@memberjunction/codegen-lib 6.1.0-edge.2 → 6.1.0-edge.3

package/dist/Database/manage-metadata.js

@@ -1,11 +1,17 @@
+import { SQLServerDialect, PostgreSQLDialect } from '@memberjunction/sql-dialect';
 import { CodeGenDatabaseProvider } from './codeGenDatabaseProvider.js';
+import { analyzeQueryForMaterialization, detectAggregationKeyColumns, detectAdditiveMeasures, MATERIALIZATION_SURROGATE_COLUMN } from './materializationAnalysis.js';
+import { evaluateMaterializationDrift } from './materializationDrift.js';
+import { classifyQueryParameters, buildHeldValues } from './materializationParamClassifier.js';
+import { buildBroadRowFilterSQL } from './materializationBroadRender.js';
+import { QueryParameterProcessor } from '@memberjunction/query-processor';
 // Side-effect import — registers `SQLServerCodeGenProvider` with `MJGlobal.ClassFactory`
 // under the `'sqlserver'` key via its `@RegisterClass` decorator. Without this import,
 // `ClassFactory.CreateInstance(CodeGenDatabaseProvider, 'sqlserver')` returns nothing
 // and the SS code path silently fails.
 import './providers/sqlserver/SQLServerCodeGenProvider.js';
 import { configInfo, currentWorkingDirectory, dbPlatform, getSettingValue, mj_core_schema, outputDir } from '../Config/config.js';
-import { EntityInfo, ExternalDataSourceReadRouter, ExtractActualDefaultValue, LogError, LogStatus, Metadata, SeverityType } from "@memberjunction/core";
+import { CodeNameFromString, EntityInfo, ExternalDataSourceReadRouter, ExtractActualDefaultValue, LogError, LogStatus, Metadata, RunQuerySQLFilterManager, SeverityType } from "@memberjunction/core";
 import { MJEntityFieldSchema } from "@memberjunction/core-entities";
 import { logError, logMessage, logStatus, startSpinner, updateSpinner, succeedSpinner } from "../Misc/status_logging.js";
 import { SQLUtilityBase } from "./sql.js";
@@ -15,7 +21,7 @@ import { CodeGenReporter } from "../Misc/codegen-reporter.js";
 import { applySearchableFieldsCap, entityLevelEnableBlockedReason, isNarrativeFieldName, normalizePredicate, normalizeSmartFieldResultShape, } from "./search-guardrails.js";
 import { mapExternalNativeTypeToMJ } from "../Misc/externalTypeMapping.js";
 import { SQLParser } from "@memberjunction/sql-parser";
-import { createDisplayName, generatePluralName, MJGlobal, stripTrailingChars, UUIDsEqual } from "@memberjunction/global";
+import { createDisplayName, generatePluralName, MJGlobal, ResolveSingleEntityResourceTarget, stripTrailingChars, UUIDsEqual } from "@memberjunction/global";
 import { v4 as uuidv4 } from 'uuid';
 import * as fs from 'fs';
 import path from 'path';
@@ -48,22 +54,30 @@ export class ManageMetadataBase {
         // ─── End Dialect Infrastructure ───────────────────────────────────
         this._sqlUtilityObject = MJGlobal.Instance.ClassFactory.CreateInstance(SQLUtilityBase);
         /**
-         * External-data-source analogue of {@link manageVirtualEntities}. For each entity backed by an
-         * external data source, introspect the REMOTE schema (via the EDS router resolved through the
-         * ClassFactory) and sync its `EntityField` rows to match — the remote equivalent of reading the
-         * local INFORMATION_SCHEMA for a virtual/view entity. No-op when there are no external entities.
+         * True if the MaterializedResult table exists. createNewEntities and detectMaterializationDrift gate their
+         * references to it on this so codegen doesn't throw on a DB lacking the materialization schema. Cached per run.
          */
+        this._materializedResultTableExists = null;
         /**
-         * Whether the current database's `Entity` table has the `ExternalDataSourceID` column. External
-         * Data Sources ship as a SQL Server migration only, so on any database/schema that predates it
-         * (PostgreSQL today) the column is absent — and ANY raw SQL referencing it throws
-         * "column does not exist" and aborts the entire CodeGen run, even for a pure MJ-DB schema with no
-         * external entities. Callers gate EDS-specific queries on this so CodeGen stays green everywhere
-         * and auto-activates once the column lands on other platforms (mirrors the PG stored-proc guard in
-         * metadataSupportObjects.ts). Cross-platform via INFORMATION_SCHEMA; cached for the run (the schema
-         * cannot change mid-run).
+         * Per-run cache of the entity NAMES (trimmed + lowercased) that an API-KEY row filter
+         * (`APIKeyScope.RowFilterID`) or an API-APPLICATION row ceiling (`APIApplicationScope.RowFilterID`) binds to.
+         *
+         * `'unknown'` — the INITIAL value — means "not loaded, or could not be enumerated", and makes
+         * {@link entityHasRowLevelRestriction} treat EVERY entity as row-restricted. That is the fail-CLOSED
+         * direction on purpose: refusing to materialize costs nothing (the entity/query stays live-only), while
+         * materializing an entity whose API-key row filter we failed to see is the precise leak these gates exist to
+         * prevent. Loaded by {@link loadAPIKeyRowFilterTargets} at the top of every materialization entry point.
+         * `protected` (not `private`) so a unit-test seam can declare "this fixture configures no API-key row filters".
          */
-        this._entityHasExternalDataSourceColumn = null;
+        this.apiKeyRowFilterTargets = 'unknown';
+        /**
+         * THE cached INFORMATION_SCHEMA column-existence probe against the MJ core schema — cross-platform, and
+         * cached for the run (the schema cannot change mid-run). Every column gate in this class routes through
+         * it: {@link entityHasExternalDataSourceColumn}, {@link queryHasExternalDataSourceColumn},
+         * {@link queryHasIsMaterializedColumn} and {@link loadAPIKeyRowFilterTargets} are each a one-line call,
+         * so a new gate has no reason to hand-roll a fourth copy of the same query and its own cache field.
+         */
+        this._coreSchemaColumnExists = new Map();
     }
     /**
      * Returns the CodeGenDatabaseProvider for the current database platform.
@@ -398,6 +412,22 @@ export class ManageMetadataBase {
             ForeignKeys: Array.isArray(ve.ForeignKeys) ? ve.ForeignKeys : undefined,
         }));
     }
+    /**
+     * Extracts the MaterializedBaseViews array from the additionalSchemaInfo config file.
+     * Each entry declares a 1:1 base-view materialization of an existing entity (plan §4.1).
+     */
+    extractMaterializedBaseViewsFromConfig(config) {
+        const decls = config.MaterializedBaseViews;
+        if (!Array.isArray(decls))
+            return [];
+        return decls.map((d) => ({
+            EntityName: d.EntityName || undefined,
+            BaseTable: d.BaseTable || undefined,
+            SchemaName: d.SchemaName || undefined,
+            RefreshSchedule: d.RefreshSchedule || undefined,
+            IntendedWorkload: d.IntendedWorkload || undefined,
+        }));
+    }
     /**
      * Extracts ISARelationships array from the additionalSchemaInfo config file.
      * The config may contain a top-level "ISARelationships" key with an array of
@@ -657,6 +687,34 @@ export class ManageMetadataBase {
         }
         return { success: true, createdCount, updatedCount };
     }
+    /**
+     * Builds the entity lookup behind {@link processISARelationshipConfig}: match on Name first, else
+     * on BaseTable, optionally constrained to a schema.
+     *
+     * The schema predicate is composed CONDITIONALLY rather than written as
+     * `(@SchemaName IS NULL OR SchemaName = @SchemaName)`. That form is fatal on PostgreSQL: the
+     * parameter's only unambiguous use is `$n IS NULL`, which gives the planner no type to infer, so
+     * the whole statement fails to prepare with "could not determine data type of parameter $n".
+     * SQL Server infers the type from the other side of the OR and never saw the problem.
+     *
+     * The failure was silent in the worst way — processISARelationshipConfig catches per-relationship
+     * errors and logs them, so CodeGen ran to completion with a zero exit code while every declared
+     * IS-A relationship on PostgreSQL was quietly skipped, leaving Entity.ParentID NULL. Downstream
+     * that means no mirrored parent fields, no parent JOIN in the child base view, and a child whose
+     * Save() never writes the parent row.
+     *
+     * Emitting the predicate only when a schema was supplied keeps the parameter list free of
+     * type-ambiguous entries and is portable to both dialects without a cast.
+     */
+    buildISAEntityLookupSQL(schema, selectBody, nameParam, schemaName) {
+        const schemaPredicate = schemaName ? ` AND SchemaName = @ISASchemaName` : '';
+        const sql = `
+                  ${this.selectTop(1, selectBody, `FROM ${this.qs(schema, 'vwEntities')}
+                  WHERE Name = @${nameParam}
+                     OR (BaseTable = @${nameParam}${schemaPredicate})`, `CASE WHEN Name = @${nameParam} THEN 0 ELSE 1 END`)}
+               `;
+        return { sql, params: schemaName ? { 'ISASchemaName': schemaName } : {} };
+    }
     /**
      * Processes IS-A relationship configurations from the additionalSchemaInfo config.
      * For each configured relationship, looks up both entities by name (or by table name
@@ -675,11 +733,11 @@ export class ManageMetadataBase {
         for (const rel of relationships) {
             try {
                 // Look up the parent entity — try by Name first, then by BaseTable within the given schema
-                const parentResult = await this.runQueryWithParams(pool, `
-                  ${this.selectTop(1, 'ID, Name', `FROM ${this.qs(schema, 'vwEntities')}
-                  WHERE Name = @ParentName
-                     OR (BaseTable = @ParentName AND (@SchemaName IS NULL OR SchemaName = @SchemaName))`, 'CASE WHEN Name = @ParentName THEN 0 ELSE 1 END')}
-               `, { 'ParentName': rel.ParentEntity, 'SchemaName': rel.SchemaName || null });
+                const parentLookup = this.buildISAEntityLookupSQL(schema, 'ID, Name', 'ParentName', rel.SchemaName);
+                const parentResult = await this.runQueryWithParams(pool, parentLookup.sql, {
+                    'ParentName': rel.ParentEntity,
+                    ...parentLookup.params,
+                });
                 if (parentResult.recordset.length === 0) {
                     logError(`    > IS-A config: parent entity "${rel.ParentEntity}" not found — skipping`);
                     continue;
@@ -687,11 +745,11 @@ export class ManageMetadataBase {
                 const parentId = parentResult.recordset[0].ID;
                 const parentName = parentResult.recordset[0].Name;
                 // Look up the child entity — same strategy
-                const childResult = await this.runQueryWithParams(pool, `
-                  ${this.selectTop(1, 'ID, Name, ParentID', `FROM ${this.qs(schema, 'vwEntities')}
-                  WHERE Name = @ChildName
-                     OR (BaseTable = @ChildName AND (@SchemaName IS NULL OR SchemaName = @SchemaName))`, 'CASE WHEN Name = @ChildName THEN 0 ELSE 1 END')}
-               `, { 'ChildName': rel.ChildEntity, 'SchemaName': rel.SchemaName || null });
+                const childLookup = this.buildISAEntityLookupSQL(schema, 'ID, Name, ParentID', 'ChildName', rel.SchemaName);
+                const childResult = await this.runQueryWithParams(pool, childLookup.sql, {
+                    'ChildName': rel.ChildEntity,
+                    ...childLookup.params,
+                });
                 if (childResult.recordset.length === 0) {
                     logError(`    > IS-A config: child entity "${rel.ChildEntity}" not found — skipping`);
                     continue;
@@ -951,6 +1009,695 @@ export class ManageMetadataBase {
         }
         return { success: true, createdCount };
     }
+    /**
+     * Processes base-view materialization declarations from additionalSchemaInfo (plan §4.1).
+     * Each declares a 1:1 snapshot of an existing entity's base view. Because the shape is
+     * identical to the source entity, NO new entity is minted — the existing entity is reused
+     * (its RLS applies unchanged, §6.1). For each declaration this:
+     *   1. resolves the source entity,
+     *   2. derives the column shape from the entity's base-view fields (SQLFullType),
+     *   3. emits the physical table (materialized_<Name>) + wrapper view (materialized_vw<Name>)
+     *      via the cross-engine DDL primitive (create-if-absent, so a migration-provided table
+     *      is reused rather than clobbered — §12),
+     *   4. upserts the "MJ: Materialized Results" row (SourceType='EntityBaseView') keyed on the
+     *      source entity.
+     * Population is the refresh driver's job (a later phase); rows land in Status='Building'.
