@memberjunction/server 6.1.0-edge.2 → 6.1.0-edge.4

package/dist/integration/CustomColumnPromoter.js

@@ -18,8 +18,10 @@
  * goes through DDLGenerator driven by provider.PlatformKey ('sqlserver' | 'postgresql'); RSU has
  * dual SS/PG setup; IOF + field map are BaseEntity. No SS-only SQL anywhere in this file.
  *
- * The MJAPI restart (so the new column is exposed over GraphQL) + the JSON spread of staged
- * values into the new columns are M3 — this stage applies the schema only (SkipRestart).
+ * The ADD COLUMN runs through RSU's normal pipeline — migration written, committed, MJAPI
+ * restarted once for the whole batch — because the restart is what exposes the new columns over
+ * GraphQL and the commit is what stops the database carrying columns the repository has no record
+ * of. The JSON spread of staged values into those columns happens here too.
  */
 import { LogError, LogStatus, LogStatusEx, Metadata, RunView, } from '@memberjunction/core';
 import { IntegrationEngine, CUSTOM_OVERFLOW_COLUMN, CONTENT_HASH_COLUMN, computeContentHash, buildOverflowStats, planPromotions, sanitizeColumnName, inferColumnTypeFromStats, } from '@memberjunction/integration-engine';
@@ -33,6 +35,23 @@ const OVERFLOW_SAMPLE_SIZE = 1000;
  * remainder is logged + deferred; it promotes on subsequent syncs (the keys stay captured).
  */
 const MAX_PROMOTIONS_PER_PASS = 25;
+/** Page size for the full-table overflow walks (spread + stale purge). */
+const OVERFLOW_PAGE_SIZE = 500;
+/**
+ * Rows the stale purge may WRITE in one pass, the row-wise sibling of {@link MAX_PROMOTIONS_PER_PASS}.
+ *
+ * Each purged row costs one `BaseEntity.Save()` — roughly nine serialized round trips, which is the
+ * only write shape MJ offers today (there is no batched-update provider capability; `TransactionGroup`
+ * gives atomicity, not batching, since both dialect implementations loop one query per item). That
+ * puts the purge around 250 rows/min, so an unbounded sweep of a large table would hold the post-sync
+ * promotion callback open for hours.
+ *
+ * Residue is inert while it waits — the field-map-first terminate check already stops a mapped key
+ * being re-offered as a new column — so draining it across several syncs costs nothing but time.
+ * The budget bounds WRITES, not the scan: the walk keeps reading (page reads are cheap and rows
+ * already purged fall through in memory) so a later pass still reaches residue further down the table.
+ */
+const MAX_PURGE_ROWS_PER_PASS = 1000;
 /** Registers the post-sync custom-column promotion hook on the IntegrationEngine singleton. */
 export function registerIntegrationCustomColumnPromoter() {
     IntegrationEngine.Instance.SetPostSyncSchemaPromotionCallback(async (ctx) => {
@@ -96,69 +115,227 @@ export class IntegrationCustomColumnPromoter {
         const integrationID = await this.resolveIntegrationID(companyIntegrationID);
         if (!integrationID)
             return NOT_PROMOTED;
-        const columnsAdded = [];
+        const plans = [];
+        const batchInputs = [];
         for (const entityName of syncedEntityNames) {
             try {
-                const added = await this.promoteEntity(companyIntegrationID, integrationID, entityName, customKeyStats?.[entityName]);
-                columnsAdded.push(...added);
+                const planned = await this.planWorkForEntity(companyIntegrationID, entityName, customKeyStats?.[entityName]);
+                if (!planned)
+                    continue; // no overflow column / no captured customs / no entity map
+                const { entityInfo, entityMap } = planned;
+                // PURGE FIRST — strip already-promoted keys from the staging JSON BEFORE any column is
+                // created, and before the RSU pass that would restart this process. A failed or skipped
+                // sync is not evidence that a column is missing, and leaving the residue in place is what
+                // makes an already-promoted key re-surface as a phantom "new column". Runs even when there
+                // is nothing new to promote, which is the only way pre-existing residue on rows the sync
+                // never rewrites ever gets cleaned rather than endlessly re-detected.
+                if (planned.stale.length > 0) {
+                    await this.purgeStaleOverflowKeys(entityName, entityMap.ID, entityInfo, planned.stale);
+                }
+                if (planned.work.length === 0)
+                    continue; // already converged
+                let work = planned.work;
+                // M4a: bound schema churn per pass — the remainder stays captured and promotes next sync.
+                if (work.length > MAX_PROMOTIONS_PER_PASS) {
+                    LogStatus(`[CustomColumnPromoter] ${work.length} candidates on ${entityName}; promoting ${MAX_PROMOTIONS_PER_PASS} this pass, ${work.length - MAX_PROMOTIONS_PER_PASS} deferred to next sync.`);
+                    work = work.slice(0, MAX_PROMOTIONS_PER_PASS);
+                }
+                const newColumns = work.filter(w => w.needsColumn);
+                const plan = { entityName, entityInfo, entityMap, work, batchIndex: -1 };
+                if (newColumns.length > 0) {
+                    plan.batchIndex = batchInputs.length;
+                    batchInputs.push(this.buildSchemaInput(entityInfo, newColumns));
+                }
+                plans.push(plan);
             }
             catch (err) {
-                // One entity's promotion failure must not abort the others, and never the sync.
