Plaid v2 writeback: MM 12-segment fan-out, atomic-graph hardening, Test Org on

Normalizes Monthly-Minimum writeback to v1-parity 12-segment fan-out, drops the racing in-graph OLI DELETEs that were tripping ENTITY_IS_DELETED rollbacks, and flips plaidWritebackPipelineV2 on for the Plaid Test Org.

1. Why this exists

Plaid v2 writeback maps Dealops engine output into Salesforce CPQ records. Three issues had been piling up in the sandbox push loop, all blocking the Plaid Test Org from going live on the v2 pipeline:

• MM count mismatch. An earlier prototype path (PLAID_MM_SINGLE_LINE) had emitted a single non-segmented MM line and let CPQ segment it, chasing a sandbox "24 MM lines" blowup. The blowup turned out to be a non-deterministic SF-side injection — 0/7 reproductions across all variants, including re-pushing the original trigger opp. Fan-out is clean in every observed push; v1 prod Q-55179 writes 12 segmented MM lines and the v2 path needs to match.
• Whole-graph rollback on re-push. The atomic composite graph was folding in explicit per-OLI DELETEs and the un-primary PATCH on the old primary quote. Un-primary triggers CPQ's own cascade-delete of synced OLIs; the explicit DELETE then raced the cascade, returned ENTITY_IS_DELETED, and rolled back the entire push (un-primary, opp PATCH, new quote, all QLIs).
• Sandbox field rejections. Custom prod fields like Product_GBT__c are present in the synced Plaid catalog's extraFields but the sandbox SBQQ__QuoteLine__c doesn't have the column — every QLI insert was failing INVALID_FIELD.

This PR closes dealops-a6g (MM normalization), folds in the already-in-tree fixes dealops-s7q / cv5 / c7u that the MM work depends on, and flips the org flag.

2. MM: from single line to 12 segments

The fan-out loop is now unconditional and walks subscriptionTerms directly. The env-gated branch is gone:

// dealops-a6g (RESOLVED): unconditional v1-parity fan-out.
for (const sfRow of mmSegmentEmissions) {
  const segmentKey = this.generatePlaidSegmentKey();
  // One segmented QLI per subscription month (v1 parity).
  const mmIterations = subscriptionTerms;
  for (let index = 0; index < mmIterations; index++) {
    // ... emit segmented QLI with SegmentIndex = index + 1
  }
}

3. Atomic graph: who deletes the stale OLIs?

The composite graph used to look like the left column below — un-primary first, then explicit per-OLI DELETEs, then the new writes. CPQ's cascade and our explicit DELETEs ran in the same atomic transaction and raced.

The asymmetric branch

OLI DELETEs are only safe to drop when an un-primary PATCH is in the same graph — that's what triggers CPQ's cascade. If there is no primary quote, there's no cascade to rely on, and we must clean up the stale OLIs ourselves or they'll co-exist with the freshly synced QLIs and corrupt opp totals. So the builder splits:

if (existingPrimaryQuoteCrmId) {
  // un-primary PATCH only — CPQ cascade handles the OLIs
  preprocessSubRequests.push(unprimarySubRequest);
} else if (oliCrmIdsToDelete.length > 0) {
  // no primary → no cascade. Emit explicit DELETEs; safe, nothing to race.
  oliCrmIdsToDelete.forEach((oliId, idx) => {
    preprocessSubRequests.push({ method: 'DELETE', url: ..., referenceId: ... });
  });
}

Discovery still computes oliCrmIdsToDelete in both branches so dry-run logs and metrics stay meaningful, and a switch back to explicit DELETEs is one branch flip away if CPQ cascade proves unreliable on the Plaid SF org.

4. Sandbox field strips (cv5 / c7u)

Custom QLI fields the Plaid prod SF org defines but the sandbox org does not. The catalog sync carries them in extraFields; every QLI insert was failing INVALID_FIELD. buildPlaidQliTailFields destructures them out before the safe-fields spread:

const {
  SBCF_Net_Unit_Price__c: _omitNetUnitPrice,
  SBQQ__TotalDiscountAmount__c: _omitTotalDiscount,
  SBQQ__ProductCode__c: _omitProductCode,
  Product_GBT__c: _omitProductGbt,   // dealops-cv5
  ...safeExtraFields
} = extraFields ?? {};

5. Feature flag

One line in packages/feature-flags/flags.ts appends the Plaid Test Org to the plaidWritebackPipelineV2 allowlist:

'e846ccc7-40b5-4897-a31b-761d6f51654c',
// Plaid Test Org - migrated to 2.0 (dealops-xq9.50;
// deploy gated on dealops-5yz wipe + reseed)

6. Tests

passing Full server mocha suite: 4347 / 0. Typecheck + lint clean. Live sandbox pushes: 12 clean MM segments across single-line, fan-out, and re-push runs.

7. What it doesn't change

8. Risks & follow-ups