+     */
+    async processBaseViewMaterializations(pool) {
+        const config = ManageMetadataBase.getSoftPKFKConfig();
+        if (!config)
+            return { success: true, processedCount: 0 };
+        const decls = this.extractMaterializedBaseViewsFromConfig(config);
+        if (decls.length === 0)
+            return { success: true, processedCount: 0 };
+        // Gate on the MaterializedResult table existing. Base-view materializations are DECLARED in config, but the
+        // target DB may not have the materialization migration applied yet (e.g. the PostgreSQL parallel world, or a
+        // fresh env sharing a config). The MaterializedResult upsert below would otherwise throw and abort the whole
+        // codegen pass. Unlike the unconditional gates on the sibling methods this one logs, because the user asked
+        // for these materializations — a silent skip would look like the feature quietly did nothing.
+        if (!(await this.materializedResultTableExists(pool))) {
+            logStatus(`    > Base-view materialization: ${decls.length} declaration${decls.length === 1 ? '' : 's'} in config but the MaterializedResult table is absent on this database — skipping (has the materialization migration been applied?)`);
+            return { success: true, processedCount: 0 };
+        }
+        // Enumerate the API-key row-filter layer BEFORE any leak gate below evaluates it. Until this runs the cache
+        // reads 'unknown' and every entity is treated as row-restricted (fail closed) — see loadAPIKeyRowFilterTargets.
+        await this.loadAPIKeyRowFilterTargets(pool);
+        // Refresh so entity.Fields reflect this run's field sync before we snapshot the shape.
+        const md = new Metadata(); // global-provider-ok: codegen runs offline against a single provider
+        await md.Refresh();
+        const coreSchema = mj_core_schema();
+        const esc = (s) => s.replace(/'/g, "''");
+        const lit = (s) => (s ? `'${esc(s)}'` : 'NULL');
+        let processedCount = 0;
+        for (const decl of decls) {
+            // 1) Resolve the source entity (by name, else by base table [+ schema]).
+            const entity = decl.EntityName
+                ? md.EntityByName(decl.EntityName)
+                : md.Entities.find((e) => !!decl.BaseTable &&
+                    (e.BaseTable ?? '').trim().toLowerCase() === decl.BaseTable.trim().toLowerCase() &&
+                    (!decl.SchemaName || (e.SchemaName ?? '').trim().toLowerCase() === decl.SchemaName.trim().toLowerCase()));
+            if (!entity) {
+                logError(`    > Base-view materialization: source entity not found (${decl.EntityName ?? `${decl.SchemaName ?? coreSchema}.${decl.BaseTable}`}) — skipping`);
+                continue;
+            }
+            // SECURITY (Leak 1 / C2 for base views): refuse to materialize an EXTERNAL entity that is
+            // read-RLS-protected. An external entity's live reads are REFUSED under RLS (the read path throws —
+            // MJ can't enforce RLS WHERE clauses on a remote system), and for external entities the read path
+            // returns at the external-dispatch branch BEFORE any materialized-view swap — so the "base-view reads
+            // re-apply source RLS" exemption is unreachable here. Mirroring the remote rows into a local table
+            // would expose, via any raw query joining the wrapper view, data the live path refuses entirely.
+            // (A LOCAL RLS base-view is safe: its read path swaps to the mirror and re-applies the entity's RLS.)
+            if (entity.ExternalDataSourceID && this.entityHasRowLevelRestriction(entity)) {
+                logError(`    > REFUSING base-view materialization of external entity "${entity.Name}": it is row-restricted (role RLS and/or an API-key row filter) and its live reads are refused under RLS, so a local mirror would leak its rows via raw queries over the wrapper view (a mirror cannot reproduce remote row restrictions). Skipping.`);
+                continue;
+            }
+            // Schema-name symmetry with the refresh path: MaterializationRefresher validates schema/table/view names
+            // as plain SQL identifiers (^[A-Za-z_][A-Za-z0-9_]*$) before building any DDL. The mint quotes at the
+            // dialect level and would otherwise accept a schema name (e.g. "Sales Data", "crm-prod") the refresh then
+            // permanently rejects — the entity would provision cleanly and never refresh (a silent nightly failure).
+            // Refuse up front, symmetrically, with a clear message. (TableName/ViewName are CodeName-derived and
+            // always identifier-safe, so only the source entity's schema can be non-conforming here.)
+            if (!/^[A-Za-z_][A-Za-z0-9_]*$/.test(entity.SchemaName ?? '')) {
+                logError(`    > REFUSING base-view materialization of entity "${entity.Name}": its schema "${entity.SchemaName}" is not a plain SQL identifier — the refresh path requires one, so materialization is only supported for entities in identifier-named schemas. Skipping.`);
+                continue;
+            }
+            // 2) Derive the snapshot column shape from the entity's base-view fields.
+            // For EXTERNAL entities, virtual fields are MJ-computed and absent from the remote source, so the
+            // refresh mirror (rebuildFromExternalEntity) only materializes non-virtual columns — the mint column
+            // set MUST match, or the wrapper view would expose virtual columns the refreshed table lacks (breaking
+            // virtual-field reads at the DB AND permanently tripping DriftHold on the next drift scan). LOCAL base
+            // views compute virtual columns in the view itself, so `SELECT * FROM <baseView>` includes them and the
+            // mint keeps all fields — the mint/refresh column sets stay symmetric on both paths.
+            const columns = this.materializedEntityFields(entity)
+                .map((f) => ({
+                Name: f.Name,
+                SQLType: f.SQLFullType,
+                Nullable: f.AllowsNull,
+                IsPrimaryKey: f.IsPrimaryKey,
+            }));
+            if (columns.length === 0) {
+                logError(`    > Base-view materialization for "${entity.Name}": no fields resolved — skipping`);
+                continue;
+            }
+            // 3) Emit the physical table + wrapper view (create-if-absent; reuse a migration-provided table).
+            const matSchema = entity.SchemaName;
+            const tableName = `materialized_${entity.CodeName}`;
+            const viewName = `materialized_vw${entity.CodeName}`;
+            const tableSQL = this.dbProvider.generateMaterializedTableSQL(matSchema, tableName, columns);
+            const viewSQL = this.dbProvider.generateMaterializedWrapperViewSQL(matSchema, viewName, tableName);
+            // includeBatchSeparator: each is a single GO-free batch (executed via ds.query), but the
+            // migration file needs a GO between statements for Flyway/sqlcmd.
+            await this.LogSQLAndExecute(pool, tableSQL, `Create materialized table for base-view materialization of entity ${entity.Name}`, false, true);
+            await this.LogSQLAndExecute(pool, viewSQL, `Create wrapper view for base-view materialization of entity ${entity.Name}`, false, true);
+            // 4) Upsert the MJ: Materialized Results row, keyed on (SourceType, SourceEntityID).
+            const existing = await this.runQueryWithParams(pool, `SELECT ID FROM ${this.qs(coreSchema, 'MaterializedResult')} WHERE SourceType = @SourceType AND SourceEntityID = @EntityID`, { SourceType: 'EntityBaseView', EntityID: entity.ID });
+            if (existing.recordset.length > 0) {
+                const sqlUpd = `UPDATE ${this.qs(coreSchema, 'MaterializedResult')}
+                              SET SchemaName='${esc(matSchema)}', TableName='${esc(tableName)}', ViewName='${esc(viewName)}',
+                                  RefreshSchedule=${lit(decl.RefreshSchedule)}, IntendedWorkload=${lit(decl.IntendedWorkload)}
+                            WHERE ID='${existing.recordset[0].ID}'`;
+                await this.LogSQLAndExecute(pool, sqlUpd, `Update MJ: Materialized Results for base-view materialization of ${entity.Name}`);
+            }
+            else {
+                const newId = this.createNewUUID();
+                const q = (n) => this.qi(n);
+                const sqlIns = `INSERT INTO ${this.qs(coreSchema, 'MaterializedResult')} (
+                                 ${q('ID')}, ${q('SourceType')}, ${q('SourceEntityID')}, ${q('SchemaName')}, ${q('TableName')}, ${q('ViewName')},
+                                 ${q('ParamMode')}, ${q('RefreshStrategy')}, ${q('RefreshSchedule')}, ${q('Status')}, ${q('IntendedWorkload')},
+                                 ${q('__mj_CreatedAt')}, ${q('__mj_UpdatedAt')} )
+                            VALUES ( '${newId}', 'EntityBaseView', '${entity.ID}', '${esc(matSchema)}', '${esc(tableName)}', '${esc(viewName)}',
+                                 'None', 'FullRebuild', ${lit(decl.RefreshSchedule)}, 'Building', ${lit(decl.IntendedWorkload)},
+                                 ${this.utcNow()}, ${this.utcNow()} )`;
+                await this.LogSQLAndExecute(pool, sqlIns, `Insert MJ: Materialized Results for base-view materialization of ${entity.Name}`);
+            }
+            processedCount++;
+            logStatus(`    > Base-view materialization ready for entity "${entity.Name}" → [${matSchema}].[${viewName}]`);
+        }
+        return { success: true, processedCount };
+    }
+    /**
+     * Phase 2d — classifies a parameterized query's parameters (deterministic render-and-diff) and,
+     * when every parameter is a safe row filter, produces the BROAD source SQL the refresh engine
+     * materializes (the query with its row-filter WHERE predicates removed). Returns the persistence
+     * fields, or a precise refusal (per-parameter reasons) so non-qualifying queries are skipped loudly.
+     */
+    async classifyParameterizedQueryForMaterialization(pool, queryId, queryName, outputColumns, currentUser) {
+        const coreSchema = mj_core_schema();
+        const md = new Metadata(); // global-provider-ok: codegen runs offline against a single provider
+        // Load the query template + parameter definitions (typed entities for the renderer).
+        const qRes = await this.runQueryWithParams(pool, `SELECT SQL, UsesTemplate FROM ${this.qs(coreSchema, 'Query')} WHERE ID = @QID`, { QID: queryId });
+        const sqlText = qRes.recordset[0]?.SQL ?? '';
+        const usesTemplate = !!qRes.recordset[0]?.UsesTemplate;
+        // NB: QueryParameter has no Sequence column (unlike QueryField) — order by Name for deterministic,
+        // stable classification output. (Ordering only affects the persisted spec's element order; the
+        // read-time predicate is an AND of all entries, so it's order-independent for correctness.)
+        const paramRows = await this.runQueryWithParams(pool, `SELECT Name, Type, SampleValue, DefaultValue, IsRequired, ValidationFilters FROM ${this.qs(coreSchema, 'QueryParameter')} WHERE QueryID = @QID ORDER BY Name`, { QID: queryId });
+        const queryParams = [];
+        for (const row of paramRows.recordset) {
+            const pe = await md.GetEntityObject('MJ: Query Parameters', currentUser);
+            pe.LoadFromData(row);
+            queryParams.push(pe);
+        }
+        // Platform-aware Nunjucks renderer — the same SQL-safe filters production uses.
+        const platform = this.dbProvider.PlatformKey;
+        RunQuerySQLFilterManager.Instance.SetPlatform(platform);
+        const dialect = platform === 'postgresql' ? new PostgreSQLDialect() : new SQLServerDialect();
+        const templateInput = { SQL: sqlText, UsesTemplate: usesTemplate, Parameters: queryParams };
+        const render = (vals) => {
+            const res = QueryParameterProcessor.processQueryTemplate(templateInput, vals, undefined, true);
+            if (!res.success) {
+                throw new Error(res.error ?? 'template render failed');
+            }
+            return res.processedSQL;
+        };
+        const paramDefs = queryParams.map((p) => ({ Name: p.Name, Type: p.Type, SampleValue: p.SampleValue }));
+        // Phase 2 is shipped: enable Bucket-1 row-filter broad materialization. The provider now auto-injects
+        // the read-time predicate from the persisted ReadFilterSpec (bound params), so a RowFilterBroad
+        // materialization is safe to mint. `allowRowFilterBroad` remains the build-level kill switch.
+        const classification = classifyQueryParameters({ queryName, params: paramDefs, outputColumns, dialect, render, allowRowFilterBroad: true });
+        if (!classification.qualification.qualifies) {
+            const detail = classification.perParam.map((pp) => `${pp.name}: ${pp.verdict.reason}`).join('; ');
+            return { qualifies: false, reason: `${classification.qualification.reason ?? 'parameters not materializable'}${detail ? ` — [${detail}]` : ''}` };
+        }
+        if (classification.qualification.paramMode !== 'RowFilterBroad') {
+            return { qualifies: false, reason: `parameterization mode '${classification.qualification.paramMode}' is not supported for materialization in v1 (only RowFilterBroad)` };
+        }
+        // Build the broad source SQL: render a concrete instance (held values) then strip the row-filter predicates.