-                LogError(`[CustomColumnPromoter] Promotion failed for entity '${entityName}': ${this.msg(err)}`);
+                // One entity's planning failure must not abort the others, and never the sync.
+                LogError(`[CustomColumnPromoter] Planning failed for entity '${entityName}': ${this.msg(err)}`);
             }
         }
-        const promoted = columnsAdded.length > 0;
-        if (promoted) {
-            // Make the freshly-created EntityFields visible in-process for the next sync's mapping.
+        if (plans.length === 0)
+            return NOT_PROMOTED;
+        // Register the follow-up DURABLY before the restart, the way the apply-objects path does
+        // (IntegrationDiscoveryResolver: `rsuInput.PendingWork = [pendingPayload]`). The restart is
+        // what loads the regenerated entity classes, so everything downstream of the DDL — the IOF
+        // rows, the field maps, the overflow spread — belongs after it, not before.
+        //
+        // Carried on the first input because RSU restarts ONCE for the whole batch; the payload
+        // describes every entity in the pass.
+        if (batchInputs.length > 0) {
+            batchInputs[0].PendingWork = [{
+                    WorkType: 'promote-columns',
+                    CompanyIntegrationID: companyIntegrationID,
+                    // Not used by the promote path (which works from PromotedColumns), but the contract
+                    // requires them and they keep the row legible to an operator reading the table.
+                    SourceObjectNames: plans.map(p => p.entityMap.ExternalObjectName),
+                    SchemaName: plans[0].entityInfo.SchemaName,
+                    CreatedAt: new Date().toISOString(),
+                    PromotedColumns: plans.map(p => ({
+                        EntityName: p.entityName,
+                        EntityMapID: p.entityMap.ID,
+                        ExternalObjectName: p.entityMap.ExternalObjectName,
+                        IntegrationID: integrationID,
+                        Columns: p.work.map(w => ({
+                            SourceKey: w.sourceKey,
+                            ColumnName: w.columnName,
+                            SchemaFieldType: w.candidate.Inferred.SchemaFieldType,
+                            MaxLength: w.candidate.Inferred.MaxLength,
+                            Coverage: w.candidate.Coverage,
+                        })),
+                    })),
+                }];
+            batchInputs[0].ContextUser = this.user;
+        }
+        // ── PHASE 2: ONE batched RSU pass for ALL entities' ADD COLUMN migrations. ──
+        // RunPipelineBatch runs every migration under one lock, then ONE CodeGen + compile +
+        // restart + git commit — which is now genuinely one restart and one commit for the whole
+        // promotion, rather than one per entity. See buildSchemaInput: the inputs no longer set
+        // SkipRestart/SkipGitCommit, so the batch commits the migration and restarts once at the
+        // end, which is what exposes the new columns over GraphQL and what keeps the repository in
+        // step with the database.
+        let batchResults = [];
+        if (batchInputs.length > 0) {
+            const batch = await RuntimeSchemaManager.Instance.RunPipelineBatch(batchInputs);
+            batchResults = batch.Results ?? [];
+            // M3: make the freshly-created EntityFields + regenerated sprocs visible in-process
+            // ONCE for the whole batch. CRITICAL: the spread below builds sproc calls from this
+            // metadata; without the refresh, row.Save() uses the STALE field list that predates
+            // the column add and mismatches the regenerated sproc → "Error executing SQL".
             try {
                 await this.provider.Refresh();
             }
             catch (err) {
-                LogError(`[CustomColumnPromoter] provider.Refresh failed: ${this.msg(err)}`);
+                LogError(`[CustomColumnPromoter] post-batch Refresh failed: ${this.msg(err)}`);
+            }
+        }
+        // ── PHASE 3: the NO-DDL case only — IOF rows + field maps + value spread, inline. ──
+        //
+        // An entity that needed a column went through the batch, and the batch RESTARTS: `pm2
+        // restart` kills this process, so nothing below runs for it. Its IOF rows, field maps and
+        // spread are completed by the post-restart consumer from the PendingWork registered above
+        // (CompletePromotion) — which is also where they belong, because only after the restart are
+        // the regenerated entity classes loaded.
+        //
+        // What is left here is the recovery-only pass: work items that need no column, so no
+        // migration, so no batch entry and no restart. Those are finished inline exactly as before.
+        // Explicitly filtered rather than left to fall through, so this reads as the branch it is
+        // instead of as dead code that happens to be unreachable.
+        const columnsAdded = [];
+        for (const plan of plans) {
+            if (plan.batchIndex >= 0) {
+                const res = batchResults[plan.batchIndex];
+                if (!res || !res.Success) {
+                    // DDL failed — leave everything captured; retry next promote (no partial commit
+                    // lost). The PendingWork row stays Pending and is re-processable.
+                    LogError(`[CustomColumnPromoter] RSU ADD COLUMN failed on ${plan.entityInfo.Name}: ${res?.ErrorMessage ?? res?.ErrorStep ?? 'unknown'} — leaving ${plan.entityName} captured for retry.`);
+                }
+                else {
+                    // Succeeded. The restart has either already ended this process or is about to,
+                    // and the consumer owns the follow-up — so do NOT do the metadata work here.
+                    // Still COUNT the columns: they were promoted, and SchemaUpdatePending is
+                    // derived from this list. The client keys its "workspace updating" state off
+                    // that flag, so dropping these would tell it nothing happened.
