Summary

• add a private, cursor-independent bounded CTE layout and visibility index for PostgreSQL, DuckDB, BigQuery, and Dremio grammar subsets
• preserve committed declarations separately from unfinished-body drafts, with correct nonrecursive order, nested shadowing, recursive withholding, duplicate blocking, and source-faithful insertion evidence
• use dialect-owned symmetric tri-state identifier comparison rather than generic folding or a second nullable comparison-key model
• preserve exact proven prefixes across lexer exhaustion and model cursor positions correctly at EOF, closing delimiters, barriers, and scope boundaries
• fail closed on hostile dialect callbacks/data, embedded regions, malformed syntax, unknown equivalence, and every checked resource boundary
• extend ADR 0005 and the provisional public relation-completion type contract for tri-state equality/prefix behavior and CTE-scope uncertainty

Bounds and performance

• active statement: 65,536 UTF-16 units
• shared lexical tokens: 16,384
• parenthesis/query depth: 128
• CTE frames: 256
• CTE declarations: 256
• identifier segment: 256 UTF-16 units

Representative local means:

• ordinary 10 KiB statement: ~0.37 ms
• 256 declarations with pairwise equivalence validation: ~3.0 ms
• 128 nested CTE frames: ~1.0 ms
• 256 sequential incomplete frames: ~0.15 ms
• cached 256-declaration projection: ~0.022 ms

Verification

• 1,587 tests passed, plus 1 governed expected failure
• changed runtime coverage: 96.96% statements, 95.88% branches, 100% functions, 96.94% lines
• all source, test, vNext exact/loose-optional, and demo typechecks pass
• oxlint and test-integrity gates pass
• package smoke, production build, demo build, and 7 Playwright browser tests pass
• three independent exact-head adversarial reviewers approved 8a1572369a171f49d3e23aeed4993aceb88a2cb2

Deferred follow-up

The bounded grammar currently authenticates SELECT query leaders. Dialect-owned support for PostgreSQL VALUES/data-modifying CTE bodies, DuckDB FROM-first queries, and additionally parenthesized BigQuery recursive terms remains an explicit follow-up before relation completion is feature-complete.

Part of #169.

Summary by cubic

Adds a bounded CTE layout and visibility index for vNext to improve relation completion across PostgreSQL, DuckDB, BigQuery, and Dremio, and now rejects invalid CTE cursor positions to avoid bogus suggestions. Introduces tri-state CTE identifier comparison and strict, fail-closed resource limits; part of #169.

• New Features

  • Private, cursor‑independent CTE index with correct order, shadowing, duplicate blocking, and recursive withholding.
  • Closed, dialect‑owned grammar per engine with a documented acceptance matrix; no cross‑dialect borrowing.
  • Tri‑state equality and prefix matching through the dialect runtime; preserves exact prefixes at EOF and scope boundaries.
  • Fast, bounded visibility via visibleSqlCtesAt; now rejects invalid cursor positions and the lexer reports limit hits with resourceAt, failing closed on malformed input.

• Migration

  • In SqlRelationCompletionDialectRuntime, replace cteIdentifiersEqual with compareCteIdentifiers returning "equal" | "distinct" | "unknown".
  • Implement cteIdentifierMatchesPrefix returning "match" | "no-match" | "unknown".

^{Written for commit 7302de4. Summary will update on new commits.}