+        let renderedHeld;
+        try {
+            renderedHeld = render(buildHeldValues(paramDefs));
+        }
+        catch (e) {
+            return { qualifies: false, reason: `could not render the query to build broad SQL: ${e instanceof Error ? e.message : String(e)}` };
+        }
+        const expectedRemovals = classification.qualification.rowFilterColumns.length;
+        const broad = buildBroadRowFilterSQL(renderedHeld, classification.qualification.rowFilterColumns, dialect, expectedRemovals);
+        if (broad.removedCount === 0) {
+            return { qualifies: false, reason: `expected to strip row-filter predicate(s) on [${classification.qualification.rowFilterColumns.join(', ')}] but none were removed from the rendered SQL — refusing to avoid a wrongly-filtered materialization` };
+        }
+        if (broad.ambiguous) {
+            return { qualifies: false, reason: `expected to strip exactly ${expectedRemovals} row-filter parameter predicate(s) on [${classification.qualification.rowFilterColumns.join(', ')}] but matched ${broad.removedCount} — the parameter predicate(s) cannot be cleanly isolated from other static or same-named predicates on those columns, so a broad materialization would include or exclude rows the live query never would. Refusing (query stays live-only).` };
+        }
+        return { qualifies: true, rowFilterColumns: classification.qualification.rowFilterColumns, broadSQL: broad.sql, readFilterSpec: classification.qualification.readFilterSpec };
+    }
+    /**
+     * Processes query materialization (CodeGen materialization phase, sub-step C — plan §4.2).
+     * Scans queries flagged `IsMaterialized = 1` and, for each that *qualifies* (unparameterized,
+     * has declared output fields — see analyzeQueryForMaterialization, §9/§10):
+     *   1. derives the materialized table's column shape + synthetic surrogate key,
+     *   2. emits the physical table (materialized_<Name>) + wrapper view (materialized_vw<Name>)
+     *      via the cross-engine DDL primitive,
+     *   3. mints a read-only Virtual Entity over the wrapper view (idempotent — reuses an existing
+     *      one), linked back to the source Query,
+     *   4. upserts the "MJ: Materialized Results" row (SourceType='Query') and its
+     *      MaterializedResultQuery join row linking it to the source Query (no direct
+     *      SourceQueryID / MaterializedResultID FK columns — the join table breaks the cycle).
+     * Runs BEFORE manageVirtualEntities so the minted entity's fields get synced from the wrapper
+     * view in the same run. Population is the refresh driver's job (later phase); Status='Building'.
+     * Parameterized queries are skipped with a log (deferred to Phase 2).
+     */
+    async processQueryMaterializations(pool, currentUser) {
+        const coreSchema = mj_core_schema();
+        // Gate the whole method on the IsMaterialized column existing — the flagged-query SELECT below filters on it,
+        // and it (like the MaterializedResult table) is added by the materialization Foundation migration. On a DB
+        // where that migration hasn't run yet the SELECT would throw and abort the entire codegen pass; with no such
+        // column there are no materializations to process anyway. Same defensive pattern as the EDS-column gate below.
+        if (!(await this.queryHasIsMaterializedColumn(pool)))
+            return { success: true, processedCount: 0, mintedCount: 0 };
+        // Gate the ExternalDataSourceID column ref so the SELECT stays valid on a DB without the EDS schema
+        // (PostgreSQL today). When absent, no query can be external, so the flag defaults false everywhere.
+        const queryHasEDSCol = await this.queryHasExternalDataSourceColumn(pool);
+        const flagged = await this.runQueryWithParams(pool, `SELECT ID, Name, SQL${queryHasEDSCol ? ', ExternalDataSourceID' : ''} FROM ${this.qs(coreSchema, 'Query')} WHERE IsMaterialized = ${this.boolLit(true)}`, {});
+        if (flagged.recordset.length === 0)
+            return { success: true, processedCount: 0, mintedCount: 0 };
+        // Symmetry with the three sibling gates: the per-query loop below also reads/writes the MaterializedResult
+        // TABLE (not just the IsMaterialized column). Don't lean on the "same Foundation migration ships both"
+        // coupling — confirm the table itself exists. If a query is flagged but the table is absent (partial schema),
+        // skip cleanly instead of throwing mid-loop and aborting the whole codegen pass.
+        if (!(await this.materializedResultTableExists(pool))) {
+            logStatus(`    > Query materialization: ${flagged.recordset.length} query(ies) flagged IsMaterialized but the MaterializedResult table is absent on this database — skipping (has the materialization migration been applied?)`);
+            return { success: true, processedCount: 0, mintedCount: 0 };
+        }
+        // Enumerate the API-key row-filter layer BEFORE the mint-time RLS gate consults it (see
+        // loadAPIKeyRowFilterTargets — an unloaded cache reads 'unknown' and refuses every materialization).
+        await this.loadAPIKeyRowFilterTargets(pool);
+        const esc = (s) => s.replace(/'/g, "''");
+        const idLit = (id) => (id ? `'${id}'` : 'NULL');
+        const md = new Metadata(); // global-provider-ok: codegen runs offline against a single provider
+        let processedCount = 0;
+        let mintedCount = 0;
+        for (const q of flagged.recordset) {
+            const queryId = q.ID;
+            const queryName = q.Name;
+            const querySQL = q.SQL ?? '';
+            // Qualifying inputs: parameterized? declared output fields?
+            const paramCountRes = await this.runQueryWithParams(pool, `SELECT COUNT(*) AS n FROM ${this.qs(coreSchema, 'QueryParameter')} WHERE QueryID = @QID`, { QID: queryId });
+            const isParameterized = (paramCountRes.recordset[0]?.n ?? 0) > 0;
+            const fieldsRes = await this.runQueryWithParams(pool, `SELECT Name, SQLFullType, IsComputed FROM ${this.qs(coreSchema, 'QueryField')} WHERE QueryID = @QID ORDER BY Sequence`, { QID: queryId });
+            const fields = fieldsRes.recordset.map((r) => ({ Name: r.Name, SQLFullType: r.SQLFullType, IsComputed: !!r.IsComputed }));
+            // Parameterized queries (Phase 2d): classify via deterministic render-and-diff. A query whose
+            // parameters are all safe row filters materializes BROAD (the row-filter predicate is removed)
+            // and the filter is re-applied at read; anything not provably a safe row filter is refused with
+            // a precise reason. Unparameterized queries skip this entirely.
+            let paramMode = 'None';
+            let rowFilterColumns = [];
+            let broadSQL = null;
+            let readFilterSpec = [];
+            if (isParameterized) {
+                const pc = await this.classifyParameterizedQueryForMaterialization(pool, queryId, queryName, fields.map((f) => f.Name), currentUser);
+                if (!pc.qualifies) {
+                    logStatus(`    > Skipping materialization of parameterized query "${queryName}": ${pc.reason}`);
+                    continue;
+                }
+                paramMode = 'RowFilterBroad';
+                rowFilterColumns = pc.rowFilterColumns;
+                broadSQL = pc.broadSQL;
+                readFilterSpec = pc.readFilterSpec;
+            }
+            // Phase 3: detect an aggregation key (the grouping columns) so the refresh can compute a stable
+            // hash surrogate (the match key for incremental / dirty-group refresh) instead of a synthetic
+            // row id. Detect against the BROAD SQL for a row-filter query (the actually-materialized shape),
+            // else the raw query SQL. NULL = not keyed → synthetic surrogate + full rebuild (always correct).
+            // Detected BEFORE the shape analysis so the surrogate column can be typed for the keyed case.
+            const detectSQL = paramMode === 'RowFilterBroad' ? (broadSQL ?? '') : querySQL;
+            const matDialect = this.dbProvider.PlatformKey === 'postgresql' ? new PostgreSQLDialect() : new SQLServerDialect();
+            const detectedKeyColumns = detectSQL ? detectAggregationKeyColumns({ sql: detectSQL, dialect: matDialect, fields }) : null;
+            // An EXTERNAL query refreshes via rebuildFromExternalQuery, which fetches all rows through the EDS
+            // driver and writes a SYNTHETIC row-index surrogate — it cannot compute the per-group combined-key
+            // hash and ignores KeyColumns entirely. So minting a keyed (hash-surrogate) entity for an external
+            // query would type the PK varchar/text while the refresh fills it with int row indices (a type +
+            // semantics mismatch). Force such queries un-keyed → synthetic surrogate + FullRebuild, which the
+            // external refresh path supports. (Incremental keyed refresh needs a local __mj_UpdatedAt watermark
+            // the EDS router doesn't expose anyway, so nothing is lost.)
+            const isExternalQuery = queryHasEDSCol && !!q.ExternalDataSourceID;
+            const keyColumns = isExternalQuery ? null : detectedKeyColumns;
+            if (isExternalQuery && detectedKeyColumns && detectedKeyColumns.length > 0) {
+                logStatus(`    > Query "${queryName}" is external + aggregation-keyed; the external refresh path uses a synthetic row-index surrogate, so materializing it un-keyed (FullRebuild) to keep the PK type/value consistent.`);
+            }
+            const isKeyed = !!keyColumns && keyColumns.length > 0;
+            const keyColumnsLit = isKeyed ? `'${esc(JSON.stringify(keyColumns))}'` : 'NULL';
+            // Surrogate PK type MUST match what the refresh actually writes into it: a KEYED materialization's
+            // refresh writes the combined-key SHA-256 hash (a fixed-width hex string), so the minted entity's
+            // PK is typed as the hash type; an un-keyed one keeps the synthetic auto-identity. Typing it int
+            // for a keyed materialization would diverge from the post-refresh varchar/text column (breaking PK
+            // typing, keyset pagination, and record lookups against the minted entity).
+            const surrogateSQLType = isKeyed
+                ? this.dbProvider.getMaterializedHashSurrogateColumnType()
+                : this.dbProvider.getMaterializedSurrogateColumnType();
+            // The materialized table's column shape is the query's output columns — independent of
+            // parameterization (a row-filter query is materialized broad over the same columns). Pass
+            // isParameterized:false so the analyzer yields the shape; the param gate above already ran.
+            const analysis = analyzeQueryForMaterialization({ queryName, isParameterized: false, fields, surrogateSQLType });
+            if (!analysis.qualifies) {
+                logStatus(`    > Skipping materialization of query "${queryName}": ${analysis.reason}`);
+                continue;
+            }
+            // SQL literals for the row-filter persistence columns (shared by the insert/update branches).
+            const rowFilterColumnsLit = rowFilterColumns.length > 0 ? `'${esc(JSON.stringify(rowFilterColumns))}'` : 'NULL';
+            const broadSQLLit = broadSQL ? `'${esc(broadSQL)}'` : 'NULL';
+            // Phase 2: the structured read-filter spec the provider injects at read time (bound params).
+            const readFilterSpecLit = readFilterSpec.length > 0 ? `'${esc(JSON.stringify(readFilterSpec))}'` : 'NULL';
+            // Phase 3/4: a keyed aggregation over a SINGLE source table is eligible for incremental in-place
+            // refresh (the refresher still guards at runtime: baseline watermark + delete detection).
+            // Joins/multi-table sources can't be localized by one watermark → keep FullRebuild (§10 bias).
+            // Among eligible ones, a PURELY ADDITIVE aggregation (only SUM/COUNT) uses the MERGE-upsert
+            // 'Incremental' path (in-place, no row churn); non-additive uses 'DirtyGroupRecompute' (delete+insert).
+            const isKeyedSingleSource = isKeyed && !!detectSQL && SQLParser.ExtractTableRefs(detectSQL, matDialect).length === 1;
+            const refreshStrategy = !isKeyedSingleSource
+                ? 'FullRebuild'
+                : detectAdditiveMeasures(detectSQL)
+                    ? 'Incremental'
+                    : 'DirtyGroupRecompute';
+            // RLS downgrade gate (§6.2, ships day one) — a materialized query entity does NOT inherit its
+            // source entities' row-level security. Refuse to materialize when RLS-safety cannot be proven, to
+            // prevent a silent privilege escalation. Asymmetric risk (§10): over-restriction is harmless; the
+            // dangerous direction (protected → unprotected) must be loud. The SAME assessment is re-run on every
+            // CodeGen pass by detectMaterializationDrift, so a source that LATER gains RLS holds the existing
+            // materialization AND revokes its read access rather than leaking (see assessQuerySourceRLSSafety).
+            const sourceLinks = await this.runQueryWithParams(pool, `SELECT EntityID FROM ${this.qs(coreSchema, 'vwQueryEntities')} WHERE QueryID = @QID`, { QID: queryId });
+            const rlsVerdict = this.assessQuerySourceRLSSafety(md, sourceLinks.recordset.map((r) => r.EntityID), detectSQL);
+            if (!rlsVerdict.safe) {
+                logError(`    > REFUSING to materialize query "${queryName}": ${rlsVerdict.reason}. Run query analysis so its source entities are linked, or use a base-view materialization (which inherits source RLS). Skipping to avoid a silent privilege escalation.`);
+                continue;
+            }
+            // The resolved source entities (all guaranteed to resolve — the gate above fails closed otherwise)
+            // scope the minted entity's read permissions to the INTERSECTION of source read access (C2). Without
+            // this, the minted entity would receive broad config-default read permissions and a user who cannot
+            // read a source entity could read its rows through the unscoped snapshot.