+                    const added = plan.work.filter(w => !w.recoverSpread);
+                    LogStatus(`[CustomColumnPromoter] Promoted ${added.length} column(s) on ${plan.entityName}: ${added.map(w => w.columnName).join(', ')} — metadata + spread deferred to the post-restart consumer.`);
+                    columnsAdded.push(...added.map(w => ({ EntityName: plan.entityName, ColumnName: w.columnName })));
+                }
+                continue;
+            }
+            try {
+                // Completion goes through the SAME method the post-restart consumer calls, given the
+                // same payload shape. There is one implementation of "finish a promotion", so the
+                // no-restart case can never drift from the restart case — which is the drift that
+                // would be impossible to notice, since only one of them runs on any given pass.
+                await this.CompletePromotion([{
+                        EntityName: plan.entityName,
+                        EntityMapID: plan.entityMap.ID,
+                        ExternalObjectName: plan.entityMap.ExternalObjectName,
+                        IntegrationID: integrationID,
+                        Columns: plan.work.map(w => ({
+                            SourceKey: w.sourceKey,
+                            ColumnName: w.columnName,
+                            SchemaFieldType: w.candidate.Inferred.SchemaFieldType,
+                            MaxLength: w.candidate.Inferred.MaxLength,
+                            Coverage: w.candidate.Coverage,
+                        })),
+                    }], new Map([[plan.entityName, plan.entityInfo]]));
+                // recoverSpread items add no column/metadata — they only finish an interrupted
+                // backfill, so they don't count as "columns added" (keeps SchemaUpdatePending honest).
+                const added = plan.work.filter(w => !w.recoverSpread);
+                columnsAdded.push(...added.map(w => ({ EntityName: plan.entityName, ColumnName: w.columnName })));
+            }
+            catch (err) {
+                // One entity's promotion failure must not abort the others, and never the sync.
+                LogError(`[CustomColumnPromoter] Promotion failed for entity '${plan.entityName}': ${this.msg(err)}`);
             }
         }
+        const promoted = columnsAdded.length > 0;
         return { Promoted: promoted, ColumnsAdded: columnsAdded, SchemaUpdatePending: promoted };
     }
-    /** Gate → plan → promote for a single target entity. Returns the columns it added. */
-    async promoteEntity(companyIntegrationID, integrationID, entityName, inRunStats) {
-        const planned = await this.planWorkForEntity(companyIntegrationID, entityName, inRunStats);
-        if (!planned || planned.work.length === 0)
-            return []; // no overflow / no entity map / already converged
-        const { entityInfo, entityMap } = planned;
-        let work = planned.work;
-        // M4a: bound schema churn per pass — the remainder stays captured and promotes next sync.
-        if (work.length > MAX_PROMOTIONS_PER_PASS) {
-            LogStatus(`[CustomColumnPromoter] ${work.length} candidates on ${entityName}; promoting ${MAX_PROMOTIONS_PER_PASS} this pass, ${work.length - MAX_PROMOTIONS_PER_PASS} deferred to next sync.`);
-            work = work.slice(0, MAX_PROMOTIONS_PER_PASS);
-        }
-        // 4. PROMOTE: ADD COLUMN+CodeGen (only the keys that need a column) → IOF rows → field maps → spread.
-        const newColumns = work.filter(w => w.needsColumn);
-        if (newColumns.length > 0 && !await this.applySchemaChange(entityInfo, newColumns)) {
-            return []; // DDL failed — leave everything captured; retry next sync (no partial commit lost)
-        }
-        await this.createIntegrationObjectFields(integrationID, entityMap.ExternalObjectName, work);
-        await this.createFieldMaps(entityMap.ID, work.filter(w => w.needsFieldMap));
-        // M3: spread staged values into the real columns + re-baseline the content hash.
-        // CRITICAL: refresh metadata FIRST. applySchemaChange just added the columns AND had RSU
-        // regenerate the spUpdate sproc to include them — but the running provider's in-memory entity
-        // metadata still predates the column add (the Refresh at the top ran before it). Without this,
-        // the spread's row.Save() builds a sproc call from the STALE field list that doesn't match the
-        // regenerated sproc → "Error executing SQL" (the spread-save failures). Re-loading metadata
-        // here aligns the entity's field set with the new DB sproc so the backfill saves succeed.
-        try {
-            await this.provider.Refresh();
-        }
-        catch (err) {
-            LogError(`[CustomColumnPromoter] pre-spread Refresh failed: ${this.msg(err)}`);
+    /**
+     * Finish a promotion AFTER the restart — the IntegrationObjectField rows, the field maps and
+     * the overflow→column spread — from the {@link RSUPendingWork} the pre-restart pass registered.
+     *
+     * Same work PHASE 3 does inline when no restart occurred; it lives here rather than in the
+     * consumer so promotion logic stays in one class. The difference is that here the regenerated
+     * entity classes ARE loaded, so the spread writes through real typed columns instead of the
+     * dynamic .Get/.Set the pre-restart path is forced into.
+     *
+     * Idempotent, because a pending row is marked Completed only after success and so stays
+     * re-processable after a crash: field maps are filtered against those already active, and the
+     * spread only fills a destination that is still empty.
+     */
+    async CompletePromotion(promoted, knownEntities) {
+        const columnsAdded = [];
+        for (const entry of promoted) {
+            try {
+                // Provider first — post-restart it holds the REGENERATED class, which is the whole
+                // reason completion was deferred. knownEntities is the inline caller handing back the
+                // EntityInfo it already resolved, for the no-DDL pass where nothing was regenerated.
