Declarative trigger guards in schema YAML (SQLite + PostgreSQL)

[MelbourneDeveloper]

TLDR

Adds declarative trigger guards to schema YAML — per-table triggers: with LQL raiseWhen predicates that become native BEFORE/AFTER row triggers on SQLite and PostgreSQL, diffed by name exactly like RLS policies. Closes #82.

What Was Added?

• Model + YAML: TriggerDefinition (Name, Timing, Events, ForEachRow, raiseWhen LQL, ErrorMessage) on TableDefinition.Triggers, with TriggerTiming/TriggerEvent enums and round-trippable YAML (defaults timing: Before, forEachRow: true omitted on write).
• Diff: SchemaDiff.Triggers — triggers missing from the database produce CreateTriggerOperation; triggers removed from YAML produce DropTriggerOperation only under MigrationOptions.AllowDestructive (logged at Warning), mirroring RLS policy semantics.
• Guard LQL transpiler (RlsPredicateTranspiler.TriggerGuard.cs): supports old.<col>/new.<col> row references, and/or composition, and [not] exists(pipeline) subqueries. On Postgres, exists-pipeline SQL gets an identifier-quoting pass so mixed-case columns survive case folding. The internal __TRG_ sentinel token is reserved and rejected in user predicates.
• SQLite emission/inspection: one trigger per event named usr_{event}_{trigger}_{table} using SELECT RAISE(ABORT, '<message>') WHERE <predicate>; read back from sqlite_master and grouped by base name so re-diff is a no-op.
• Postgres emission/inspection: plpgsql guard function {table}_{trigger}_trgfn + trigger usr_{trigger}. RAISE EXCEPTION USING MESSAGE avoids % format interpretation; the dollar-quote tag is chosen to never occur in the body; identifiers over PostgreSQL's 63-byte limit fail loudly. The inspector reads only usr_-prefixed triggers, so unmanaged triggers (e.g. Sync change tracking) are never dropped by destructive runs, and *_trgfn guard functions are excluded from support-function read-back.
• Validation (all fail the migration loudly at DDL-gen time): forEachRow: false rejected (guards are row-level on every platform); old. references with an Insert event and new. references with a Delete event rejected.
• Shared test infrastructure: SqliteTestDb/PostgresTestDb helpers extracted from the RLS tests (moved, not copied).

What Was Changed or Deleted?

• SqliteRlsSchemaInspector refactored onto the new shared SqliteTriggerNames parsing (behaviour unchanged).
• SqliteRlsMigrationTests/PostgresRlsE2ETests now use the shared test-db helpers (pure moves; assertions unchanged).
• MigrationRunner.IsDestructive now includes DropTriggerOperation.
• PostgresSupportSchemaInspector excludes %_trgfn functions from function read-back (they are owned by the trigger lifecycle).
• .gitignore: added .deslop-cache/.
• Lql/LqlExtension/src/test/runTest.ts: the VS Code test host now uses a short --user-data-dir under the OS temp dir. The default dir lives under the repo, and on deep checkouts its IPC socket path exceeds the macOS 103-char AF_UNIX limit (listen EINVAL), making make ci fail on macOS. No test behaviour or assertions changed.

How Do The Automated Tests Prove It Works?

make ci (lint + test + build) is fully green locally, including all 533 Migration tests, Rust coverage 95.27% ≥ 95%, and the 70 VS Code extension tests. Highlights:

• SqliteDeclarativeTriggerTests (real temp-file SQLite): deleting/demoting the last owner of a tenant fails with the declared errorMessage and the row survives; succeeds when a second owner exists; guard is tenant-scoped; re-diff after apply yields zero operations; removing the trigger from YAML is a no-op without allowDestructive and drops it with; old.-on-Insert, forEachRow: false, and reserved __TRG_ predicates fail loudly.
• PostgresDeclarativeTriggerE2ETests (Testcontainers, real Postgres): same last-owner guard end-to-end; PgTriggerGuard_PascalCaseIdentifiers_EnforcedAtRuntime proves mixed-case columns work at trigger-fire time; PgTrigger_DestructiveRerun_PreservesUnmanagedTriggerAndGuardFunction proves unmanaged triggers survive destructive re-runs and the guard keeps enforcing.
• PostgresTriggerDdlTests: dollar-quote tag collision picks an alternate tag; trigger objects get the usr_ managed prefix; identifiers over 63 bytes throw with a clear message.
• TriggerYamlSerializerTests: YAML round-trip preserves the declaration and omits defaults.

Spec / Doc Changes

New docs/specs/declarative-triggers-spec.md with hierarchical spec IDs ([MIG-TRIGGER-YAML], [MIG-TRIGGER-MODEL], [MIG-TRIGGER-DIFF], [MIG-TRIGGER-GUARD-LQL], [MIG-TRIGGER-SQLITE], [MIG-TRIGGER-PG]), referenced from the implementing code and tests.

Follow-up filed as #84: the core LQL→PostgreSQL transpiler emits unquoted identifiers in pipeline SQL (latent in the RLS exists() path); this PR works around it for trigger guards via a quoting pass.

Breaking Changes

• None — triggers: is a new optional key; existing schemas are unaffected.

🤖 Generated with Claude Code