+            const resolvedSourceEntities = sourceLinks.recordset
+                .map((r) => md.EntityByID(r.EntityID))
+                .filter((e) => !!e);
+            // 2) Physical table + wrapper view (create-if-absent; reuse a migration-provided table — §12).
+            const codeName = CodeNameFromString(queryName);
+            const tableName = `materialized_${codeName}`;
+            const viewName = `materialized_vw${codeName}`;
+            // CodeName-collision guard (C3): two query names that normalize to the SAME CodeName (e.g. "Sales
+            // Report" / "Sales-Report" / "sales report") derive the same physical table. Minting the second would
+            // reuse the first's create-if-absent table (wrong columns) and its minted entity, then register a
+            // second MaterializedResult pointing at the SAME table — two refreshers clobbering one snapshot from
+            // different source SQL. Refuse if the derived table is already owned by a different materialization.
+            const tableOwners = await this.runQueryWithParams(pool, `SELECT mr.SourceType, mrq.QueryID FROM ${this.qs(coreSchema, 'MaterializedResult')} mr
+                LEFT JOIN ${this.qs(coreSchema, 'MaterializedResultQuery')} mrq ON mrq.MaterializedResultID = mr.ID
+                WHERE mr.SchemaName = @S AND mr.TableName = @T`, { S: coreSchema, T: tableName });
+            const foreignOwner = tableOwners.recordset.find((row) => !(row.SourceType === 'Query' && UUIDsEqual(row.QueryID, queryId)));
+            if (foreignOwner) {
+                logError(`    > REFUSING to materialize query "${queryName}": its derived table [${coreSchema}].[${tableName}] (from CodeName "${codeName}") is already owned by a different materialization (SourceType=${foreignOwner.SourceType}, QueryID=${foreignOwner.QueryID ?? 'n/a'}). Two sources whose names normalize to the same CodeName would clobber each other's snapshot — rename one so their CodeNames differ. Skipping.`);
+                continue;
+            }
+            const tableSQL = this.dbProvider.generateMaterializedTableSQL(coreSchema, tableName, analysis.columns);
+            const viewSQL = this.dbProvider.generateMaterializedWrapperViewSQL(coreSchema, viewName, tableName);
+            await this.LogSQLAndExecute(pool, tableSQL, `Create materialized table for query "${queryName}"`, false, true);
+            await this.LogSQLAndExecute(pool, viewSQL, `Create wrapper view for query "${queryName}"`, false, true);
+            // 3) Mint the read-only Virtual Entity over the wrapper view (idempotent by view).
+            const existingVE = await this.runQueryWithParams(pool, `SELECT ID FROM ${this.qs(coreSchema, 'vwEntities')} WHERE BaseView = @V AND SchemaName = @S`, { V: viewName, S: coreSchema });
+            let generatedEntityId = existingVE.recordset[0]?.ID;
+            if (!generatedEntityId) {
+                try {
+                    const createRes = await pool.executeStoredProcedure(`${this.qs(coreSchema, 'spCreateVirtualEntity')}`, {
+                        Name: queryName,
+                        BaseView: viewName,
+                        SchemaName: coreSchema,
+                        PrimaryKeyFieldName: analysis.surrogateColumnName,
+                        Description: `Materialized result of query "${queryName}".`,
+                    });
+                    generatedEntityId = createRes.recordset?.[0]?.[''] || createRes.recordset?.[0]?.ID || createRes.recordset?.[0]?.Column0;
+                    if (generatedEntityId) {
+                        mintedCount++;
+                        await this.addEntityToApplicationForSchema(pool, generatedEntityId, queryName, coreSchema, currentUser);
+                        await this.addMaterializedQueryEntityPermissions(pool, generatedEntityId, queryName, resolvedSourceEntities);
+                        logStatus(`    > Minted materialized entity "${queryName}" (ID: ${generatedEntityId}) over [${coreSchema}].[${viewName}]`);
+                    }
+                }
+                catch (err) {
+                    logError(`    > Failed to mint materialized entity for query "${queryName}": ${err instanceof Error ? err.message : String(err)}`);
+                    continue;
+                }
+            }
+            // If the entity was neither pre-existing nor successfully minted-and-identified, we CANNOT safely link
+            // a MaterializedResult — its GeneratedEntityID would be NULL and DataSource:'Materialized' reads could
+            // never resolve it — and spCreateVirtualEntity may have created a physical entity we failed to read the
+            // ID of (an unrecognized result shape). Fail loud and skip rather than silently writing a broken
+            // MaterializedResult row; a human may need to check for an orphaned virtual entity over the view.
+            if (!generatedEntityId) {
+                logError(`    > Could not resolve the entity ID for materialized query "${queryName}" (spCreateVirtualEntity returned an unrecognized result shape). Skipping to avoid a MaterializedResult with a NULL GeneratedEntityID; check for an orphaned virtual entity over [${coreSchema}].[${viewName}].`);
+                continue;
+            }
+            // 4) Upsert the MJ: Materialized Results row and its MaterializedResultQuery join row. The MR<->Query
+            //    link lives in the join table — there is no MaterializedResult.SourceQueryID or
+            //    Query.MaterializedResultID column (those direct FKs formed a circular dependency CodeGen rejects;
+            //    the join table carries the relationship with both FKs pointing outward).
+            const existing = await this.runQueryWithParams(pool, `SELECT mr.ID FROM ${this.qs(coreSchema, 'MaterializedResult')} mr
+                JOIN ${this.qs(coreSchema, 'MaterializedResultQuery')} mrq ON mrq.MaterializedResultID = mr.ID
+                WHERE mr.SourceType = @T AND mrq.QueryID = @QID`, { T: 'Query', QID: queryId });
+            let matResultId;
+            if (existing.recordset.length > 0) {
+                matResultId = existing.recordset[0].ID;
+                const sqlUpd = `UPDATE ${this.qs(coreSchema, 'MaterializedResult')}
+                              SET GeneratedEntityID=${idLit(generatedEntityId)}, SchemaName='${esc(coreSchema)}', TableName='${esc(tableName)}', ViewName='${esc(viewName)}',
+                                  ${this.qi('ParamMode')}='${paramMode}', ${this.qi('RowFilterColumns')}=${rowFilterColumnsLit}, ${this.qi('BroadSQL')}=${broadSQLLit}, ${this.qi('ReadFilterSpec')}=${readFilterSpecLit}, ${this.qi('KeyColumns')}=${keyColumnsLit}, ${this.qi('RefreshStrategy')}='${refreshStrategy}'
+                            WHERE ID='${matResultId}'`;
+                await this.LogSQLAndExecute(pool, sqlUpd, `Update MJ: Materialized Results for query "${queryName}"`);
+            }
+            else {
+                matResultId = this.createNewUUID();
+                const c = (n) => this.qi(n);
+                const sqlIns = `INSERT INTO ${this.qs(coreSchema, 'MaterializedResult')} (
+                                 ${c('ID')}, ${c('SourceType')}, ${c('GeneratedEntityID')}, ${c('SchemaName')}, ${c('TableName')}, ${c('ViewName')},
+                                 ${c('ParamMode')}, ${c('RowFilterColumns')}, ${c('BroadSQL')}, ${c('ReadFilterSpec')}, ${c('KeyColumns')}, ${c('RefreshStrategy')}, ${c('Status')}, ${c('__mj_CreatedAt')}, ${c('__mj_UpdatedAt')} )
+                            VALUES ( '${matResultId}', 'Query', ${idLit(generatedEntityId)}, '${esc(coreSchema)}', '${esc(tableName)}', '${esc(viewName)}',
+                                 '${paramMode}', ${rowFilterColumnsLit}, ${broadSQLLit}, ${readFilterSpecLit}, ${keyColumnsLit}, '${refreshStrategy}', 'Building', ${this.utcNow()}, ${this.utcNow()} )`;
+                await this.LogSQLAndExecute(pool, sqlIns, `Insert MJ: Materialized Results for query "${queryName}"`);
+                // Link the new materialization to its source Query via the join table.
+                const joinId = this.createNewUUID();
+                const sqlJoinIns = `INSERT INTO ${this.qs(coreSchema, 'MaterializedResultQuery')} (
+                                 ${c('ID')}, ${c('MaterializedResultID')}, ${c('QueryID')}, ${c('__mj_CreatedAt')}, ${c('__mj_UpdatedAt')} )
+                            VALUES ( '${joinId}', '${matResultId}', '${queryId}', ${this.utcNow()}, ${this.utcNow()} )`;
+                await this.LogSQLAndExecute(pool, sqlJoinIns, `Link MJ: Materialized Results to Query "${queryName}" via join row`);
+            }
+            processedCount++;
+        }
+        // Refresh so manageVirtualEntities (next step) syncs fields for the newly-minted entities.
+        if (mintedCount > 0) {
+            await md.Refresh();
+        }
+        return { success: true, processedCount, mintedCount };
+    }
+    /**
+     * Phase 4 (§13/§17.2): scan active materializations and flag those whose source shape has drifted
+     * as `DriftHold` (stop refreshing, surface for review). Runs after the entity/field re-sync, so it
+     * compares each materialization's provenance against the CURRENT metadata. Flag-and-hold only — never
+     * auto-rebuilds. Already-held (`DriftHold`) and `Disabled` rows are skipped.
+     */
+    async detectMaterializationDrift(pool) {
+        const coreSchema = mj_core_schema();
+        // Gate on MaterializedResult existing — the SELECT below reads it. Absent (materialization migration not yet
+        // applied on this DB) ⇒ no materializations exist to check for drift; skip cleanly rather than throwing and
+        // aborting the codegen pass.
+        if (!(await this.materializedResultTableExists(pool)))
+            return { success: true, heldCount: 0 };
+        // Enumerate the API-key row-filter layer BEFORE the per-row C1 re-check / leak guard consults it (see
+        // loadAPIKeyRowFilterTargets — an unloaded cache reads 'unknown' and holds every materialization).
+        await this.loadAPIKeyRowFilterTargets(pool);
+        const rows = await this.runQuery(pool, `SELECT mr.ID, mr.SourceType, mr.SourceEntityID, mrq.QueryID AS SourceQueryID, mr.GeneratedEntityID, mr.SchemaName, mr.TableName
+             FROM ${this.qs(coreSchema, 'MaterializedResult')} mr
+             LEFT JOIN ${this.qs(coreSchema, 'MaterializedResultQuery')} mrq ON mrq.MaterializedResultID = mr.ID
+             WHERE mr.Status NOT IN ('Disabled', 'DriftHold')`);
+        if (rows.recordset.length === 0)
+            return { success: true, heldCount: 0 };
+        const md = new Metadata(); // global-provider-ok: codegen runs offline against a single provider
+        await md.Refresh(); // reflect this run's schema/field sync before judging drift
+        let heldCount = 0;
+        let failedCount = 0;
+        // Per-row isolation: a throw while processing ONE materialization (a transient query read, a dropped
+        // source table, a permission edge) must NOT abort reconciliation for the REST — otherwise a single
+        // failing row indefinitely defers the C1 RLS re-check + C2 grant re-narrow for every materialization
+        // after it (a latent fail-open). Log, count, and continue; the pass reports PARTIAL reconciliation.
+        for (const r of rows.recordset) {
+            try {
+                if (await this.evaluateAndHoldDriftRow(pool, md, r, coreSchema)) {
+                    heldCount++;
+                }
+            }
+            catch (rowErr) {
+                failedCount++;
+                logError(`    > Drift/security re-check FAILED for materialization "${r.TableName}" (ID ${r.ID}) — isolated and skipped, continuing with the rest: ${rowErr instanceof Error ? rowErr.message : String(rowErr)}`);
+            }
+        }
+        if (failedCount > 0) {
+            logError(`    > detectMaterializationDrift: ${failedCount} materialization(s) failed the drift/security re-check and were skipped — reconciliation is PARTIAL this run (a source that gained RLS on a skipped row may still be granting read). Investigate the logged errors.`);
+        }
+        return { success: failedCount === 0, heldCount };
+    }
+    /**
+     * Processes ONE materialization row: the C1 RLS re-check, the C2 read-grant re-narrow, the external
+     * base-view leak guard, and generic shape/provenance drift. Returns true if the row was held. Extracted
+     * from {@link detectMaterializationDrift} so a throw on one row can be isolated by the caller's try/catch
+     * instead of aborting reconciliation for every remaining materialization.