+                const entityInfo = this.provider.EntityByName(entry.EntityName) ?? knownEntities?.get(entry.EntityName);
+                if (!entityInfo) {
+                    LogError(`[CustomColumnPromoter] Post-restart: entity '${entry.EntityName}' not found; skipping.`);
+                    continue;
+                }
+                // Destination names are CARRIED, never recomputed: uniqueColumnName may have
+                // suffixed one to dodge a collision, and re-deriving it here could pick a different
+                // name than the column the migration actually created.
+                const existingMaps = await this.activeFieldMaps(entry.EntityMapID);
+                const named = entry.Columns.map(c => ({
+                    candidate: {
+                        Key: c.SourceKey,
+                        Coverage: c.Coverage,
+                        // SqlServerType/PostgresType drive DDL only, which ran before the restart.
+                        // The phase-3 work below reads SchemaFieldType and MaxLength.
+                        Inferred: {
+                            SchemaFieldType: c.SchemaFieldType,
+                            MaxLength: c.MaxLength,
+                            SqlServerType: '',
+                            PostgresType: '',
+                        },
+                    },
+                    sourceKey: c.SourceKey,
+                    columnName: c.ColumnName,
+                    needsColumn: false,
+                    needsFieldMap: !existingMaps.has(c.SourceKey.toLowerCase()),
+                }));
+                await this.createIntegrationObjectFields(entry.IntegrationID, entry.ExternalObjectName, named);
+                await this.createFieldMaps(entry.EntityMapID, named.filter(n => n.needsFieldMap));
+                await this.spreadAndRebaseline(entry.EntityName, entry.EntityMapID, entityInfo, named);
+                LogStatus(`[CustomColumnPromoter] Post-restart: completed ${named.length} promoted column(s) on ${entry.EntityName}: ${named.map(n => n.columnName).join(', ')}`);
+                columnsAdded.push(...named.map(n => ({ EntityName: entry.EntityName, ColumnName: n.columnName })));
+            }
+            catch (err) {
+                // One entity must not abort the others; the pending row stays open for retry.
+                LogError(`[CustomColumnPromoter] Post-restart completion failed for '${entry.EntityName}': ${this.msg(err)}`);
+            }
         }
-        const refreshedEntityInfo = this.provider.EntityByName(entityName) ?? entityInfo;
-        await this.spreadAndRebaseline(entityName, entityMap.ID, refreshedEntityInfo, work);
-        LogStatus(`[CustomColumnPromoter] Promoted/recovered ${work.length} custom column(s) on ${entityName}: ${work.map(w => w.columnName).join(', ')}`);
-        return work.map(w => ({ EntityName: entityName, ColumnName: w.columnName }));
+        return columnsAdded;
     }
     /**
      * Dry-run of GATE → scan → PLAN → resolve-work for ONE entity, WITHOUT applying any schema change.
-     * Shared by {@link promoteEntity} (which then PROMOTES the work) and {@link ListCandidates} (which only
+     * Shared by {@link PromoteForSync} (phase 1, which then batch-PROMOTES the work) and {@link ListCandidates} (which only
      * reports it). Returns null when the entity has no overflow column / no captured customs / no entity map.
      * Because the work list is computed live (overflow keys minus already-column-and-mapped), re-running is
      * inherently deduped — a concurrent discovery that already promoted a key yields no work item for it.
@@ -176,6 +353,15 @@ export class IntegrationCustomColumnPromoter {
         //     out-of-band capture; the hash basis excludes overflow so skips never write it),
         //  3. candidates persisted from prior runs (survive restarts for on-demand listing).
         const overflowJson = await this.scanOverflow(entityName);
+        // Every key literally present in the staging JSON right now. Kept separate from the candidate
+        // set below, which drops low-coverage keys — a stale key must be purged regardless of coverage.
+        const liveOverflowKeys = new Set();
+        for (const raw of overflowJson) {
+            const parsed = this.parseOverflow(raw);
+            if (parsed)
+                for (const k of Object.keys(parsed))
+                    liveOverflowKeys.add(k);
+        }
         // U3 note (rkihm-BC review, #3061): this in-repo promotion path passes no `LockUntilFullSync`, so it
         // does NOT yet enforce "hold promotion until a full sync since the last schema change." The lever
         // exists on `PromotionPlanOptions`, but pulling it here requires this caller to know whether a full
@@ -192,7 +378,10 @@ export class IntegrationCustomColumnPromoter {
             if (!byKey.has(c.Key))
                 byKey.set(c.Key, c);
         const passing = [...byKey.values()];
-        if (passing.length === 0)
+        // Nothing captured AND nothing staged — genuinely nothing to look at. A key can be staged
+        // without being a candidate (coverage filtered it out), and that key still needs purging, so
+        // this deliberately does NOT return early on `passing.length === 0` alone.
+        if (passing.length === 0 && liveOverflowKeys.size === 0)
             return null;
         const entityMap = await this.findEntityMap(companyIntegrationID, entityName);
         if (!entityMap) {
@@ -200,9 +389,19 @@ export class IntegrationCustomColumnPromoter {
             return null;
         }
         // Skip fully-terminated keys (column + field map both exist); keep promote (needs column) / recover.
