fix: honor $parameters in rel inline filters (#96), RETURN created rel vars (#95), caller-assigned edge ids for upsert_edge (#97)

[dylanbstorey]

Summary

• Inline relationship pattern property filter is ignored when the value is a parameter (WHERE clause / node patterns unaffected) #96 — MATCH ()-[r:T {k: $param}]->() silently skipped parameter-valued inline relationship property filters and matched every edge of the type (dangerous for scoped SET/DELETE). transform_match.c now emits the same OR-of-EXISTS parameter condition the node path already used.
• CREATE...RETURN r commits despite raising 'Unknown variable' when relationship connects pre-existing MATCH-bound nodes #95 — MATCH (x),(y) CREATE (x)-[r:T]->(y) RETURN r raised Unknown variable: r after the CREATE had committed (the handler re-ran MATCH+RETURN, which can't see CREATE-introduced variables). execute_multi_match_create_query can now hand back one variable map per processed MATCH row, and the MATCH+CREATE+RETURN handler projects CREATE-only variables from those maps (bare r, r.k, aggregates, SKIP/LIMIT; one result row per MATCH row). Queries returning only MATCH-bound variables keep the legacy path unchanged. The concurrency/WAL docs note requested in the same issue is included.
• upsert_edge has no equivalent to upsert_node's caller-assigned id, so it can't address parallel edges on the same (source, target, type) triple #97 — caller-assigned edge ids for upsert_edge. Prerequisite fix: the MERGE match phase ignored $param inline properties for both nodes and edges (matching any node of the label / any edge on the triple), and node creation dropped them — fixed in executor_merge.c (find_node_by_pattern, find_edge_by_pattern, node-create property phase). On top: Python upsert_edge(..., edge_id=None) and Rust upsert_edge_with_id(...) merge on an id relationship property, so parallel edges on the same (source, target, type) triple are individually addressable and upsertable in place. Default (no-id) behavior is unchanged; the existing Rust upsert_edge signature is untouched.

Test plan

• make test-unit — 947/947 pass (3 new CUnit regression tests: rel $param filter, MATCH+CREATE+RETURN created var, MERGE $param properties)
• make test-functional — clean, including new tests/functional/39_issue_regression_tests.sql (17 hard assertions; CHECK-constrained temp table aborts under sqlite3 -bail on any mismatch)
• python -m tests.tck --backend extension — 3778 pass; per-scenario diff against a pre-change run of the same tree: zero regressions
• make test-python — 357 pass (3 new edge_id tests)
• make test-rust — 244 integration tests pass (1 new upsert_edge_with_id test)
• Reproduced each issue's exact repro script before the fix and re-ran it after

Semantic coverage matrix

• Cells added/modified in docs/testing/semantic-coverage-matrix.md:
  • Section 1: MATCH (a) CREATE (a)-[:R]->(b) literal cell now covers RETURN r/r.k read-back on the created rel var (39:GH-95)
  • Section 3 (param vs literal symmetry): new rows for MATCH rel inline prop filter (read) (39:GH-96), MERGE node inline prop match phase (39:GH-97 3.1), MERGE rel inline prop match phase (39:GH-97 3.2)
  • Dated coverage-update section describing all three fixes
• Regression tests added/updated in tests/functional/:
  • tests/functional/39_issue_regression_tests.sql (new — the file the matrix process references)
• If neither: reviewer may apply skip-coverage-matrix label (pure refactors only).

Checklist

• angreal test functional green locally (run via make test-functional equivalent — sqlite3 -bail over tests/functional/*.sql)
• angreal test unit green locally (run via make test-unit)
• PR title follows conventional-commit style (fix:, feat:, chore: …)

Fixes #95, fixes #96, closes #97.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QuzbsuTddFm245sZ9egekS

---

Generated by Claude Code