fix: payments money-safety round-2 — webhook booking gate + M2 stranded-charge refund, sync-path status guards, gift-card cancel re-issue reconcile, till-sweep clawback lock, sweep VAT rescue, refund credit routing

- webhook payment.updated now re-reads the booking FOR UPDATE: non-payable booking -> payment failed + stranded-charge refund row + flood-capped alert; payable booking -> full completion side-effects; gift-card rows stay pending (C6 same-key retry); unknown events -> 503 so Square retries
- sync-path completion flips (saved-card, tip, online) guarded AND status='pending' + post-flip re-read; postChargeRecheck failed-mark guarded — no webhook-first double-processing, no phantom split rows
- CancelGiftCard resume reconciles ALL Square refunds (pending blocks re-issue; COMPLETED sum >= entitlement resolves+neutralizes; else re-issues only the difference under a fresh key) — closes double-refund
- sweep till_sales fail/clawback takes crussell:till:<key> lock + post-lock status re-read; recordUntrackedTillSalePayment applies VAT; rescue align clears VAT fields before re-apply (split-accurate VAT, tip rows stay VAT-free)
- refunds: chargeAggKey widened, redeemed-card refunds route to user_giftcard_balances, guest cash refunds recorded failed + notification
- handlers: tip split-records excluded from VAT loop, splitIdempotencyKey hashed, cross-booking key reuse 409, refunded payments excluded from existingCount, SCA-mint exemption for save_card, non-COMPLETED results routed

Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)

backend/handlers/payments/charge_helpers.go

@@ -294,9 +294,22 @@ func postChargeRecheck(ctx context.Context, w http.ResponseWriter, tx pgx.Tx, bo
 	if !payable {
 		log.Printf("CRITICAL: Square payment %s (ID=%s) for booking %s was processed but booking is now %q — marking %s failed; money taken at Square MUST be refunded manually",
 			sqStatus, sqPayID, bookingID, recheckStatus, chargeNoun)
-		if _, upErr := tx.Exec(ctx, `UPDATE payments SET status = 'failed' WHERE id = $1`, paymentID); upErr != nil {
+		// R10: guard the failed mark on status='pending' so it can never
+		// clobber a row the Square payment.completed webhook already resolved
+		// to 'completed'. In the race where the webhook wins the booking
+		// FOR UPDATE lock between the Square call and this recheck, the
+		// payment is genuinely completed — the cancellation refund path
+		// computes refunds from completed payments and will reverse the money
+		// there. Marking such a row 'failed' would strand a completed charge
+		// off the refund ledger entirely. A 0-row update (row already
+		// completed/failed) is left as-is; the CRITICAL log below still
+		// alerts ops that money landed on a no-longer-payable booking.
+		if res, upErr := tx.Exec(ctx, `UPDATE payments SET status = 'failed' WHERE id = $1 AND status = 'pending'`, paymentID); upErr != nil {
 			log.Printf("CRITICAL: Square payment %s (ID=%s) landed on %q booking %s but marking %s failed errored: %v — manual reconciliation required",
 				sqStatus, sqPayID, recheckStatus, bookingID, chargeNoun, upErr)
+		} else if res.RowsAffected() == 0 {
+			log.Printf("Square payment %s (ID=%s) landed on %q booking %s but payment row %s is no longer 'pending' (webhook/sweep already resolved it) — leaving the row as-is; not marking it failed",
+				sqStatus, sqPayID, bookingID, chargeNoun, paymentID)
 		}
 		if cErr := tx.Commit(ctx); cErr != nil {
 			log.Printf("CRITICAL: Square payment %s (ID=%s) landed on %q booking %s and committing the failed mark errored: %v — manual reconciliation required",