-        const fieldMapSources = await this.activeFieldMapSources(entityMap.ID);
-        const work = this.resolveWorkItems(passing, entityInfo, fieldMapSources);
-        return { entityInfo, entityMap, work };
+        const fieldMaps = await this.activeFieldMaps(entityMap.ID);
+        const work = this.resolveWorkItems(passing, entityInfo, fieldMaps);
+        // Keys that are ALREADY promoted (active field map) yet whose value is still sitting in the
+        // staging JSON. These are not candidates — they are residue, and they cannot clear themselves:
+        // the next sync only rewrites a row whose content hash changed, and the hash basis excludes the
+        // overflow column, so a row untouched since before the promotion keeps the key indefinitely.
+        const stale = [];
+        for (const key of liveOverflowKeys) {
+            const columnName = fieldMaps.get(key.toLowerCase());
+            if (columnName)
+                stale.push({ sourceKey: key, columnName });
+        }
+        return { entityInfo, entityMap, work, stale };
     }
     /**
      * Lists the custom-column CANDIDATES for one entity — the "new columns found" awaiting promotion,
@@ -213,7 +412,9 @@ export class IntegrationCustomColumnPromoter {
         const planned = await this.planWorkForEntity(companyIntegrationID, entityName);
         if (!planned)
             return [];
-        return planned.work.map(w => ({
+        // recoverSpread items are not "new columns found" — they are an internal backfill-recovery
+        // signal (column + field map already exist), so they must not surface as UI candidates.
+        return planned.work.filter(w => !w.recoverSpread).map(w => ({
             EntityName: entityName,
             SourceKey: w.sourceKey,
             ColumnName: w.columnName,
@@ -326,16 +527,30 @@ export class IntegrationCustomColumnPromoter {
      * map is missing becomes a RECOVERY item (covers the crash window between ADD COLUMN and the
      * field-map write) — needsColumn=false, needsFieldMap=true.
      */
-    resolveWorkItems(passing, entityInfo, fieldMapSources) {
+    resolveWorkItems(passing, entityInfo, fieldMaps) {
         const existingByLower = new Map(entityInfo.Fields.map(f => [f.Name.toLowerCase(), f.Name]));
         const taken = new Set(entityInfo.Fields.map(f => f.Name.toLowerCase()));
         const items = [];
         for (const candidate of passing) {
-            const existingCol = existingByLower.get(sanitizeColumnName(candidate.Key).toLowerCase());
+            // An ACTIVE field map is the authoritative proof that this key was already promoted, and it
+            // names the column that was created. Trust it ahead of a re-sanitized guess against the
+            // in-memory field list: that list can predate the ADD COLUMN in THIS process (the promoter
+            // refreshes its own provider, not every other one), and the real column may carry a
+            // collision suffix the guess cannot reproduce. Both misses used to read as "no column yet",
+            // re-offering an already-promoted key to the operator as a brand-new column — and, on
+            // promotion, minting a duplicate `_2` column beside the working one.
+            const mappedDest = fieldMaps.get(candidate.Key.toLowerCase());
+            const hasFieldMap = mappedDest !== undefined;
+            const existingCol = (mappedDest ? existingByLower.get(mappedDest.toLowerCase()) ?? mappedDest : undefined) ??
+                existingByLower.get(sanitizeColumnName(candidate.Key).toLowerCase());
             const hasColumn = !!existingCol;
-            const hasFieldMap = fieldMapSources.has(candidate.Key.toLowerCase());
-            if (hasColumn && hasFieldMap)
-                continue; // terminated — nothing to do
+            if (hasColumn && hasFieldMap) {
+                // Schema + mapping are terminated — but see WorkItem.recoverSpread: an interrupted
+                // spread leaves rows carrying the value only in overflow, and this was the one exit
+                // that made that state permanent. Keep it as a spread-recovery item instead.
+                items.push({ candidate, sourceKey: candidate.Key, columnName: existingCol, needsColumn: false, needsFieldMap: false, recoverSpread: true });
+                continue;
+            }
             const columnName = existingCol ?? this.uniqueColumnName(sanitizeColumnName(candidate.Key), taken);
             if (!hasColumn)
                 taken.add(columnName.toLowerCase());
@@ -343,16 +558,33 @@ export class IntegrationCustomColumnPromoter {
         }
         return items;
     }
-    /** Active field-map SOURCE field names for an entity map (lowercased) — for the terminate/recovery check. */
-    async activeFieldMapSources(entityMapID) {
+    /**
+     * Active field maps for an entity map, as `lowercased SourceFieldName -> DestinationFieldName`.
+     *
+     * The destination name is what makes this authoritative: it records the column promotion ACTUALLY
+     * created, which may carry a collision suffix (`_2`) that re-sanitizing the source key can never
+     * reproduce. One query serves the terminate/recovery check, the stale-key detection and the hash
+     * re-baseline, where there used to be two identical queries and a lossier projection.
+     */
+    async activeFieldMaps(entityMapID) {
         const rv = new RunView();
         const res = await rv.RunView({
             EntityName: 'MJ: Company Integration Field Maps',
             ExtraFilter: `EntityMapID='${entityMapID}' AND Status='Active'`,
-            Fields: ['SourceFieldName'],
+            Fields: ['SourceFieldName', 'DestinationFieldName'],
             ResultType: 'simple',
         }, this.user);
-        return new Set(res.Success ? (res.Results ?? []).map(r => (r.SourceFieldName ?? '').toLowerCase()) : []);
+        const maps = new Map();
+        if (res.Success) {
+            for (const r of res.Results ?? []) {
+                const source = (r.SourceFieldName ?? '').toLowerCase();
+                // '' destination keeps the key MAPPED for the terminate check while signalling that no
+                // column name is recoverable from it — a `has()` miss and a `get()` of '' are different answers.
