frontend parity: Gremlin and SQL meet at the same Quack Query (test-only witness + parity matrix)

[AdaWorldAPI]

What this tests

Do independently designed query languages lower onto the same population algebra? This PR measures it for SQL ↔ Gremlin and reads the Cypher (#1306 plan) and SurrealQL (AST) sides against it.

No production code changes. No new operator in Quack, no new op in mask-risc, no upstream file touched.

Changes

• crates/lance-graph-quack/tests/gremlin_parity.rs (test only):
  • typed Gremlin steps (V, has, out, in, where, dedup, values, count, sum, groupCount, and the refused ones);
  • a declared schema over the DuckDB fixture: billedTo and partOf are fk relations, tradesWith is M:N with line as its edge table, ownedBy is a 3-row company table built only to write a two-deep chain;
  • an adapter that lowers onto lance_graph_quack::Query, with typed refusals;
  • an independent row-at-a-time Gremlin oracle with bulk;
  • 13 tests.
• .claude/plans/frontend-parity-witness-v1.md: inventory, parity matrix, Gremlin step→primitive table, SurrealQL AST harvest, gaps, falsifiers.
• Board: entry, STATUS_BOARD D-FPW rows, INTEGRATION_PLANS, AGENT_LOG.

Results

• Convergence below the frontend. The Gremlin traversal and its SQL equivalent produce == Query values, and execution matches DuckDB's committed answers: join_sum_country = 1237848 and join_group_count_country. The forward traversal (where(out(..))) and the reverse one (in(..)) land on the same anchor and the same atom multiset.
• Anchor rule. A functional hop keeps the anchor and reads the target through the fk (EqU32Via, GroupKey::Via). One fan-out re-anchors on the table whose rows are the paths, either the child table or the edge table. So one M:N hop is one Program, and its bag counts are exact.
• Input for plan: cypher-mask-lowering-v2 — mask engine beside upstream, one switch, refusals #1306. RF-BAG ("any non-DISTINCT count after a hop") is broader than the semantics require. After a functional hop, or one hop over an edge population, the bag count is exact.
• Five gaps pass the two-independent-witness rule. None is built.
  • composed functional read (= lance-graph-dir-sim: directory desired-state simulation over SoA + Quack (no writes) #1308 option (a));
  • barrier + workspace after a fan-out (= lance-graph-dir-sim: directory desired-state simulation over SoA + Quack (no writes) #1308 (b)/(d), board: chained-hop exploration map, round 2 (premise audit of (c), falsifier on (a)+(d)) #1310);
  • ordered compare through an fk;
  • sum of a foreign value;
  • ordered compare on u32.
• SurrealQL.
  • The shape fits: a hop is Part::Graph(Lookup), a mini-SELECT over a first-class edge table, and results are bag-flattened. That is the same geometry as the edge-population re-anchor.
  • It cannot be used as a dependency: sql::* is pub(crate), the public arena AST is unpublished, and the license is BSL 1.1. A fork-side lowering hook is D-FPW-3, and it is an operator decision.
• DuckGQL / DuckPG. No code or plan by those names exists in the tree.

Verification

• cargo test -p lance-graph-quack: all suites pass (36 + 39 + 13).
• clippy --tests -D warnings is clean; fmt is clean.
• Five disable runs each turned the suite red, then were restored:
  • dedup does not mark the set;
  • re-anchor drops carried predicates;
  • a bag emitted as a mask;
  • an edge hop drops the target-exists semijoin;
  • carry keeps a stale via atom.
• One limitation: the edge-hop disable is caught only by the structural Query check. The fixture has no dangling fk, so execution cannot see it.
• Adapter tax (debug build, printed, not asserted): adapter ≈ 1.9 µs, Quack lowering ≈ 1.0 µs, execution over 4,096 rows ≈ 167 µs.

Environment note: the workspace needs the ndarray sibling at ../ndarray. I cloned it read-only for the build; nothing in it changed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Eo5eHGADmeKNE3xYhgdK76

---

Generated by Claude Code

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Warning

Billing warning: we have not been able to collect payment for this subscription for more than 72 hours. Please update the payment method or pay any pending invoices in Billing to avoid service interruption.

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Autopilot could not be updated. Open Coding to check access and billing.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d4444f29b7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Isolate foreign-value summation with an i32 field

When this test is used to justify the G4 “sum of a foreign value” gap, Country conflates foreign addressing with an unsupported value type: terminal only permits Sum for Kind::I32, so even V(Partner).values(Country).sum() cannot lower. Consequently this case does not show that the foreign hop is the missing capability. Add an i32 field on the target table and demonstrate that its local sum succeeds while the through-FK form is refused.

AGENTS.md reference: AGENTS.md:L24-L25

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Reject unsupported steps after values()

When values() is followed by anything other than sum(), this assignment leaves the vertex cursor unchanged, so a traversal such as [V(Line), Values(Amount), Emit] is accepted and lowered to Agg::Rows, returning vertices rather than property values. The purportedly independent oracle repeats the same mistake by recording Values without changing trav and returning elements for Emit, so the two witnesses can agree on an incorrect Gremlin result; either model value traversers or reject every unsupported continuation.

Useful? React with 👍 / 👎.