backend/handlers/payments/giftcards.go

@@ -109,8 +109,17 @@ type BuyGiftCardRequest struct {
 	RecipientType  string  `json:"recipient_type"`
 	RecipientEmail string  `json:"recipient_email,omitempty"`
 	CardID         *string `json:"card_id,omitempty"`
-	NewCardToken   *string `json:"new_card_token,omitempty"`
-	SaveCard       bool    `json:"save_card"`
+	// SavedCardID (saved_card_id) is the SCA path's reference to the stored
+	// saved card (user_saved_cards.id), mirroring CreateBookingPayment /
+	// CreateTerminalPayment / TillSaleRequest. It coexists with NewCardToken
+	// when the frontend sends the SCA tokenize-result —
+	// card.tokenize(verificationDetails, cardId) — as new_card_token: the
+	// tokenize-result token is a fresh one-time source_id and the saved-card
+	// row supplies the Square customer (resolveChargeSource). Without a token
+	// it behaves exactly like card_id (legacy saved-card charge).
+	SavedCardID   *string `json:"saved_card_id,omitempty"`
+	NewCardToken  *string `json:"new_card_token,omitempty"`
+	SaveCard      bool    `json:"save_card"`
 	// IdempotencyKey is optional (M1): if not provided, a deterministic key is
 	// generated server-side based on user_id + amount + recipient_type + card_id.
 	// This ensures retries of the same logical purchase use the same key while
@@ -1492,12 +1501,25 @@ func BuyGiftCard(w http.ResponseWriter, r *http.Request) {
 	// distinct no-key purchases never collapse onto one dedup key. Clients who
 	// need full control still supply their own idempotency_key; that path is
 	// unchanged.
+
+	// savedCardRef is the effective saved-card reference for this request:
+	// the legacy card_id field OR the SCA path's saved_card_id (they are the
+	// same user_saved_cards.id; card_id wins when both are sent). Mirrors
+	// CreateBookingPayment. scaTokenizedSavedCard is the SCA tokenize-result
+	// wire contract: the token is the one-time charge source and the saved-card
+	// row supplies the Square customer.
+	savedCardRef := req.CardID
+	if savedCardRef == nil || *savedCardRef == "" {
+		savedCardRef = req.SavedCardID
+	}
+	scaTokenizedSavedCard := req.NewCardToken != nil && *req.NewCardToken != "" && savedCardRef != nil && *savedCardRef != ""
+
 	clientSuppliedKey := req.IdempotencyKey != ""
 	var noKeyCardPart string
 	if !clientSuppliedKey {
 		noKeyCardPart = "new"
-		if req.CardID != nil && *req.CardID != "" {
-			noKeyCardPart = *req.CardID
+		if savedCardRef != nil && *savedCardRef != "" {
+			noKeyCardPart = *savedCardRef
 		}
 	}
 	if err := validators.Validate.Struct(&req); err != nil {
@@ -1524,12 +1546,24 @@ func BuyGiftCard(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	if err := ValidateCardInfo(req.CardID, nil, req.NewCardToken); err != nil {
+	if err := ValidateCardInfo(req.CardID, req.SavedCardID, req.NewCardToken); err != nil {
 		log.Printf("Failed to process request: %v", err)
 		http.Error(w, "Invalid request", http.StatusBadRequest)
 		return
 	}
 
+	// The SCA tokenize-result wire contract requires the saved-card reference
+	// in saved_card_id — the frontend sends it there, never card_id (the
+	// verification_token-less legacy buy keeps card_id alone). A token riding
+	// alongside the legacy card_id field is the ambiguous shape that used to
+	// silently drop saved_card_id and charge the SCA tokenize-result as a
+	// new-card one-off without a customer binding — reject it outright.
+	if req.CardID != nil && *req.CardID != "" && req.NewCardToken != nil && *req.NewCardToken != "" {
+		log.Printf("Gift-card buy rejected: card_id + new_card_token coexist — the SCA tokenize-result path requires saved_card_id")
+		http.Error(w, "Invalid card configuration: use saved_card_id for an SCA tokenized saved-card purchase", http.StatusBadRequest)
+		return
+	}
+
 	if err := ValidateVerificationToken(req.VerificationToken); err != nil {
 		log.Printf("Failed to process request: %v", err)
 		http.Error(w, "Invalid request", http.StatusBadRequest)
@@ -1709,7 +1743,7 @@ func BuyGiftCard(w http.ResponseWriter, r *http.Request) {
 	if req.VerificationToken != nil {
 		giftCardVerificationToken = *req.VerificationToken
 	}
-	if (req.CardID != nil && *req.CardID != "") || req.SaveCard {
+	if (savedCardRef != nil && *savedCardRef != "" && !scaTokenizedSavedCard) || req.SaveCard {
 		if gateOK, _ := requireTwoFactorForCardAccess(w, r, paymentService, userID, giftCardVerificationToken, reusePendingID == ""); !gateOK {
 			return
 		}
@@ -1720,8 +1754,11 @@ func BuyGiftCard(w http.ResponseWriter, r *http.Request) {
 	var savedCardCustomerID string
 	// Resolve the new-card-vs-saved-card Square source — shared with
 	// CreateBookingPayment/CreateTipPayment (see resolveChargeSource for the
-	// R6 rationale).
-	sourceID, savedCardID, savedCardCustomerID, sourceOK := resolveChargeSource(ctx, w, paymentService, userID, req.NewCardToken, req.CardID, req.SaveCard, "Card not found")
+	// R6 rationale). savedCardRef (card_id OR saved_card_id) is passed as the
+	// card reference; when an SCA tokenize-result token rides along in
+	// NewCardToken, resolveChargeSource uses the token as the source and the
+	// card row for the customer.
+	sourceID, savedCardID, savedCardCustomerID, sourceOK := resolveChargeSource(ctx, w, paymentService, userID, req.NewCardToken, savedCardRef, req.SaveCard, "Card not found")
 	if !sourceOK {
 		return
 	}
@@ -1885,7 +1922,7 @@ func BuyGiftCard(w http.ResponseWriter, r *http.Request) {
 	// saving a new card during this purchase — is customer-initiated: Square
 	// requires customer_details on stored-credential payments. A one-off cnon:
 	// nonce charge is not a stored credential and needs none.
-	if req.CardID != nil || req.SaveCard {
+	if (savedCardRef != nil && *savedCardRef != "") || req.SaveCard {
 		paymentReq.CustomerDetails = &square.CreateCustomerDetails{CustomerInitiated: true}
 	}
 