+                if (source)
+                    maps.set(source, r.DestinationFieldName ?? '');
+            }
+        }
+        return maps;
     }
     /** Suffixes _2, _3, … until the sanitized name does not collide with an existing/assigned one. */
     uniqueColumnName(base, taken) {
@@ -365,25 +597,52 @@ export class IntegrationCustomColumnPromoter {
         }
         return `${base}_${Date.now() % 100000}`; // pathological fallback (never expected)
     }
-    /** Generates ADD COLUMN DDL for the new columns and runs it through RSU (CodeGen reflects them). */
-    async applySchemaChange(entityInfo, named) {
+    /**
+     * Builds the RSUPipelineInput (ADD COLUMN DDL) for one entity's new columns — pure, runs NO
+     * pipeline. PromoteForSync collects these across all entities and runs ONE RunPipelineBatch so
+     * the whole promote is a single CodeGen/compile/restart pass.
+     */
+    buildSchemaInput(entityInfo, named) {
         const platform = this.dbProvider.PlatformKey;
         const ddl = new DDLGenerator();
         const statements = named.map(n => ddl.GenerateAlterTableAddColumn(entityInfo.SchemaName, entityInfo.BaseTable, this.toTargetColumn(n, platform), platform));
-        const input = {
+        return {
             MigrationSQL: statements.join('\n'),
             Description: `Promote ${named.length} custom column(s) on ${entityInfo.Name}`,
             AffectedTables: [`${entityInfo.SchemaName}.${entityInfo.BaseTable}`],
-            // M3 owns the restart-signal + restart; this stage applies schema only. Runtime
-            // promotion creates no git commit (the dev RSU flow does; a per-sync commit is noise).
-            SkipRestart: true,
-            SkipGitCommit: true,
+            // Neither SkipRestart nor SkipGitCommit is set, so RSU runs its normal pipeline. Both
+            // are optional and RSU gates on `!inputs.every(i => i.SkipGitCommit)`, so omitting them
+            // IS the default — no caller or platform change is needed to get commit + restart.
+            //
+            // `pm2 restart` kills this process, so PHASE 3 below does NOT run when the restart
+            // happens. That is why the follow-up is registered as PendingWork above: the
+            // post-restart consumer completes it, with the regenerated entity classes actually
+            // loaded. PHASE 3 remains as the fallback for the case where no DDL was needed (no
+            // batch, so no restart) — mirroring the apply path, which likewise finishes inline only
+            // when the restart did not occur.
+            //
+            // Both flags were previously hardcoded true here, the only place in the repo either was
+            // forced rather than passed in. Every integration entry point takes them as arguments
+            // defaulting to false, so add/remove tables, refresh schema and first-time setup all
+            // commit and restart; promotion was the outlier. Without the restart the columns never
+            // reached GraphQL — "already usable (metadata refreshed)" conflated metadata with CODE,
+            // since Refresh() reloads EntityField rows but does not load regenerated classes into a
+            // running process. Without the commit the database carried columns git had no record of.
+            // SkipGitCommit, by contrast, was wrong and is now dropped, so the migration and the
+            // regenerated code reach the repository. It is the only place in the repo either flag
+            // was forced rather than passed in; every integration entry point takes them as
+            // arguments defaulting to false, so add/remove tables, refresh schema and first-time
+            // setup all commit already. Without the commit the database carries columns git has no
+            // record of — observed live, where a workspace's promoted columns were present only
+            // because a LATER schema refresh happened to re-emit them as ADD COLUMN IF NOT EXISTS.
+            // The commit does not touch the process, so it is safe where the restart is not.
+            //
+            // The restart itself remains genuinely unsolved: it has to happen AFTER phase 3, and
+            // nothing performs it — `sync.schema_update` carries restartRequiredForGraphQL: true
+            // and has no subscriber, while the client arms its RSU poll and waits for a restart
+            // that never comes. Until that is closed, the columns do not reach GraphQL. Deliberately
+            // NOT fixed by restarting here, which would break promotion outright.
         };
-        const result = await RuntimeSchemaManager.Instance.RunPipeline(input);
-        if (!result.Success) {
-            LogError(`[CustomColumnPromoter] RSU ADD COLUMN failed on ${entityInfo.Name}: ${result.ErrorMessage ?? result.ErrorStep ?? 'unknown'}`);
-        }
-        return result.Success;
     }
     /** Builds a per-platform TargetColumnConfig from a planned candidate. */
     toTargetColumn(n, platform) {
@@ -476,55 +735,176 @@ export class IntegrationCustomColumnPromoter {
     /**
      * Spreads the staged overflow JSON values into the freshly-created real columns, then
      * re-baselines the content hash (gaps.md §2 step 3). JS per-row pass — dialect-agnostic (no
-     * cast SQL, which is where PG bugs hide); BaseEntity handles the dialect on write. The overflow
-     * column is intentionally NOT cleared: once the field map exists the key is no longer "unmapped",
-     * so the next sync stops re-capturing it and planPromotions skips the now-existing column — the
-     * stale value self-heals. Bounded to rows that carry overflow, paged, once on the discovery sync.
+     * cast SQL, which is where PG bugs hide); BaseEntity handles the dialect on write. Each promoted
+     * key is also STRIPPED from the staging JSON as it is spread: it is no longer unmapped, and
+     * leaving it behind is what later re-surfaces it as a phantom new column (see
+     * {@link purgeStaleOverflowKeys}). Bounded to rows that carry overflow, paged.
