[stugorf]

Problem and change

A complete valid query ending in ; is rejected as unexpected trailing input.

Consume one optional terminator at the shared parser entry point. String literals still parse normally; repeated terminators, empty statements and appended statements remain errors.

Refs lance-format/lance-graph#101. This is a lance-csr fix; it does not close the upstream issue.

Validation

• Failing before on the verified 0.3.1 debug baseline: 32 failed / 4 passed. Passing after: 36 focused cases, 1,299 complete Python tests in both debug and isolated production-wheel verification.
• Passed: cargo fmt --all --check, cargo test --workspace --locked, workspace Clippy with warnings denied, make install, make test, make lint, executable documentation snippets, strict Zensical build and typos@1.26.0.
• All three candidate debug gates passed before any final production build. Fresh locked consumers outside the checkouts verified wheel/package payloads and native hashes, then ran each complete suite with Rust absent from PATH.

Standalone scalar/null RETURN, UNION ALL, MATCH/WITH, aggregation, pagination, surrounding whitespace, semicolons in strings, collected/streaming execution, equivalent ASTs and multi-statement rejection.

Workload	Mode	Correct queries	First-pass geometric speedup
30-query SF1	memory	30	1.000x
30-query SF1	lazy	30	1.020x
IC SF1	memory	14	1.004x
IC SF1	lazy	14	1.021x

Above 1 is faster. These original suite means are retained unchanged; initial >10% per-query flags and their paired investigations are recorded in the final gate record. Both complete workloads preserve the original golden answers. These are local engineering measurements, not an audited SNB result or performance guarantee.

Measured serially after builds stopped, on Apple M4 Max (16 cores, 128 GiB), macOS arm64, Rust 1.96.0 and Python 3.13.17. The suites are pinned at 6260689 (30 queries) and c185123 (IC); settings, dependencies, SF1 manifests, timings and peak process RSS are recorded. Per-query before/after tables, reproduction instructions, and hashed raw evidence provide the commands and receipts.

User impact

Calls accept one optional trailing semicolon. The Cypher guide documents the one-query-per-call rule. No Python signatures, stubs or dependencies change.

Independently based on f5a986e822032a3aa8c52af55ad08e92504ee73c, head 28096751241a8707042091e5012f99ce1f32f594. Latest main 6effd0db0c002912cf969fd7447ea7a3ce2631df adds documentation links/package URLs only; runtime sources and locks are unchanged, and this PR applies cleanly. Other platforms and hosted CI are not included in the local verification; check the PR checks for their current status.

Submitted by David Hughes (@stugorf), with AI Collaboration.