@@ -1949,7 +1986,7 @@ func BuyGiftCard(w http.ResponseWriter, r *http.Request) {
 	// records the completed charge (the gate verified without consuming, so a
 	// failed/ambiguous Square charge did not burn the code and a same-key
 	// retry could re-verify the SAME code).
-	if (req.CardID != nil && *req.CardID != "") || req.SaveCard {
+	if (savedCardRef != nil && *savedCardRef != "" && !scaTokenizedSavedCard) || req.SaveCard {
 		if consErr := twofa.ConsumePendingCode(ctx, issueTx, userID); consErr != nil {
 			log.Printf("CRITICAL: Square payment succeeded (ID=%s) but consuming the 2FA code for user %s failed: %v — manual reconciliation required", paymentResult.SquarePayID, userID, consErr)
 			http.Error(w, "internal server error", http.StatusInternalServerError)
@@ -2865,36 +2902,137 @@ func cancelGiftCardForUser(ctx context.Context, w http.ResponseWriter, r *http.R
 			refundAmount -= priorAmount
 			refundKey = paymentID + "-gccancel-diff-" + strconv.FormatInt(int64(math.Round(refundAmount*100)), 10)
 			log.Printf("Gift-card purchase %s has a prior partial refund of £%.2f — issuing the £%.2f remainder", paymentID, priorAmount, refundAmount)
-		case "pending":
-			// In-flight refund for this payment. Only resume if the pending
-			// row is OURS; a foreign pending row must not be re-issued.
-			if !strings.HasPrefix(priorKey, paymentID+"-gccancel-") {
+		case "pending", "failed":
+			// Resume an in-flight (pending) or previously-failed refund for
+			// this payment. Only resume a pending row that is OURS; a foreign
+			// pending row must not be re-issued.
+			if priorStatus == "pending" && !strings.HasPrefix(priorKey, paymentID+"-gccancel-") {
 				log.Printf("Gift-card cancel rejected: payment %s has a pending refund row %s from another flow", paymentID, priorRefundID)
 				http.Error(w, "A refund for this gift card purchase is already being processed — please try again later", http.StatusConflict)
 				return
 			}
 			refundID = priorRefundID
-			refundKey = priorKey
-			refundAmount = priorAmount
 			resumingRefund = true
-		case "failed":
-			// Money provably never moved — safe to re-issue with the stored
-			// key (or persist the deterministic key when a legacy row has
-			// none, mirroring ensureRefundKey).
-			refundID = priorRefundID
-			refundAmount = priorAmount
-			resumingRefund = true
-			if priorKey != "" {
-				refundKey = priorKey
-			} else {
-				if _, uErr := tx.Exec(ctx, `
-					UPDATE refunds SET idempotency_key = $1
-					WHERE id = $2 AND idempotency_key IS NULL`, refundKey, refundID); uErr != nil {
-					log.Printf("Failed to persist refund idempotency key for %s: %v", refundID, uErr)
-					http.Error(w, "internal server error", http.StatusInternalServerError)
+			// currentEntitlement is what THIS request's eligibility re-verified
+			// (the full purchase value, or the spend-verified unspent remainder).
+			// The prior attempt's amount may predate a till spend, so never
+			// re-issue more than min(priorAmount, currentEntitlement).
+			currentEntitlement := refundAmount
+
+			// Reconcile EVERY refund Square holds for this payment (pending and
+			// completed, any amount), not just an exact-amount COMPLETED one.
+			// The prior attempt may still be PENDING (in-flight) at Square —
+			// invisible to an exact-amount reconcile — and re-issuing under a
+			// fresh amount-derived key while it is in flight would mint a
+			// SECOND Square refund (double refund once the first completes).
+			// Only the full picture of the payment's Square refunds can decide
+			// whether a re-issue is safe, how much is still owed, and whether
+			// the money has already moved.