      */
     async spreadAndRebaseline(entityName, entityMapID, entityInfo, named) {
         const hasHashCol = entityInfo.Fields.some(f => f.Name === CONTENT_HASH_COLUMN);
         const mappedDestFields = hasHashCol ? await this.activeDestinationFields(entityMapID) : [];
-        let startRow = 0;
-        const pageSize = 500;
+        await this.forEachOverflowRow(entityName, entityInfo, row => this.spreadOneRow(row, named, hasHashCol, mappedDestFields));
+    }
+    /**
+     * Removes already-promoted keys from the staging JSON across the WHOLE table, spreading each
+     * value into its real column first if that column is still empty (a row promoted before the
+     * spread ever reached it must not lose the value on the way out).
+     *
+     * This exists because the residue cannot self-heal. The sync rewrites a row only when its content
+     * hash changes, and the hash basis deliberately excludes the overflow column — so a row that has
+     * not changed since before the promotion is never rewritten, keeps the promoted key forever, and
+     * keeps that key showing up in the candidate listing as a new column to add.
+     */
+    async purgeStaleOverflowKeys(entityName, entityMapID, entityInfo, stale) {
+        const columns = new Set(entityInfo.Fields.map(f => f.Name.toLowerCase()));
+        const purgeable = stale.filter(k => columns.has(k.columnName.toLowerCase()));
+        if (purgeable.length === 0) {
+            // Mapped but no such column in this process's metadata: the map may be newer than the
+            // metadata, so DON'T strip the staged value — that would be the only copy of it.
+            LogStatus(`[CustomColumnPromoter] ${stale.length} mapped key(s) on ${entityName} have no column in current metadata; purge skipped this pass.`);
+            return;
+        }
+        // Same re-baseline as the spread: the promoted columns are already in the field maps, so they
+        // are already part of what the next sync hashes. Backfilling one without re-baselining would
+        // leave every purged row hash-mismatched and force a needless rewrite on the next sync.
+        const hasHashCol = entityInfo.Fields.some(f => f.Name === CONTENT_HASH_COLUMN);
+        const mappedDestFields = hasHashCol ? await this.activeDestinationFields(entityMapID) : [];
+        let written = 0;
+        let removedFromSet = 0;
+        await this.forEachOverflowRow(entityName, entityInfo, async (row) => {
+            const result = await this.purgeOneRow(row, purgeable, hasHashCol, mappedDestFields);
+            if (result.wrote)
+                written++;
+            if (result.leftSet)
+                removedFromSet++;
+            return result.leftSet;
+        }, () => written >= MAX_PURGE_ROWS_PER_PASS);
+        const keys = purgeable.map(k => k.sourceKey).join(', ');
+        if (written >= MAX_PURGE_ROWS_PER_PASS) {
+            LogStatus(`[CustomColumnPromoter] Purged ${keys} from ${written} row(s) on ${entityName} — per-pass budget reached; any remaining residue is purged on the next sync.`);
+        }
+        else {
+            LogStatus(`[CustomColumnPromoter] Purged ${keys} from the staging JSON of ${written} row(s) on ${entityName} (already promoted).`);
+        }
+    }
+    /**
+     * Walks every row of `entityName` whose overflow JSON is non-null, in pages, applying `visit`.
+     *
+     * `visit` reports whether the row LEFT the filtered set (its JSON went null). Those removals shift
+     * every later row toward the front, so the offset advances by rows-seen-minus-rows-removed rather
+     * than by page size — otherwise the walk skips exactly as many rows as it cleans. Ordering is
+     * pinned to the primary key so the offsets refer to a stable sequence across the paged queries.
+     *
+     * `stop` is checked before each row and ends the walk early — a write budget, not a scan limit,
+     * so a capped caller still reads past rows it has nothing to do on. See {@link MAX_PURGE_ROWS_PER_PASS}.
+     */
+    async forEachOverflowRow(entityName, entityInfo, visit, stop) {
+        const orderBy = entityInfo.PrimaryKeys.map(pk => pk.Name).join(', ');
+        let seen = 0;
+        let removed = 0;
         for (;;) {
             const rv = new RunView();
             const res = await rv.RunView({
                 EntityName: entityName,
                 ExtraFilter: `${CUSTOM_OVERFLOW_COLUMN} IS NOT NULL`,
+                OrderBy: orderBy.length > 0 ? orderBy : undefined,
                 ResultType: 'entity_object',
-                MaxRows: pageSize,
-                StartRow: startRow,
+                MaxRows: OVERFLOW_PAGE_SIZE,
+                StartRow: seen - removed,
             }, this.user);
             if (!res.Success) {
-                LogError(`[CustomColumnPromoter] Spread scan failed for ${entityName}: ${res.ErrorMessage}`);
+                LogError(`[CustomColumnPromoter] Overflow row scan failed for ${entityName}: ${res.ErrorMessage}`);
                 return;
             }
             const rows = res.Results ?? [];
             for (const row of rows) {
-                await this.spreadOneRow(row, named, hasHashCol, mappedDestFields);
+                if (stop?.())
+                    return;
+                if (await visit(row))
+                    removed++;
+                seen++;
             }
-            if (rows.length < pageSize)
+            if (rows.length < OVERFLOW_PAGE_SIZE)
                 break;
-            startRow += pageSize;
         }
     }
-    /** Applies the staged values + re-baselined hash to a single row entity and saves it. */
+    /**
+     * Strips the given already-promoted keys from one row's staging JSON, backfilling any real column
+     * that is still empty.