+     */
+    async evaluateAndHoldDriftRow(pool, md, r, coreSchema) {
+        // RLS re-check (C1): a QUERY materialization whose source has SINCE gained a read RLS filter (or lost
+        // its source-entity provenance) would keep serving the full unscoped snapshot to every user — the
+        // mint-time gate cannot see a change made after minting. Re-run the same assessment here and, when it
+        // now fails, hold the materialization AND revoke the minted entity's read access: a precomputed
+        // snapshot cannot enforce per-user scoping, so fail closed until a human authors protection. Base-view
+        // materializations re-apply the source entity's RLS at read time (they reuse the source entity), so
+        // they are exempt from this check.
+        if (r.SourceType === 'Query' && !r.SourceQueryID) {
+            // FAIL CLOSED on LOST PROVENANCE. `SourceQueryID` comes from a LEFT JOIN on MaterializedResultQuery.
+            // `spDeleteQuery` cascade-deletes that join row, but everything the join row protected SURVIVES: the
+            // MaterializedResult, the minted read-only virtual entity, its EntityPermission read grants, and the
+            // already-POPULATED materialized_<Name> table. With a NULL QueryID the C1 RLS re-check and the C2 grant
+            // re-narrow below are both unreachable, `gatherDriftFacts`' Query branch returns all-empty arrays,
+            // `evaluateMaterializationDrift` returns `{drift:false}`, and Status stays 'Active' — so the guard whose
+            // whole job is to revoke read + hold could never fire again and the unscoped snapshot would keep serving
+            // indefinitely. A query materialization that cannot name its source query has, by definition, lost the
+            // provenance every one of those safety checks is computed from, so treat it as drift and drive it down
+            // the established revoke-then-hold path. Over-holding is harmless (a human can relink and re-activate);
+            // under-holding is the leak.
+            await this.revokeReadAndHoldMaterialization(pool, r, coreSchema, 'PROVENANCE DRIFT', 'its MaterializedResultQuery link is gone (the source query was deleted), so the RLS re-check and read-grant re-narrow can no longer be computed for it');
+            return true;
+        }
+        if (r.SourceType === 'Query' && r.SourceQueryID) {
+            const qe = await this.runQueryWithParams(pool, `SELECT EntityID FROM ${this.qs(coreSchema, 'QueryEntity')} WHERE QueryID = @Q`, { Q: r.SourceQueryID });
+            const sourceEntityIds = qe.recordset.map((x) => x.EntityID);
+            // Load the query text so the P1 under-linking guard can parse it for source tables the QueryEntity
+            // links may have missed (see assessQuerySourceRLSSafety). Best-effort: if the read fails, the
+            // linked-source checks still run — only a POSITIVE unlinked-RLS detection changes the verdict.
+            const qSqlRes = await this.runQueryWithParams(pool, `SELECT SQL FROM ${this.qs(coreSchema, 'Query')} WHERE ID = @Q`, { Q: r.SourceQueryID });
+            const driftSQL = qSqlRes.recordset?.[0]?.SQL;
+            const rlsVerdict = this.assessQuerySourceRLSSafety(md, sourceEntityIds, driftSQL);
+            if (!rlsVerdict.safe) {
+                await this.revokeReadAndHoldMaterialization(pool, r, coreSchema, 'RLS DRIFT', rlsVerdict.reason ?? 'source RLS drift');
+                return true; // already held for the leak; the shape-drift check below is moot
+            }
+            // Leak 2 (C2 ongoing): RLS is still safe, but a role may have LOST plain read on a source since mint.
+            // The intersection grant is computed once at mint and never re-narrowed by the RLS check above, so
+            // re-narrow it here to the CURRENT source-read intersection — otherwise a role that lost read on a
+            // source keeps reading the snapshot's rows it can no longer read live. Re-narrow only (never re-grant).
+            if (r.GeneratedEntityID) {
+                const sourceEntities = sourceEntityIds.map((id) => md.EntityByID(id)).filter((e) => !!e);
+                const revoked = await this.reconcileMaterializedQueryEntityReadGrants(pool, r.GeneratedEntityID, r.TableName, sourceEntities);
+                if (revoked > 0) {
+                    logStatus(`    > PERMISSION DRIFT: materialized entity "${r.TableName}" → revoked ${revoked} over-broad read grant(s) (a role can no longer read every source).`);
+                }
+            }
+        }
+        // SECURITY (Leak 1 drift side): an existing base-view materialization of an EXTERNAL read-RLS-protected
+        // entity leaks (the mirror is readable via raw queries, but the entity's live reads are RLS-refused and
+        // the read path never re-applies RLS to an external mirror). The mint gate refuses NEW ones; hold any
+        // that already exist (or whose entity became external/RLS after minting) so refresh stops — and flag for
+        // a human to drop the mirror. (Local base-view RLS is safe and not held.)
+        if (r.SourceType === 'EntityBaseView' && r.SourceEntityID) {
+            const bvEntity = md.EntityByID(r.SourceEntityID);
+            if (bvEntity && bvEntity.ExternalDataSourceID && this.entityHasRowLevelRestriction(bvEntity)) {
+                // EMPTY the mirror now — it holds remote rows the live path refuses under RLS, and DriftHold alone
+                // would only stop FUTURE refreshes while leaving the already-populated rows queryable via raw SQL
+                // over the wrapper view. Emptying (not dropping) removes the leaked data without breaking any object
+                // dependency; the wrapper view remains but returns nothing. Then DriftHold so refresh never refills it.
+                await this.LogSQLAndExecute(pool, `DELETE FROM ${this.qs(r.SchemaName, r.TableName)}`, `Empty external RLS base-view mirror "${r.TableName}" (leak guard — mirror exposed RLS-refused rows)`);
+                await this.LogSQLAndExecute(pool, `UPDATE ${this.qs(coreSchema, 'MaterializedResult')} SET ${this.qi('Status')}='DriftHold' WHERE ID='${r.ID}'`, `Flag external RLS base-view materialization "${r.TableName}" as DriftHold (leak guard)`);
+                logError(`    > RLS LEAK GUARD: base-view materialization "${r.TableName}" mirrors an EXTERNAL row-restricted entity ("${bvEntity.Name}" — role RLS and/or an API-key row filter) → mirror EMPTIED + DriftHold (refresh stopped). It exposed rows the live path refuses under RLS.`);
+                return true;
+            }
+        }
+        const facts = await this.gatherDriftFacts(pool, md, r, coreSchema);
+        const verdict = evaluateMaterializationDrift(facts);
+        if (verdict.drift) {
+            // Fail-closed on ANY hold of a QUERY materialization: once held, the row drops out of
+            // detectMaterializationDrift's `Status NOT IN ('DriftHold', ...)` scan, so the C1 RLS re-check and the
+            // C2 grant re-narrow above NEVER run for it again. A query virtual entity's reads are NOT status-gated
+            // (they always read `materialized_vw…`), so a still-readable held snapshot whose source LATER gains RLS
+            // (or whose role loses source read) would leak the full unscoped rows. Revoke read access on the hold to
+            // close that window now, mirroring the RLS-drift revoke above. Base-view mats reuse the source entity
+            // (RLS re-applied at read time via the status-gated effective base view), so they have no minted grant here.
+            if (r.SourceType === 'Query' && r.GeneratedEntityID) {
+                await this.revokeMaterializedEntityReadAccess(pool, r.GeneratedEntityID, r.TableName, `drift hold — ${verdict.reason ?? 'shape/provenance drift'}`);
+            }
+            await this.LogSQLAndExecute(pool, `UPDATE ${this.qs(coreSchema, 'MaterializedResult')} SET ${this.qi('Status')}='DriftHold' WHERE ID='${r.ID}'`, `Flag materialization "${r.TableName}" as DriftHold`);
+            logStatus(`    > DRIFT: materialization "${r.TableName}" → DriftHold — ${verdict.reason}`);
+            return true;
+        }
+        return false;
+    }
+    /**
+     * THE single rule for "which of an entity's fields get materialized" in a base-view materialization. Every
+     * site that computes a base-view column set — the MINT (physical table + wrapper view), the DRIFT comparison,
+     * and the runtime REFRESH mirror — must use this one predicate, or they judge each other's output as drift.
+     *
+     * Rule: for an EXTERNAL entity, virtual fields are MJ-computed and absent from the remote source, so the
+     * refresh mirror only materializes non-virtual columns and the mint must match. A LOCAL base view computes its
+     * virtual columns in the view itself, so `SELECT * FROM <baseView>` includes them and every field is kept.
+     * (`MaterializationRefresher.rebuildFromExternalEntity` applies `!f.IsVirtual` on the external path only — the
+     * same rule, expressed in a context where "external" is already established.)
+     */
+    materializedEntityFields(entity) {
+        return entity.Fields.filter((f) => !(entity.ExternalDataSourceID && f.IsVirtual));
+    }
+    /**
+     * The established fail-closed "stop serving this materialization" sequence, shared by every branch that must
+     * take a query materialization out of service.
+     *
+     * Revoke read FIRST, then flag DriftHold. The revoke is the security-critical action (it closes the leak — a
+     * snapshot serving unscoped source rows); DriftHold only stops FUTURE refreshes. Revoke-first means that if
+     * the second statement fails, the readable window is already closed (fail-safe) — whereas DriftHold-then-revoke
+     * would leave read OPEN if the revoke failed. A row with no minted entity has no grant row to flip; it is still
+     * held.
+     */
+    async revokeReadAndHoldMaterialization(pool, r, coreSchema, label, reason) {
+        if (r.GeneratedEntityID) {
+            await this.revokeMaterializedEntityReadAccess(pool, r.GeneratedEntityID, r.TableName, reason);
+        }
+        await this.LogSQLAndExecute(pool, `UPDATE ${this.qs(coreSchema, 'MaterializedResult')} SET ${this.qi('Status')}='DriftHold' WHERE ID='${r.ID}'`, `Flag materialization "${r.TableName}" as DriftHold (${label.toLowerCase()})`);
+        logError(`    > ${label}: materialization "${r.TableName}" → read access revoked + DriftHold — ${reason}`);
+    }
+    /** Gathers the drift-relevant existence facts for one materialization against current metadata. */
+    async gatherDriftFacts(pool, md, r, coreSchema) {
+        if (r.SourceType === 'EntityBaseView') {
+            const entity = r.SourceEntityID ? md.EntityByID(r.SourceEntityID) : undefined;
+            if (!entity) {
+                return { sourceType: 'EntityBaseView', baseView: { sourceEntityExists: false, currentEntityFields: [], materializedColumns: [] } };
+            }
+            const materializedColumns = await this.getMaterializedTableColumns(pool, r.SchemaName, r.TableName);
+            return {
+                sourceType: 'EntityBaseView',
+                // MUST use the same field-exclusion rule the MINT and the REFRESH use ({@link materializedEntityFields}).
+                // Comparing the UNFILTERED field list against the mirror's columns made every external entity with an
+                // IsVirtual field look like it had drifted the very first run after minting (added=[<virtual fields>])
+                // → Status='DriftHold' → permanent live-only fallback, and DriftHold rows are excluded from the sweep,
+                // so it could never recover. The comparison has to be apples-to-apples with what was actually built.
+                baseView: { sourceEntityExists: true, currentEntityFields: this.materializedEntityFields(entity).map((f) => f.Name), materializedColumns },
+            };
+        }
+        // Query case — provenance via QueryEntity / QueryField / QueryDependency, plus output-shape facts.
+        const missingSourceEntities = [];
+        const missingSourceFields = [];
+        const missingComposedQueries = [];
+        // Output-shape facts (§13 drift): the query's CURRENT declared output columns (ALL fields) vs the
+        // snapshot table's actual DATA columns. Detects a SELECT-list edit on a create-if-absent table that
+        // provenance existence checks can't see. Both empty-tolerant downstream.
+        let currentOutputColumns = [];
+        let materializedColumns = [];
+        if (r.SourceQueryID) {
+            const qid = r.SourceQueryID;
+            const qe = await this.runQueryWithParams(pool, `SELECT EntityID FROM ${this.qs(coreSchema, 'QueryEntity')} WHERE QueryID = @Q`, { Q: qid });
+            for (const row of qe.recordset) {
+                if (!md.EntityByID(row.EntityID))
+                    missingSourceEntities.push(row.EntityID);
+            }
+            // ONE QueryField read serves BOTH needs: the output-shape column set (ALL Names) and the provenance
+            // field check (rows whose SourceEntityID+SourceFieldName are both known — only those are judged, to
+            // avoid false positives from incomplete/LLM-inferred provenance; filtered in JS below).
+            const qf = await this.runQueryWithParams(pool, `SELECT Name, SourceEntityID, SourceFieldName FROM ${this.qs(coreSchema, 'QueryField')} WHERE QueryID = @Q`, { Q: qid });
+            // Both sides of the output-shape comparison must exclude the SAME system columns: the synthetic
+            // surrogate PK and any __mj_* columns are always present in the materialized table but are stripped
+            // from materializedColumns below. A query that legitimately PROJECTS a __mj_* column (e.g. SELECT
+            // __mj_UpdatedAt) would otherwise appear only in currentOutputColumns → a phantom "added" drift that
+            // false-flags DriftHold. Applying the identical filter here keeps the comparison apples-to-apples.
