the uri_map; builtins emitted as WHERE filters (no phantom join) in both the inference and materialization builders; Postgres dialect mapping plus the Lakebase backend handing out a postgres-dialect translator; and a guard that the Databricks dialect is unchanged.
• scripts/repro_swrl_builtin_defects.py — standalone red/green harness (validates the rule is safe SWRL, builds a 3-row fixture with a known answer, prints the generated SQL, scores inferred-vs-expected) so the defect and fix are independently checkable end to end.

Verified: fixture 0→2/2; a real 2,846-instance graph 0→263/263 against an independently computed expected count; the existing suite passes unchanged, including test_greater_than_uses_try_cast (the dialect mapping happens post-format, so the pinned Databricks templates are untouched).

Note on scope: TRY_CAST-style dialect handling could arguably live in the builtin descriptors rather than a post-hoc string mapping; we kept the change minimal and localized. Happy to rework to maintainer preference. The cypher_template field currently holds SQL-flavoured text rather than Cypher — left untouched, but flagging it.

[CLAassistant]

All committers have signed the CLA.

[benoitcayladbx]

Thanks for this — the diagnosis in #171 is correct, the four source hunks are the right shape, and the per-defect regression file is exactly the coverage the existing suite was missing. Independently confirmed on current main as well as the v0.8.0 path.

Happy to land this once the items below are addressed. None of them change the approach.

Blockers

1. CLA — the assistant bot is still pending. We cannot merge until that’s signed.
2. scripts/repro_swrl_builtin_defects.py — please drop it from the PR, or rewrite it to take connection info from env / CLI flags. As written it hardcodes host=localhost port=5433 … password=changeme, which we should not put in a public repo. The unit file (tests/units/ontology/test_swrl_builtin_defects.py) is the piece we want in-tree.
3. CI — no checks reported on this branch yet. Once CLA + the script cleanup are in, we need the usual unit suite green (pytest -q -m "not scenario").

Requested changes (keep the fix, tighten the edges)

4. Builtin name normalisation — rsplit(":", 1)[-1] is fine for the editor’s swrlb:greaterThanOrEqual, but it mangles a full IRI (http://www.w3.org/2003/11/swrlb#greaterThanOrEqual). Prefer stripping a known prefix set (swrlb:, swrl:) and/or taking the fragment after #, rather than splitting on the last colon.
5. Reuse SWRLParser.partition_rule_atoms in build_inference_sql / build_materialization_sql instead of hand-rolling the class/prop/builtin split. That duplication is how defect 3 existed while violation SQL was already correct.
6. getattr(self, "_dialect", "databricks") — with __init__ setting _dialect, just use self._dialect.
7. Dialect handling — the post-format TRY_CAST → CAST / AS DOUBLE → AS DOUBLE PRECISION mapping is the smallest hammer and is enough for the comparison builtins in the repro. Two notes to either fix or call out explicitly:
   • TRY_CAST vs CAST is a semantic change: Databricks yields NULL on junk literals, Postgres raises and can fail the whole inference run.
   • DATEDIFF, CURRENT_TIMESTAMP(), and some string templates are still Databricks-shaped; date/math builtins can still explode on Lakebase after this patch. Comparison (>=) is what #171 needs; a follow-up for the rest is fine if you document the gap. Longer-term, dialect-specific templates on the builtin descriptors (as you offered) is the cleaner home.

Nits / out of scope (no need to block)

• You’re right that cypher_template currently holds SQL-ish CAST(… AS DOUBLE), not Cypher — leave it, but thanks for flagging.
• build_materialization_sql still only joins direct properties of the primary var; inference already BFS-walks the chain. Unrelated to #171.

Verified locally that test_greater_than_uses_try_cast can stay green if the Databricks dialect is the default and mapping happens post-format — that part of the design is good.

[rickynho-rccl]

Thanks for the fast and precise review. Follow-up commit pushed; item by item:

Blockers

1. CLA: in progress on my side (employer sign-off needed for a work-derived contribution); separate from the code below.
2. Repro script: dropped from the PR. The unit file is the only test artifact now. The credential it carried was rotated the same day.
3. CI: no checks have ever run on this fork branch (first-time contributor), so it needs a maintainer to approve the workflow run. Locally, against this branch: pytest tests/ --ignore=tests/e2e -m "not e2e and not property and not eval and not external" → 5846 passed; flake8 --max-line-length=100 clean on the changed files (the one E501 it reports in SWRLBuiltinRegistry.py is a pre-existing template line, unchanged); ruff --select F821 clean; changed files are black-clean. A changelogs/v0.8.0/ entry is included for the changelog-presence gate.

Requested changes

4. Builtin name normalisation: now uri_local_name() (fragment after #, else last path segment) followed by stripping a known prefix set (swrlb:, swrl:), no more rsplit(":"). Full-IRI test added. One honest note: SWRL_ATOM_RE excludes :, so text rules already reach is_builtin with the bare name; the registry-level normalisation covers raw-token callers and IRIs.
5. partition_rule_atoms: both build_inference_sql and build_materialization_sql now use it (same five lines as build_violation_sql). One consequence worth stating: those two builders previously left negated atoms in class_atoms/prop_atoms (pristine too), i.e. not(p(?x,?y)) was joined as a positive pattern. With the shared partition they now go through _build_negated_atoms and emit NOT EXISTS, exactly like violation SQL; parametrized test added. build_antecedent_count_sql still hand-rolls the identical split - happy to fold it here or in a follow-up, your call.
6. self._dialect: done, all four sites.
7. Dialect gap: documented at the rewrite site and in the class docstring: TRY_CAST→CAST raises on non-numeric literals where Databricks yields NULL; matches (RLIKE), dateDiff (DATEDIFF) and now (CURRENT_TIMESTAMP()) remain Databricks-shaped and still error on Lakebase. Per-dialect templates on SWRLBuiltin as the follow-up, as you suggested.

Also in this commit: the _build_uri_map fix now writes datatype properties first and object properties last, so an object property wins a same-name clash - same precedence as AggregateRuleEngine._build_uri_map, test added. Previously my hunk had the opposite tie-break.

Nits: agreed on both; left cypher_template alone, and the materialization direct-only join is untouched.

Two things noticed while in there, out of scope, no action taken: (a) the '%%' in the startsWith/endsWith/contains templates survives str.format as a literal %% in the emitted SQL on both dialects; (b) SPARQLRuleEngine and DecisionTableEngine build their own uri_maps and still skip class dataProperties. Happy to open issues for either if useful.