fix(shaclgen): emit sh:pattern for pattern constraints inside any_of

[jdsika]

Dependencies: None (standalone PR)

Summary

The SHACL generator (shaclgen.py) currently ignores pattern constraints specified inside any_of branches. When a slot uses any_of with a branch that has pattern: (with or without range:), the generated SHACL produces an empty blank node [ ] instead of [ sh:pattern "..." ].

This means any value trivially satisfies the branch (an empty node has no constraints), making the sh:or validation too permissive.

Root Cause

In shaclgen.py (the for any in s.any_of: loop, ~L325–373), branch translation dispatches solely on any.range:

• If range maps to a class → sh:class
• If range maps to a type → sh:datatype
• If range maps to an enum → sh:in
• Else → empty BNode (via add_simple_data_type(func, None))

The loop never reads any.pattern from the AnonymousSlotExpression, so pattern-only branches (no range) produce empty constraint nodes, and range+pattern branches lose the pattern constraint.

Note: the top-level s.pattern check (~L394) sits in the else: arm of if s.any_of:, so it only runs when any_of is absent and cannot cover patterns inside branches.

Fix

After the if/elif/else dispatch that creates each branch BNode, add:

if any.pattern:
    g.add((range_list[-1], SH.pattern, Literal(any.pattern)))

This correctly handles all three cases:

1. Pattern-only branch (no range): empty node gets sh:pattern → [ sh:pattern "^..." ]
2. Range + pattern: node gets both sh:datatype and sh:pattern → [ sh:datatype xsd:string ; sh:pattern "^..." ]
3. Range-only (no pattern): no-op — existing behavior preserved

This mirrors the existing top-level handling (if s.pattern: prop_pv(SH.pattern, Literal(s.pattern))), so both paths emit sh:pattern consistently.

Test Coverage

A 3-class test schema (tests/linkml/test_generators/input/shaclgen/any_of_pattern.yaml) exercises:

• PatternOnlyBranch: enum + URI + pattern-only (no range) in any_of
• RangeWithPattern: range: string + pattern on same branch
• MixedBranches: combination of range-only, pattern-only, and range+pattern branches

Two tests consume it:

• test_any_of_with_pattern() — asserts on the generated RDF triples (which branch node carries sh:pattern, and that range-only branches do not).
• test_any_of_with_pattern_pyshacl_end_to_end() — behavioural guard: runs pyshacl over conforming and non-conforming instances for all three classes. Without the fix a pattern-only branch serialises as an empty shape [ ], which every value node trivially satisfies, so sh:or accepts anything and the non-conforming assertions fail.

Both tests were verified to fail on main without the one-line change and pass with it. All existing tests continue to pass.

Real-World Use Case

This fix is needed for SPDX license validation in the Gaia-X / ENVITED-X ecosystem. The gx:license slot uses any_of to accept:

1. Known SPDX identifiers (enum)
2. Arbitrary URIs (range: uri)
3. Custom license refs matching ^LicenseRef-[a-zA-Z0-9\-\.]+$ (pattern-only branch)

Without this fix, branch (3) generates an empty SHACL node that accepts anything, making validation meaningless. The corresponding upstream model change is tracked in service-characteristics fix/spdx-license-ref-pattern.

Specification References

• SHACL W3C Rec §4.6.2 (sh:pattern): defines pattern as a string constraint on value nodes
• LinkML Spec §5.6.4 (pattern): The ordinal string value of the slot must match the given pattern
• LinkML Spec §5.6.14 (any_of): holds if at least one of the expressions hold

sh:pattern is a value-node string constraint, so it is legal (and idiomatic) on the branch shapes inside sh:or; for literals it matches the lexical form, for IRIs the IRI string.

---

This fix was developed and validated using the Gaia-X credential model as a real-world test case. AI-assisted development (GitHub Copilot).