+            const surrogate = MATERIALIZATION_SURROGATE_COLUMN.toLowerCase();
+            const isOutputShapeColumn = (name) => {
+                const lc = name.trim().toLowerCase();
+                return lc !== surrogate && !lc.startsWith('__mj_');
+            };
+            currentOutputColumns = qf.recordset.map((row) => String(row.Name)).filter(isOutputShapeColumn);
+            for (const row of qf.recordset) {
+                if (row.SourceEntityID == null || row.SourceFieldName == null)
+                    continue;
+                const ent = md.EntityByID(row.SourceEntityID);
+                const fieldName = String(row.SourceFieldName);
+                if (!ent) {
+                    missingSourceFields.push(`${row.SourceEntityID}.${fieldName}`);
+                }
+                else if (!ent.Fields.find((f) => f.Name.trim().toLowerCase() === fieldName.trim().toLowerCase())) {
+                    missingSourceFields.push(`${ent.Name}.${fieldName}`);
+                }
+            }
+            const qd = await this.runQueryWithParams(pool, `SELECT DependsOnQueryID FROM ${this.qs(coreSchema, 'QueryDependency')} WHERE QueryID = @Q`, { Q: qid });
+            const depIds = qd.recordset.map((row) => row.DependsOnQueryID).filter((d) => !!d);
+            if (depIds.length > 0) {
+                // ONE set-based existence check instead of a SELECT per dependency (was an N+1 cascade — K
+                // round trips per materialization). DependsOnQueryID values are DB-sourced UUIDs; single-quote
+                // escaped defensively. Case-insensitive membership handles SS-uppercase vs stored casing.
+                const inList = depIds.map((d) => `'${d.replace(/'/g, "''")}'`).join(', ');
+                const existing = await this.runQuery(pool, `SELECT ID FROM ${this.qs(coreSchema, 'Query')} WHERE ID IN (${inList})`);
+                const existingSet = new Set(existing.recordset.map((row) => String(row.ID).toLowerCase()));
+                for (const d of depIds) {
+                    if (!existingSet.has(d.toLowerCase()))
+                        missingComposedQueries.push(d);
+                }
+            }
+            const rawCols = await this.getMaterializedTableColumns(pool, r.SchemaName, r.TableName);
+            // Same exclusion as currentOutputColumns (see isOutputShapeColumn above) — surrogate + __mj_* columns
+            // are not part of the query's declared output shape.
+            materializedColumns = rawCols.filter(isOutputShapeColumn);
+        }
+        return { sourceType: 'Query', query: { missingSourceEntities, missingSourceFields, missingComposedQueries, currentOutputColumns, materializedColumns } };
+    }
+    /** Actual column names of a materialized table (via INFORMATION_SCHEMA), for drift comparison. */
+    async getMaterializedTableColumns(pool, schema, table) {
+        const esc = (s) => s.replace(/'/g, "''");
+        const res = await this.runQuery(pool, `SELECT COLUMN_NAME AS cn FROM INFORMATION_SCHEMA.COLUMNS WHERE TABLE_SCHEMA='${esc(schema)}' AND TABLE_NAME='${esc(table)}' ORDER BY ORDINAL_POSITION`);
+        return res.recordset.map((row) => row.cn);
+    }
     /**
      * Derives an entity name from a view name by removing common prefixes (vw, v_)
      * and converting to a human-friendly format.
@@ -1108,6 +1855,13 @@ export class ManageMetadataBase {
             const md = new Metadata(); // global-provider-ok: codegen runs offline against a single provider
             await md.Refresh();
         }
+        // Config-driven query materialization — mint read-only Virtual Entities over materialized
+        // query results. Runs BEFORE manageVirtualEntities so the minted entities' fields get synced
+        // from their wrapper views in this same run (plan §4.2). Internally refreshes metadata when it mints.
+        const queryMatResult = await this.processQueryMaterializations(pool, currentUser);
+        if (queryMatResult.processedCount > 0) {
+            logStatus(`    > Query materialization: processed ${queryMatResult.processedCount} quer${queryMatResult.processedCount === 1 ? 'y' : 'ies'} (${queryMatResult.mintedCount} new entit${queryMatResult.mintedCount === 1 ? 'y' : 'ies'})`);
+        }
         const veResult = await this.manageVirtualEntities(pool);
         if (!veResult.success) {
             logError('   Error managing virtual entities');
@@ -1158,6 +1912,21 @@ export class ManageMetadataBase {
         if (organicKeyResult.createdCount > 0 || organicKeyResult.updatedCount > 0) {
             logStatus(`    > Organic keys: ${organicKeyResult.createdCount} created, ${organicKeyResult.updatedCount} updated from config`);
         }
+        // Config-driven base-view materialization — emit the physical table + wrapper view and
+        // attach an "MJ: Materialized Results" row to the existing entity. Runs AFTER entity fields
+        // are managed so the snapshot shape reflects the current base view (plan §4.1).
+        const baseViewMatResult = await this.processBaseViewMaterializations(pool);
+        if (baseViewMatResult.processedCount > 0) {
+            logStatus(`    > Base-view materialization: processed ${baseViewMatResult.processedCount} declaration${baseViewMatResult.processedCount === 1 ? '' : 's'} from config`);
+        }
+        // Phase 4 (§13/§17.2): flag materializations whose source shape drifted (source entity/field
+        // dropped/renamed, or a composed inner query removed) as DriftHold and stop refreshing them —
+        // flag-and-hold, never silent auto-rebuild. Runs after the entity/field re-sync so it sees the
+        // current schema.
+        const driftResult = await this.detectMaterializationDrift(pool);
+        if (driftResult.heldCount > 0) {
+            logStatus(`    > Materialization drift: held ${driftResult.heldCount} materialization${driftResult.heldCount === 1 ? '' : 's'} for review (Status=DriftHold)`);
+        }
         start = new Date();
         logStatus('   Syncing schema info from database...');
         if (!await this.updateSchemaInfoFromDatabase(pool, excludeSchemas)) {
@@ -1255,19 +2024,56 @@ export class ManageMetadataBase {
         }
         return { success: bSuccess, anyUpdates: anyUpdates };
     }
+    /**
+     * External-data-source analogue of {@link manageVirtualEntities}. For each entity backed by an
+     * external data source, introspect the REMOTE schema (via the EDS router resolved through the
+     * ClassFactory) and sync its `EntityField` rows to match — the remote equivalent of reading the
+     * local INFORMATION_SCHEMA for a virtual/view entity. No-op when there are no external entities.
+     */
+    /**
+     * Whether the current database's `Entity` table has the `ExternalDataSourceID` column. External
+     * Data Sources ship as a SQL Server migration only, so on any database/schema that predates it
+     * (PostgreSQL today) the column is absent — and ANY raw SQL referencing it throws
+     * "column does not exist" and aborts the entire CodeGen run, even for a pure MJ-DB schema with no
+     * external entities. Callers gate EDS-specific queries on this so CodeGen stays green everywhere
+     * and auto-activates once the column lands on other platforms (mirrors the PG stored-proc guard in
+     * metadataSupportObjects.ts). Cross-platform via INFORMATION_SCHEMA; cached for the run (the schema
+     * cannot change mid-run).
+     */
     async entityHasExternalDataSourceColumn(pool) {
-        if (this._entityHasExternalDataSourceColumn === null) {
-            // Check the VIEW the gated queries actually read (vwEntities), not the base Entity table — a
-            // schema where the table has the column but the view wasn't refreshed to expose it would still
-            // throw. (Matches the PG guard in metadataSupportObjects.ts, which also checks the view.)
-            const sql = `SELECT COUNT(*) AS ColExists FROM INFORMATION_SCHEMA.COLUMNS ` +
-                `WHERE TABLE_SCHEMA = '${mj_core_schema()}' AND TABLE_NAME = 'vwEntities' AND COLUMN_NAME = 'ExternalDataSourceID'`;
+        // Check the VIEW the gated queries actually read (vwEntities), not the base Entity table — a
+        // schema where the table has the column but the view wasn't refreshed to expose it would still
+        // throw. (Matches the PG guard in metadataSupportObjects.ts, which also checks the view.)
+        return await this.columnExistsInCoreSchema(pool, 'vwEntities', 'ExternalDataSourceID');
+    }
+    /**
+     * Whether the current database's `Query` table has the `ExternalDataSourceID` column. Same rationale as
+     * {@link entityHasExternalDataSourceColumn}: External Data Sources ship as a SQL-Server-only migration, so
+     * on a DB without it (PostgreSQL today) the column is absent and any raw SQL referencing it aborts the
+     * CodeGen run. processQueryMaterializations gates its ExternalDataSourceID reference on this. Cached per run.
+     */
+    async queryHasExternalDataSourceColumn(pool) {
+        return await this.columnExistsInCoreSchema(pool, 'Query', 'ExternalDataSourceID');
+    }
+    /**
+     * True if the Query table has the IsMaterialized column (added by the materialization Foundation migration).
+     * processQueryMaterializations gates on this so codegen doesn't throw on a DB where that migration hasn't run
+     * yet (e.g. the PostgreSQL parallel world's object-availability lag) — same defensive pattern as
+     * queryHasExternalDataSourceColumn. Cached per run.
+     */
+    async queryHasIsMaterializedColumn(pool) {
+        return await this.columnExistsInCoreSchema(pool, 'Query', 'IsMaterialized');
+    }
+    async materializedResultTableExists(pool) {
+        if (this._materializedResultTableExists === null) {
+            const sql = `SELECT COUNT(*) AS TblExists FROM INFORMATION_SCHEMA.TABLES ` +
+                `WHERE TABLE_SCHEMA = '${mj_core_schema()}' AND TABLE_NAME = 'MaterializedResult'`;
             const result = await this.runQuery(pool, sql);
             const row = (result.recordset?.[0] ?? {});
-            const cnt = row.ColExists ?? row.colexists ?? 0;
-            this._entityHasExternalDataSourceColumn = Number(cnt) > 0;
+            const cnt = row.TblExists ?? row.tblexists ?? 0;
+            this._materializedResultTableExists = Number(cnt) > 0;
         }
-        return this._entityHasExternalDataSourceColumn;
+        return this._materializedResultTableExists;
     }
     async manageExternalEntities(pool, currentUser) {
         let bSuccess = true;
@@ -2070,9 +2876,12 @@ export class ManageMetadataBase {
             }
             // Check the DATABASE for existing field record — in-memory metadata may be stale
             // (e.g. createNewEntityFieldsFromSchema may have already added this field from the view)
-            const existsResult = await this.runQueryWithParams(pool, `SELECT ID, IsVirtual, Type, Length, Precision, Scale, AllowsNull, AllowUpdateAPI
+            // Identifiers are quoted via qi(): __mj.EntityField's columns are mixed-case, and an
+            // unquoted reference folds to lower case on PostgreSQL — `column "length" does not exist`.
+            // SQL Server's case-insensitive resolution hid this for as long as IS-A ran only there.
+            const existsResult = await this.runQueryWithParams(pool, `SELECT ${this.qi('ID')}, ${this.qi('IsVirtual')}, ${this.qi('Type')}, ${this.qi('Length')}, ${this.qi('Precision')}, ${this.qi('Scale')}, ${this.qi('AllowsNull')}, ${this.qi('AllowUpdateAPI')}
                     FROM ${this.qs(mj_core_schema(), 'EntityField')}
-                    WHERE EntityID = @EntityID AND Name = @FieldName`, { 'EntityID': childEntity.ID, 'FieldName': parentField.Name });
+                    WHERE ${this.qi('EntityID')} = @EntityID AND ${this.qi('Name')} = @FieldName`, { 'EntityID': childEntity.ID, 'FieldName': parentField.Name });
             if (existsResult.recordset.length > 0) {
                 // Field already exists — update it to ensure it's marked as a virtual IS-A field
                 const existingRow = existsResult.recordset[0];
@@ -2085,14 +2894,14 @@ export class ManageMetadataBase {
                     !existingRow.AllowUpdateAPI;
                 if (needsUpdate) {
                     const sqlUpdate = `UPDATE ${this.qs(mj_core_schema(), 'EntityField')}
-                  SET IsVirtual=${this.boolLit(true)},
-                      Type='${parentField.Type}',
-                      Length=${parentField.Length},
-                      Precision=${parentField.Precision},
-                      Scale=${parentField.Scale},
-                      AllowsNull=${this.boolLit(parentField.AllowsNull)},
-                      AllowUpdateAPI=${this.boolLit(true)}
-                  WHERE ID='${existingRow.ID}'`;
+                  SET ${this.qi('IsVirtual')}=${this.boolLit(true)},
+                      ${this.qi('Type')}='${parentField.Type}',
+                      ${this.qi('Length')}=${parentField.Length},
+                      ${this.qi('Precision')}=${parentField.Precision},
+                      ${this.qi('Scale')}=${parentField.Scale},
+                      ${this.qi('AllowsNull')}=${this.boolLit(parentField.AllowsNull)},
+                      ${this.qi('AllowUpdateAPI')}=${this.boolLit(true)}
+                  WHERE ${this.qi('ID')}='${existingRow.ID}'`;
                     await this.LogSQLAndExecute(pool, sqlUpdate, `Update IS-A parent field ${parentField.Name} on ${childEntity.Name}`);
                     bUpdated = true;
                 }
@@ -3919,7 +4728,27 @@ export class ManageMetadataBase {
     }
     async createNewEntities(pool, currentUser) {
         try {
-            const sSQL = `SELECT * FROM ${this.qs(mj_core_schema(), 'vwSQLTablesAndEntities')} WHERE ${this.qi('EntityID')} IS NULL ` + this.createExcludeTablesAndSchemasFilter('');
+            // Exclude the PHYSICAL materialization tables (materialized_<Name>) from new-entity auto-creation.
