[MelbourneDeveloper]

TLDR

Code deduplication across DataProvider, LQL, Migration and Sync: measured duplication is now 14.76%, and the Deslop budget drops from 39.44% on main to 15%.

What Was Added?

• Shared production helpers that replace copy-pasted logic:
  • DataProvider codegen: CodeGenerationGuards, DataAccessGenerator partials (Bulk, BulkMethods, KeyWrites, SqlMethods).
  • CLI: CliOptions, CliFileSystem, CliErrorWriter, SqlServerCli, Other/Nimblesite.Cli.Shared/CommandDispatch.
  • LQL: StatementRendering, StatementBuilderOperations, SelectContextBase, CommonFunctionMappings.
  • Migration: SchemaInspectionSupport, ForeignKeyActionSql, SqlIdentifierRewriter, SqlSelectionExtractor, IndexExpressionSplitter, FunctionBodyValidation, SqliteColumnChecks, SqlServerExpressionNormalizer, PostgresCatalogQuery.
  • Sync: SyncClientStorage, SyncCommandExecution, SyncLogEntryFactory, TriggerGenerationWorkflow, SqliteCommandExecution, RelationalSyncCommands.
• Nimblesite.Sync.SqlServer: change applier, sync schema, store and trigger generator built on the shared relational sync commands.
• Tests.Shared/: shared test infrastructure (ResultAssertions, MigrationAssertions, MappingAssertions, DatabaseCommands, PostgresDatabaseTest, GeneratedCodeCompiler, CodeGenerationTestDoubles, ExtensionTestCases, PersonCommandParameters).
• Data-only files split out of algorithm files: *.Fixtures.cs / *.fixtures.rs scenario data, *.SqlSpecifications.cs / *.CatalogQueryText.cs SQL text, and *.EmissionText.cs code templates.
• Docs/reports/deslop-resources-2026-10-04*: duplication audit evidence.

What Was Changed or Deleted?

• Duplicated generators, inspectors, sync repositories, LQL renderers and test setups now call the shared helpers above.
• The DataProvider SQL Server generator (SqlServerCodeGenerator, SqlServerParser, SqlServerTableMetadata) is consolidated onto the shared codegen and uses Microsoft's ScriptDom T-SQL parser. PlaceholderMarker.cs is removed.
• Repeated test methods are collapsed into data-driven theories over shared scenario runners.
• Lql/Nimblesite.Lql.Core/DetailedLexerErrorListener.cs is merged into the generic DetailedParserErrorListener<TSymbol>, which now handles lexer and parser errors.
• .deslop.toml:
  • max_duplication_percent is lowered from 39.44 to 15.
  • Data-only files are excluded: scenario fixtures, SQL text specifications, emission templates, public DTO and model contracts, and the standalone example projects. Every exclusion has a stated reason, and every runner, algorithm and assertion is still analysed.
• .github/workflows/ci.yml: Deslop is bumped from 0.36.0 to 0.37.0.
• CustomCodeGenerationTests: fixed the test-local repository generator. It emitted "" around its SQL after the template moved out of a verbatim string. The expected fixture is unchanged.
• SqlModelCoverageTests.BuilderWhere.cs (new): runs SelectStatementBuilder.Where<T> predicates with AND, OR (wrapped in parentheses) and != against SQLite and asserts the exact rows returned. Moving codegen into new DataProvider.Core partials had pushed DataProvider.Example coverage to 39.44%, under its 40% gate; it is now 40.72%.
• RlsPredicateTranspiler: fixed a regression from moving bare-identifier quoting into the shared SqlIdentifierRewriter. The shouldQuote callback returned the keep-bare condition without negating it, so policies emitted tenant_id = "current_setting"(…)::"uuid". It now quotes columns and leaves functions, casts, keywords and qualified names bare, as on main. This change brought 127 failing Migration tests (RLS, trigger and policy round-trip) back to passing.
• SyncCoordinatorTests.Mapped.cs (new): the first tests for MappedSyncCoordinator. They run mapped pull and push between real SQLite replicas, with mapping state and record hashes stored by the SQLite MappingStateRepository. Covered: version tracking (applied once, then skipped), hash tracking (pushed once, then skipped while unchanged), echo prevention, strict unmapped tables, and apply, send and suppression failures. After the later Deslop commits Sync.Core coverage had fallen to 69.47%, under its 74% gate; it is now 88.79%.
• coverage-thresholds.json: the coverage gate ratcheted every threshold up to measured coverage: DataProvider.Core 92→93, Lql.Core 74→75, Migration.Core 88→89, Sync.Core 74→88, Sync.SQLite 43→46, Sync.Postgres 28→29, Sync.Integration 11→13, Sync.Http 43→44. No threshold was lowered.

How Do The Automated Tests Prove It Works?

make ci (lint, then every test project with its coverage gate, then build) exits 0 locally. make fmt CHECK=1 and deslop . also pass.

Project	Coverage	Gate
DataProvider/Nimblesite.DataProvider.Core	93.03%	93%
DataProvider/Nimblesite.DataProvider.Example	40.72%	40%
Lql/Nimblesite.Lql.Core	75.62%	75%
Lql/Nimblesite.Lql.TypeProvider.FSharp	49.16%	49%
Migration/Nimblesite.DataProvider.Migration.Core	89.27%	89%
Sync/Nimblesite.Sync.Core	88.79%	88%
Sync/Nimblesite.Sync.SQLite	46.62%	46%
Sync/Nimblesite.Sync.Postgres	29.39%	29%
Sync/Nimblesite.Sync.Integration	13.24%	13%
Sync/Nimblesite.Sync.Http	44.22%	44%
Reporting/Nimblesite.Reporting	78.82%	78%
Reporting/Nimblesite.Reporting.Integration	76.38%	76%
Lql/lql-lsp-rust	98.57%	98%
Lql/LqlExtension	54.48%	54%

• .NET: 2163 tests passed across 12 projects; 3 skipped (the Native AOT smoke tests, which need a published AOT binary).
• Rust (lql-lsp-rust): 442 tests passed.
• VS Code extension: 70 passing.
• Regression tests that caught the fixes: RlsCurrentSettingLqlTests, RlsPredicateTranspilerTests, RlsLqlExhaustiveTests and the Postgres/SQLite RLS E2E suites (127 failures before the quoting fix), and CustomCodeGenerationTests.CustomTableOperationGenerator_GeneratesRepositoryPattern.

Spec / Doc Changes

• docs/specs/*: adds IDs for each shared helper and the rules for keeping independent data, for example [CI-TEST-SHARED-SCENARIOS], [CI-TEST-SHARED-RESULT], [CI-DESLOP-INDEPENDENT-DATA], [CI-DESLOP-CONTRACT-DATA], [CI-INDEPENDENT-SQL-SPECIFICATIONS], [DP-CODEGEN-SHARED-BULK-METHODS], [DP-CODEGEN-SHARED-SQL-METHODS], [DP-CLI-SHARED-EMISSION-TEXT] and [LQL-VSCODE-SHARED-DIAGNOSTIC-SCENARIOS].
• Agent instructions move into AGENTS.md, and CLAUDE.md now points to it.

Breaking Changes

• None