+     *
+     * Reports both facts the caller needs and they are NOT the same: `wrote` is what the per-pass
+     * budget spends (one Save), `leftSet` is whether the row dropped out of the `IS NOT NULL` filter
+     * and so shifts the paged offsets. A row that keeps other unmapped keys is written but stays.
+     */
+    async purgeOneRow(row, stale, hasHashCol, mappedDestFields) {
+        const overflow = this.parseOverflow(row.Get(CUSTOM_OVERFLOW_COLUMN));
+        if (!overflow)
+            return { wrote: false, leftSet: false };
+        let changed = false;
+        for (const k of stale) {
+            if (!Object.prototype.hasOwnProperty.call(overflow, k.sourceKey))
+                continue;
+            const current = row.Get(k.columnName);
+            if (current === null || current === undefined)
+                row.Set(k.columnName, overflow[k.sourceKey]);
+            delete overflow[k.sourceKey];
+            changed = true;
+        }
+        // Nothing of ours in this row — a pure read, so it costs no budget. This is the case that makes
+        // a capped pass still able to reach residue further down: rows an earlier pass already cleaned
+        // fall through here instead of being re-written.
+        if (!changed)
+            return { wrote: false, leftSet: false };
+        const remaining = Object.keys(overflow).length > 0 ? JSON.stringify(overflow) : null;
+        row.Set(CUSTOM_OVERFLOW_COLUMN, remaining);
+        if (hasHashCol) {
+            const mapped = {};
+            for (const dest of mappedDestFields)
+                mapped[dest] = row.Get(dest);
+            row.Set(CONTENT_HASH_COLUMN, computeContentHash(mapped));
+        }
+        if (!await row.Save()) {
+            LogError(`[CustomColumnPromoter] Overflow purge save failed: ${row.LatestResult?.CompleteMessage ?? 'unknown'}`);
+            return { wrote: false, leftSet: false };
+        }
+        return { wrote: true, leftSet: remaining === null };
+    }
+    /**
+     * Applies the staged values + re-baselined hash to a single row entity and saves it. Returns true
+     * when the row's overflow JSON went null (it left the filtered set — see {@link forEachOverflowRow}).
+     */
     async spreadOneRow(row, named, hasHashCol, mappedDestFields) {
         // Dynamic .Get/.Set is REQUIRED here: these columns were created at runtime and have no
         // generated typed property in this still-running process (full typed access arrives on the
         // post-promotion restart). This is the sanctioned exception to the no-.Get/.Set rule.
         const overflow = this.parseOverflow(row.Get(CUSTOM_OVERFLOW_COLUMN));
         if (!overflow)
-            return;
+            return false;
         let changed = false;
         for (const n of named) {
             if (Object.prototype.hasOwnProperty.call(overflow, n.sourceKey)) {
-                row.Set(n.columnName, overflow[n.sourceKey]);
+                // Idempotent spread: only WRITE when the destination is still unset. A freshly-added
+                // column is null on first spread (so it fills); an already-backfilled column is left
+                // alone, so a re-run after an interrupted spread finishes the gap without overwriting
+                // a settled value.
+                const current = row.Get(n.columnName);
+                if (current === null || current === undefined) {
+                    row.Set(n.columnName, overflow[n.sourceKey]);
+                }
+                // The STRIP is unconditional, and deliberately outside the guard above: the key now has
+                // a real column and an active field map, so it is no longer unmapped whether or not this
+                // pass was the one that filled it. Leaving it behind relies on a later sync to evict it,
+                // which never happens for a row whose content hash does not change (the hash basis
+                // excludes this column). Stripping is what makes the recovery pass converge — once the
+                // key is gone from the JSON the row drops out of the scan entirely.
+                delete overflow[n.sourceKey];
                 changed = true;
             }
         }
         if (!changed)
-            return;
+            return false;
+        const remaining = Object.keys(overflow).length > 0 ? JSON.stringify(overflow) : null;
+        row.Set(CUSTOM_OVERFLOW_COLUMN, remaining);
         if (hasHashCol) {
             // Re-baseline to the next-sync value: hash over all active mapped destination columns
             // (now incl. the new ones) as they sit on the row — matches what the next sync computes.
@@ -535,18 +915,13 @@ export class IntegrationCustomColumnPromoter {
         }
         if (!await row.Save()) {
             LogError(`[CustomColumnPromoter] Spread save failed: ${row.LatestResult?.CompleteMessage ?? 'unknown'}`);
+            return false;
         }
+        return remaining === null;
     }
     /** Active field-map destination column names for an entity map (for hash re-baseline). */
     async activeDestinationFields(entityMapID) {
-        const rv = new RunView();
-        const res = await rv.RunView({
-            EntityName: 'MJ: Company Integration Field Maps',
-            ExtraFilter: `EntityMapID='${entityMapID}' AND Status='Active'`,
-            Fields: ['DestinationFieldName'],
-            ResultType: 'simple',
-        }, this.user);
-        return res.Success ? (res.Results ?? []).map(r => r.DestinationFieldName).filter(Boolean) : [];
+        return [...(await this.activeFieldMaps(entityMapID)).values()].filter(Boolean);
     }
     parseOverflow(raw) {
         if (typeof raw !== 'string' || raw.length === 0)