+            // They are internal snapshot storage, surfaced ONLY through the read-only virtual entity minted over
+            // their wrapper view. On any codegen re-run after a materialization exists, this scan would otherwise
+            // find the physical table (which has no entity of its own) and auto-mint a spurious CRUD entity over
+            // it — user-visible AND editable, letting someone mutate the snapshot the refresher overwrites. The
+            // wrapper view is already excluded (it carries the minted entity, so EntityID IS NOT NULL). The outer
+            // view is aliased `t` so the correlated MaterializedResult subquery can't resolve columns ambiguously.
+            const coreSchema = mj_core_schema();
+            // Gate the MaterializedResult reference on the table actually existing. Referencing it unconditionally
+            // would throw — breaking new-entity creation for EVERY entity — on any DB where MaterializedResult
+            // isn't present yet (e.g. the PostgreSQL parallel world's object-availability lag). If it's absent we
+            // simply skip the exclusion (the physical materialized table can't exist without it either, so nothing
+            // to exclude). Uses the shared cached helper so all three materialization-schema gates stay consistent.
+            const matResultExists = await this.materializedResultTableExists(pool);
+            const matExclusion = matResultExists
+                ? `AND NOT EXISTS (SELECT 1 FROM ${this.qs(coreSchema, 'MaterializedResult')} mr WHERE mr.${this.qi('SchemaName')} = t.${this.qi('SchemaName')} AND mr.${this.qi('TableName')} = t.${this.qi('TableName')}) `
+                : '';
+            const sSQL = `SELECT t.* FROM ${this.qs(coreSchema, 'vwSQLTablesAndEntities')} t `
+                + `WHERE t.${this.qi('EntityID')} IS NULL `
+                + matExclusion
+                + this.createExcludeTablesAndSchemasFilter('t.');
             const newEntitiesResult = await this.runQuery(pool, sSQL);
             const newEntities = newEntitiesResult.recordset;
             if (newEntities && newEntities.length > 0) {
@@ -4416,6 +5245,308 @@ export class ManageMetadataBase {
             }
         }
     }
+    /**
+     * RLS-safety assessment for a QUERY materialization, shared by the mint-time gate and the per-run drift
+     * re-check (plan §6.2 / §10). A materialized query entity does NOT inherit its source entities' row-level
+     * security, so it is only safe to serve an unscoped snapshot when we can PROVE no source is RLS-protected.
+     * Fails closed (returns `{ safe:false }`) on three conditions, in order of decreasing severity of what we
+     * cannot prove:
+     *  1. **No provenance** — zero linked source entities (raw SQL / un-analyzed query): we cannot inspect any
+     *     source's RLS, so we cannot prove safety.
+     *  2. **Unresolvable link** — a linked EntityID that doesn't resolve to a known `EntityInfo`: if it secretly
+     *     carried a read RLS filter we'd never see it.
+     *  3. **Row-restricted source** — any resolved source protected by role RLS *or* an API-key row filter (see
+     *     {@link entityHasRowLevelRestriction}; the API-key layer matters because such an entity carries no
+     *     `ReadRLSFilterID`, and the minted entity's NEW EntityID would not match the key's EntityID binding).
+     * Over-restriction here is harmless (the query stays live-only, or an existing materialization is held); the
+     * reverse — serving a protected source's rows unscoped — is the leak this guard exists to prevent.
+     */
+    assessQuerySourceRLSSafety(md, sourceEntityIds, sql) {
+        if (sourceEntityIds.length === 0) {
+            return { safe: false, reason: 'no source-entity provenance is recorded (its QueryEntity links are empty), so RLS-safety cannot be verified' };
+        }
+        const resolved = sourceEntityIds.map((id) => ({ id, entity: md.EntityByID(id) }));
+        const unresolved = resolved.filter((x) => !x.entity);
+        if (unresolved.length > 0) {
+            return { safe: false, reason: `${unresolved.length} source-entity link(s) [${unresolved.map((u) => u.id).join(', ')}] do not resolve to a known entity, so their RLS status can't be verified` };
+        }
+        const rlsProtected = resolved
+            .map((x) => x.entity)
+            .filter((e) => this.entityHasRowLevelRestriction(e));
+        if (rlsProtected.length > 0) {
+            const names = rlsProtected.map((e) => `"${e.Name}"`).join(', ');
+            return { safe: false, reason: `source entit${rlsProtected.length === 1 ? 'y' : 'ies'} ${names} ${rlsProtected.length === 1 ? 'is' : 'are'} read-RLS-protected (role RLS and/or an API-key row filter), and a materialized query entity does NOT inherit source row restrictions (plan §6.2)` };
+        }
+        // P1 hardening — catch RLS sources the LINKER MISSED. The checks above inspect only the LINKED sources
+        // (vwQueryEntities); a source reached via a wrapping view / CTE / function / aliased columns that
+        // query-analysis under-linked would be invisible to them. Parse the query SQL directly, map each table ref
+        // to an entity, and fail closed if a referenced entity is read-RLS-protected but NOT in the linked set —
+        // the precise leak (an unverified RLS source). An unlinked NON-RLS table is not a leak, so it doesn't trip
+        // this (avoids over-refusing legitimate materializations). Parse failure is ignored (the linked-only checks
+        // above still stand); only a POSITIVE detection of an unlinked RLS source refuses.
+        if (sql && sql.trim().length > 0) {
+            const linkedSet = new Set(sourceEntityIds.map((id) => id.trim().toLowerCase()));
+            let refs = [];
+            try {
+                refs = SQLParser.ExtractTableRefs(sql, this.dialect) ?? [];
+            }
+            catch {
+                refs = [];
+            }
+            // CTE self-references are emitted by ExtractTableRefs as unqualified (dbo-defaulted) refs, but they are
+            // NOT base-table sources. Since findEntityByBaseObject now falls back to a schema-agnostic name match for
+            // unqualified refs (to catch under-linked __mj/app-schema sources), a CTE that happens to share an
+            // entity's base-table name would otherwise be misread as an under-linked source and over-refuse. Collect
+            // the CTE names and skip any unqualified ref that matches one. (Parse failure → no exclusion; the guard
+            // still fails closed, only slightly more conservatively.)
+            const cteNames = new Set();
+            try {
+                // Extract each CTE's name from its "name AS (...)" definition. Handle all three quotings SQLParser can
+                // emit ([bracket], "double-quote", bare) plus an optional column list — mirroring the parser's own CTE
+                // header regex. A missed name only means a (safe) over-refusal, but bracket names are the common T-SQL
+                // form, so cover them.
+                for (const def of SQLParser.ExtractCTEs(sql, this.dialect)?.CTEDefinitions ?? []) {
+                    const m = /^\s*(?:\[([^\]]+)\]|"([^"]+)"|([A-Za-z_][\w$]*))\s*(?:\([^)]*\))?\s+AS\b/i.exec(def);
+                    const name = m?.[1] ?? m?.[2] ?? m?.[3];
+                    if (name)
+                        cteNames.add(name.trim().toLowerCase());
+                }
+            }
+            catch { /* no CTE exclusion available — guard stays fail-closed */ }
+            for (const ref of refs) {
+                if (!ref.TableName)
+                    continue;
+                const refSchema = (ref.SchemaName ?? '').trim().toLowerCase();
+                // Skip CTE self-references (only unqualified / dbo-defaulted refs can be a CTE name).
+                if ((refSchema === '' || refSchema === 'dbo') && cteNames.has(ref.TableName.trim().toLowerCase()))
+                    continue;
+                // Refuse ANY under-linked entity source, not only RLS-protected ones. Both the RLS-safety check above
+                // AND the read-grant role intersection (addMaterializedQueryEntityPermissions) are computed over the
+                // LINKED set only, so a source query-analysis did NOT link is invisible to both: an RLS source would
+                // leak its rows unscoped, and a merely CanRead-restricted (non-RLS) source would let the minted
+                // entity's read grant exceed "can read every source" — a privilege escalation. A parsed ref maps to an
+                // entity when it IS that entity's base view/table; derived-table aliases map to nothing, and CTE
+                // self-references are excluded above, so a mapped-but-unlinked source is genuine under-linking and
+                // refusing is correct; over-refusal is harmless (the query stays live-only — link the source via full
+                // query analysis, or use a base-view materialization which inherits source read access).
+                // Consider ALL candidate entities the ref can map to and fail closed if ANY is unlinked: an unqualified
+                // ref is schema-ambiguous (the parser collapses it to 'dbo'), so a same-base-name entity in another
+                // schema could be the true referent — picking only the first match could silently pass an unlinked one.
+                const candidates = this.findEntitiesByBaseObject(md, ref.SchemaName, ref.TableName);
+                const unlinked = candidates.find((e) => !linkedSet.has(e.ID.trim().toLowerCase()));
+                if (unlinked) {
+                    const rlsNote = this.entityHasRowLevelRestriction(unlinked) ? ' row-restricted (role RLS and/or an API-key row filter)' : ' read-restricted';
+                    return { safe: false, reason: `query references${rlsNote} source "${unlinked.Name}" (${ref.SchemaName ?? ''}.${ref.TableName}) that query-analysis did NOT link — an unscoped snapshot cannot prove its per-source read access is honored (P1 under-linking guard)` };
+                }
+            }
+        }
+        return { safe: true };
+    }
+    /** Maps a physical (schema, table/view) reference to the MJ entities backed by it — by BaseView or BaseTable,
+     *  case/whitespace-insensitive. Returns ALL candidates so the P1 guard can fail closed on an AMBIGUOUS ref.
+     *  Resolution: an explicit, non-default schema yields only that schema's exact matches; an empty schema OR the
+     *  parser-default `'dbo'` — which {@link SQLParser.ExtractTableRefs} assigns to every UNQUALIFIED ref, while MJ
+     *  core/app entities live in `__mj`/app schemas (never `dbo`) — is treated as unqualified and yields every
+     *  schema-agnostic name match. That way an unqualified reference to a `__mj`/app-schema entity is still caught,
+     *  and a same-base-name collision across schemas surfaces every candidate rather than just the first (so the
+     *  guard refuses if any of them is under-linked). Used by the P1 under-linking guard. */
+    findEntitiesByBaseObject(md, schema, table) {
+        const t = table.trim().toLowerCase();
+        const s = (schema ?? '').trim().toLowerCase();
+        const nameMatch = (e) => {
+            const bv = (e.BaseView ?? '').trim().toLowerCase();
+            const bt = (e.BaseTable ?? '').trim().toLowerCase();
+            return bv === t || bt === t;
+        };
+        if (s && s !== 'dbo') {
+            // Explicit, non-default schema → only that schema's entity qualifies (no cross-schema ambiguity).
+            return md.Entities.filter((e) => (e.SchemaName ?? '').trim().toLowerCase() === s && nameMatch(e));
+        }
+        // Empty schema / parser-default 'dbo' (i.e. UNQUALIFIED) → every name match across schemas is a candidate.
+        return md.Entities.filter(nameMatch);
+    }
+    /**
+     * Grants read permissions to a newly-minted materialized QUERY entity scoped to the INTERSECTION of source
+     * read access (C2). A precomputed snapshot has no per-row scoping, so a role may read it only if it can read
+     * EVERY source entity — otherwise a user who cannot read a source could read its rows through the snapshot.
+     * Starts from the same config-default role list as {@link addDefaultPermissionsForEntity} (never grants a
+     * role the defaults wouldn't) and drops any role that lacks explicit read on every source. Always read-only
+     * (the entity is a virtual entity with no CRUD sprocs). If the intersection is empty the entity receives no
+     * role-based read grants — the fail-closed direction (§10); a human can grant read after review.
