(SPARQL) Full W3C Suite in CI w/ Registers for Skip/Ignore

[aaj3f]

Summary

This PR makes the entire W3C SPARQL test suite (the rdf-tests submodule) an enforced, durable CI gate. Before this change, the testsuite-sparql CI job ran cargo test --no-run — it compiled the harness but never executed a single test — and every evaluation category was #[ignore]d. The harness had also fallen behind the engine: SPARQL UPDATE execution (#509, #1288) and TriG/named-graph support (#1278, #1279) landed in the API, but the harness still had "not implemented" stubs and loaded named-graph data into the default graph.

Now cargo test in testsuite-sparql/ runs all 36 suites — ~1,420 tests — green in ~15 s, and CI runs exactly that. Every test either passes or appears in an explicit, root-cause-grouped skip register that is policed in both directions.

Standing after this PR: 815 passing / 538 registered engine or feature gaps / 67 registered not-applicable (protocol, graph-store protocol, service-description, SERVICE-needs-endpoints, entailment regimes).

The enforcement model

• Every manifest in the submodule is registered as a test fn in tests/w3c_sparql.rs — including the previously unregistered SPARQL 1.2 tree (254 tests). Nothing is #[ignore]d; nothing is silently un-run.
• Per-suite skip registers live in tests/registers/mod.rs, grouped by root cause with rationale comments and pointers into the audit doc.
• check_testsuite enforces the register both ways: an unregistered failure fails the suite (regression), and a registered test that now passes also fails the suite (stale entry — it must be removed in the same change that fixes the feature). This mechanism immediately proved itself: the old property-path skip list had rotted to 14 stale entries out of 18 (13 features had since landed, plus pp06 freed by a harness fix in this PR).
• Registers can therefore only shrink. The burn-down plan (Phases B–D, with per-cluster perf-safety prescriptions) is documented in docs/audit/2026-07-sparql-testsuite-audit.md.

Harness changes

• mf:UpdateEvaluationTest implemented end-to-end through the public sparql_update transact surface: load initial graph-store state (ut:data/ut:graphData), apply the update, compare resulting default-graph and named-graph states isomorphically, and check for unexpected non-empty named graphs. Update-eval went from 24 → 67 passing.
• Named graphs load as real named graphs: qt:graphData/ut:graphData files are wrapped as TriG GRAPH blocks (directives hoisted) and loaded through the transact builder. Also fixes a latent bug where labeled graphData blank nodes were silently dropped (get_graph_data never read ut:graph).
• CSV/TSV expected-result support: RFC 4180 CSV and SPARQL-TSV parsers plus lossy CSV-space projection for comparison. csv-tsv suite went 0 → 5 of 6.
• SPARQL 1.2 suites registered per category, and the un-versioned mf:PositiveUpdateSyntaxTest/mf:NegativeUpdateSyntaxTest types they use are now handled.
• Duplicate combined-manifest test fns removed; Makefile and docs/contributing/sparql-compliance.md rewritten to match the registers model.

New engine findings (registered, not fixed here)

The suite immediately surfaced precise engine gaps, now sitting in the registers with grouped rationale:

• GRAPH ?g exposes the default graph as a graph named by the ledger alias, and binds ?g as a plain literal rather than an IRI (breaks most sparql10/graph tests even though the data now loads correctly).
• GRAPH blocks inside DELETE WHERE are rejected at lowering; INSERT into a not-yet-existing named graph silently loses triples; USING semantics are incomplete; combined DELETE/INSERT WHERE applies inserts without the deletes.
• The UPDATE grammar lacks all graph-management verbs (LOAD/CLEAR/CREATE/DROP/COPY/MOVE/ADD + SILENT, ~70 tests). Modify does parse WITH/USING.
• Turtle lexer rejects a blank-node label immediately followed by . (_:o6.) — blocks all 4 json-res tests and affects real user data of that shape.
• The ~76-test SPARQL 1.0 expression cluster: XSD type promotion in comparisons, open-world equality, builtin edge semantics.
• RDF-star: a code-level sub-audit (audit doc §4.3) confirmed Fluree's edge-annotation/reified-edge model covers parse → IR → match → storage; the actual 1.2 gaps are triple-term functions, bare triple-terms-as-values (pending design decision), Turtle-star ingest, and CONSTRUCT annotation projection.

Perf safety

This PR contains zero engine changes — it touches only the out-of-workspace testsuite-sparql crate, CI config, and docs. Hot-path risk is nil by construction. The audit doc (§6) prescribes the perf-safety discipline for the follow-up engine fixes: parse/prepare-time fixes preferred, common-type fast paths preserved byte-identical, query_hot_bsbm*/insert_formats bench guardrails within regression-budget.json budgets per PR.

Test plan

• cd testsuite-sparql && cargo test — 36 suites, 0 failed, 0 ignored (~15 s).
• cargo fmt --all -- --check and cargo clippy --all-targets -- -D warnings clean.
• CI job runs the full suite with a 45-minute timeout backstop (actual runtime is minutes, dominated by build).
• No workspace crates touched; workspace CI unaffected.