+     */
+    async addMaterializedQueryEntityPermissions(pool, entityId, entityName, sourceEntities) {
+        if (!configInfo.newEntityDefaults.PermissionDefaults?.AutoAddPermissionsForNewEntities) {
+            return;
+        }
+        // Fail-closed (C2): with no resolved source entities we cannot prove any role can read EVERY source, so
+        // grant nothing. `sourceEntities.every(...)` is vacuously true on an empty array and would otherwise grant
+        // read to every default role — the exact "read the snapshot without read on a source" leak this method
+        // exists to prevent. Callers already refuse on empty provenance, but guard here too so the security
+        // invariant holds in isolation, not only by a distant caller's gate.
+        if (sourceEntities.length === 0) {
+            logStatus(`    > Materialized entity "${entityName}": no resolved source entities — granting no role-based read (fail-closed).`);
+            return;
+        }
+        const md = new Metadata(); // global-provider-ok: codegen runs offline against a single provider
+        const roleCanReadAllSources = (roleId) => sourceEntities.every((e) => e.Permissions.some((p) => UUIDsEqual(p.RoleID, roleId) && p.CanRead));
+        for (const p of configInfo.newEntityDefaults.PermissionDefaults.Permissions) {
+            const roleId = md.Roles.find((r) => r.Name.trim().toLowerCase() === p.RoleName.trim().toLowerCase())?.ID;
+            if (!roleId) {
+                LogError(`   >>>> ERROR: Unable to find Role ID for role ${p.RoleName} to add permissions for materialized entity ${entityName}`);
+                continue;
+            }
+            if (!p.CanRead || !roleCanReadAllSources(roleId)) {
+                logStatus(`    > Materialized entity "${entityName}": role "${p.RoleName}" NOT granted read (it cannot read every source entity; the snapshot has no row-level scoping).`);
+                continue;
+            }
+            const sSQLInsert = `INSERT INTO ${this.qs(mj_core_schema(), 'EntityPermission')}
+                              (${this.qi('EntityID')}, ${this.qi('RoleID')}, ${this.qi('CanRead')}, ${this.qi('CanCreate')}, ${this.qi('CanUpdate')}, ${this.qi('CanDelete')}, ${this.qi('__mj_CreatedAt')}, ${this.qi('__mj_UpdatedAt')}) VALUES
+                              ('${entityId}', '${roleId}', ${this.boolLit(true)}, ${this.boolLit(false)}, ${this.boolLit(false)}, ${this.boolLit(false)}, ${this.utcNow()}, ${this.utcNow()})`;
+            await this.LogSQLAndExecute(pool, sSQLInsert, `SQL generated to add read permission for materialized entity ${entityName} for role ${p.RoleName}`);
+        }
+    }
+    /**
+     * Revokes read access on a minted materialized entity by setting `CanRead=0` on all its EntityPermission
+     * rows (the __mj_UpdatedAt trigger stamps the timestamp). Used by the drift re-check when a query source
+     * gains RLS after minting: the snapshot can no longer be safely served, so it is made unreadable until a
+     * human authors protection. Non-destructive (rows are kept, just flipped) so the grant can be restored.
+     */
+    async revokeMaterializedEntityReadAccess(pool, entityId, entityLabel, reason) {
+        const sql = `UPDATE ${this.qs(mj_core_schema(), 'EntityPermission')} SET ${this.qi('CanRead')}=${this.boolLit(false)} WHERE ${this.qi('EntityID')}='${entityId}' AND ${this.qi('CanRead')}=${this.boolLit(true)}`;
+        await this.LogSQLAndExecute(pool, sql, `Revoke read access on materialized entity "${entityLabel}" (${reason})`);
+    }
+    /**
+     * ROLE-RLS LAYER ONLY — true if any of the entity's role permissions carries a non-empty `ReadRLSFilterID`.
+     *
+     * This is deliberately NOT a gate on its own, and no materialization gate may call it directly. MJ enforces
+     * row restrictions in TWO layers (see `EntityInfo.GetEffectiveRowFilterWhereClause`, the sanctioned composer):
+     * role RLS *and* API-key row filters, AND-composed. An entity bound ONLY by an API-key row filter carries no
+     * `ReadRLSFilterID` at all, so a gate that stopped at this layer would judge it unprotected and fail OPEN —
+     * minting a materialized entity with a NEW EntityID, which the API-key binding (which binds by EntityID) no
+     * longer matches, handing a filtered principal a full unscoped snapshot of rows it cannot read live.
+     *
+     * `EntityInfo` exposes no user-free accessor for either layer (`GetEffectiveRowFilterWhereClause` needs a
+     * session `UserInfo`, and CodeGen has no session), so the composition is done explicitly here — by
+     * {@link entityHasRowLevelRestriction}, which is what every gate calls.
+     */
+    entityHasRowLevelSecurity(entity) {
+        return ManageMetadataBase.EntityHasRoleReadRLS(entity);
+    }
+    /** Pure form of the role-RLS layer (see {@link entityHasRowLevelSecurity}). IO-free so it can be reused
+     *  verbatim by the runtime refresher's equivalent gate. */
+    static EntityHasRoleReadRLS(entity) {
+        return entity.Permissions.some((p) => !!p.ReadRLSFilterID && p.ReadRLSFilterID.trim().length > 0);
+    }
+    /**
+     * THE row-restriction gate for materialization. True when the entity is protected by ANY row-level
+     * restriction — role RLS **or** an API-key / API-application row filter — because a materialized snapshot
+     * reproduces neither. Fails closed when the API-key layer could not be enumerated (see
+     * {@link apiKeyRowFilterTargets}).
+     */
+    entityHasRowLevelRestriction(entity) {
+        return ManageMetadataBase.EntityHasRowLevelRestriction(entity, this.apiKeyRowFilterTargets);
+    }
+    /**
+     * Pure, IO-free core of {@link entityHasRowLevelRestriction} — the predicate any other layer (e.g. the
+     * runtime `MaterializationRefresher`) should mirror, supplying the API-key target set however it can obtain
+     * it. `'unknown'` for the target set ⇒ true (fail closed).
+     */
+    static EntityHasRowLevelRestriction(entity, apiKeyRowFilterTargets) {
+        if (ManageMetadataBase.EntityHasRoleReadRLS(entity))
+            return true;
+        if (apiKeyRowFilterTargets === 'unknown')
+            return true; // cannot prove the key layer is empty → fail closed
+        return apiKeyRowFilterTargets.has((entity.Name ?? '').trim().toLowerCase());
+    }
+    /**
+     * Loads {@link apiKeyRowFilterTargets} once per CodeGen run. Called at the top of every materialization entry
+     * point, BEFORE any gate runs, so the gates never evaluate against an unloaded cache in production.
+     *
+     * Resolution rules (all biased fail-closed):
+     *  - The `RowFilterID` column absent on a scope table ⇒ that binding layer cannot exist on this database
+     *    (pre-v6 schema / the PostgreSQL parallel world) ⇒ contributes nothing. This is CORRECT, not fail-open.
+     *  - A filtered rule's `ResourcePattern` must name ONE exact entity (enforced at rule save). Mappability is
+     *    decided by {@link ResolveSingleEntityResourceTarget}, shared verbatim with the runtime refresher's
+     *    identical gate; a rule it cannot resolve collapses the whole set to `'unknown'` — every entity is then
+     *    treated as restricted.
+     *  - Any error enumerating the rules ⇒ `'unknown'` (never a silently-empty set).
+     * Permission type is deliberately NOT narrowed to `Read`: mapping a scope rule to a permission type requires
+     * the APIScope path taxonomy that lives outside CodeGen, and over-restriction here is harmless.
+     */
+    async loadAPIKeyRowFilterTargets(pool) {
+        if (this.apiKeyRowFilterTargets !== 'unknown')
+            return; // already loaded this run
+        try {
+            const targets = new Set();
+            for (const table of ['APIKeyScope', 'APIApplicationScope']) {
+                if (!(await this.columnExistsInCoreSchema(pool, table, 'RowFilterID')))
+                    continue;
+                const res = await this.runQuery(pool, `SELECT DISTINCT ${this.qi('ResourcePattern')} AS ResourcePattern FROM ${this.qs(mj_core_schema(), table)} WHERE ${this.qi('RowFilterID')} IS NOT NULL`);
+                for (const row of res.recordset ?? []) {
+                    const r = row;
+                    const pattern = (r.ResourcePattern ?? r.resourcepattern) ?? '';
+                    // The mappability rule is SHARED with the runtime refresher's identical gate
+                    // (ResolveSingleEntityResourceTarget in @memberjunction/global) rather than copied. The two
+                    // gates must agree exactly: a copy that drifts open here silently re-opens the leak the
+                    // runtime gate closes, and vice versa — with nothing in the build to notice.
+                    const target = ResolveSingleEntityResourceTarget(pattern);
+                    if (target === null) {
+                        logError(`    > API-key row-filter enumeration: a filtered scope rule in ${table} has an unmappable ResourcePattern ` +
+                            `("${pattern.trim()}") — it cannot be resolved to a single entity, so EVERY entity is treated as row-restricted ` +
+                            `for materialization this run (fail closed). Fix the rule to name one exact entity.`);
+                        this.apiKeyRowFilterTargets = 'unknown';
+                        return;
+                    }
+                    targets.add(target);
+                }
+            }
+            this.apiKeyRowFilterTargets = targets;
+        }
+        catch (err) {
+            this.apiKeyRowFilterTargets = 'unknown';
+            logError(`    > API-key row-filter enumeration FAILED — every entity will be treated as row-restricted for materialization ` +
+                `this run (fail closed): ${err instanceof Error ? err.message : String(err)}`);
+        }
+    }
+    async columnExistsInCoreSchema(pool, table, column) {
+        const key = `${table}.${column}`.toLowerCase();
+        const cached = this._coreSchemaColumnExists.get(key);
+        if (cached !== undefined)
+            return cached;
+        const sql = `SELECT COUNT(*) AS ColExists FROM INFORMATION_SCHEMA.COLUMNS ` +
+            `WHERE TABLE_SCHEMA = '${mj_core_schema()}' AND TABLE_NAME = '${table}' AND COLUMN_NAME = '${column}'`;
+        const result = await this.runQuery(pool, sql);
+        const row = (result.recordset?.[0] ?? {});
+        const cnt = row.ColExists ?? row.colexists ?? 0;
+        const exists = Number(cnt) > 0;
+        this._coreSchemaColumnExists.set(key, exists);
+        return exists;
+    }
+    /**
+     * Leak-2 (C2 ongoing): re-narrows a minted materialized QUERY entity's read grants to the CURRENT source-read
+     * intersection. The mint-time grant ({@link addMaterializedQueryEntityPermissions}) is computed once; if a role
+     * later LOSES plain `CanRead` on a source, its grant on the snapshot must be revoked too — otherwise it keeps
+     * reading snapshot rows it can no longer read live. Called from the drift pass on every codegen run. Only ever
+     * REVOKES (the safe direction); re-granting a role that regained access is a human decision after review.
+     * Returns the number of grants revoked.
+     */
+    async reconcileMaterializedQueryEntityReadGrants(pool, entityId, entityLabel, sourceEntities) {
+        if (sourceEntities.length === 0)
+            return 0; // no sources to intersect against — the RLS/empty gate handles this
+        const roleCanReadAllSources = (roleId) => sourceEntities.every((e) => e.Permissions.some((p) => UUIDsEqual(p.RoleID, roleId) && p.CanRead));
+        const grants = await this.runQuery(pool, `SELECT ${this.qi('RoleID')} FROM ${this.qs(mj_core_schema(), 'EntityPermission')} WHERE ${this.qi('EntityID')}='${entityId}' AND ${this.qi('CanRead')}=${this.boolLit(true)}`);
+        let revoked = 0;
+        for (const g of grants.recordset) {
+            const roleId = g.RoleID;
+            if (!roleId)
+                continue;
+            if (!roleCanReadAllSources(roleId)) {
+                await this.LogSQLAndExecute(pool, `UPDATE ${this.qs(mj_core_schema(), 'EntityPermission')} SET ${this.qi('CanRead')}=${this.boolLit(false)} WHERE ${this.qi('EntityID')}='${entityId}' AND ${this.qi('RoleID')}='${roleId}' AND ${this.qi('CanRead')}=${this.boolLit(true)}`, `Narrow read grant on materialized entity "${entityLabel}": role ${roleId} can no longer read every source (C2 intersection re-narrowed)`);
+                revoked++;
+            }
+        }
+        return revoked;
+    }
     createNewEntityInsertSQL(newEntityUUID, newEntityName, newEntity, newEntitySuffix, newEntityDisplayName) {
         const newEntityDefaults = configInfo.newEntityDefaults;
         const newEntityDescriptionEscaped = newEntity.EntityDescription ? `'${newEntity.EntityDescription.replace(/'/g, "''")}'` : null;