+			sqRefunds, recErr := reconcileAllRefundsAtSquare(ctx, squarePaymentID)
+			if recErr != nil {
+				// Square state unknown — do not re-issue (double-refund risk)
+				// and do not resolve: leave the row pending for the sweep.
+				log.Printf("Gift-card cancel resume aborted for %s: could not reconcile prior refund %s against Square: %v", code, refundID, recErr)
+				http.Error(w, "Refund state could not be verified — please try again", http.StatusInternalServerError)
+				return
+			}
+
+			// Any still-in-flight (PENDING) Square refund for this payment
+			// means money MAY still land. Re-issuing a second refund now would
+			// double-refund once the first completes. Leave the row 'pending'
+			// (this request makes no writes) for the sweep's refund
+			// reconciliation to settle, and do not touch the card.
+			for i := range sqRefunds {
+				if sqRefunds[i].Status == "PENDING" {
+					log.Printf("Gift-card cancel resume for %s: Square refund %s for payment %s is still PENDING (in flight) — leaving refund row %s pending for sweep reconciliation", code, sqRefunds[i].ID, squarePaymentID, refundID)
+					http.Error(w, "This refund is still being processed by the payment provider — please try again later", http.StatusConflict)
 					return
 				}
 			}
+
+			// Every Square refund for this payment is terminal. FAILED /
+			// REJECTED / CANCELED refunds never moved money and count as zero;
+			// sum the COMPLETED (and APPROVED — Square's terminal-completed)
+			// amounts to learn what has actually landed.
+			entitlementPence := int64(math.Round(currentEntitlement * 100))
+			alreadyRefundedPence := int64(0)
+			completedRefundID := ""
+			for i := range sqRefunds {
+				switch sqRefunds[i].Status {
+				case "COMPLETED", "APPROVED":
+					alreadyRefundedPence += sqRefunds[i].Amount
+					if completedRefundID == "" {
+						completedRefundID = sqRefunds[i].ID
+					}
+				}
+			}
+
+			// Never end up refunded more than the prior attempt claimed OR the
+			// re-verified entitlement, whichever is lower: a till spend between
+			// attempts shrinks the entitlement, and the prior attempt never had
+			// the right to more than its own amount.
+			maxRefund := math.Min(priorAmount, currentEntitlement)
+
+			if alreadyRefundedPence >= int64(math.Round(maxRefund*100)) {
+				// The money has already moved — at least the full cancellation
+				// entitlement has landed at Square. Mark the row completed with
+				// the Square refund id and neutralise the card so the returned
+				// value can never be spent on top of it.
+				if remaining > 0 {
+					if cerr := cancelGiftCardFunding(ctx, tx, code, userID, refundID, refundAmount); cerr != nil {
+						log.Printf("CRITICAL: gift card %s was already refunded at Square but funding reversal failed: %v — MANUAL RECONCILIATION REQUIRED", code, cerr)
+						http.Error(w, "internal server error", http.StatusInternalServerError)
+						return
+					}
+				}
+				if _, upErr := tx.Exec(ctx, `UPDATE refunds SET status = 'completed', square_refund_id = $1 WHERE id = $2`, completedRefundID, refundID); upErr != nil {
+					log.Printf("CRITICAL: gift card %s was already refunded at Square but refund row %s could not be resolved: %v — MANUAL RECONCILIATION REQUIRED", code, refundID, upErr)
+					http.Error(w, "internal server error", http.StatusInternalServerError)
+					return
+				}
+				if cerr := tx.Commit(ctx); cerr != nil {
+					log.Printf("Failed to commit gift-card cancel transaction: %v", cerr)
+					http.Error(w, "internal server error", http.StatusInternalServerError)
+					return
+				}
+				w.WriteHeader(http.StatusOK)
+				if err := json.NewEncoder(w).Encode(map[string]any{
+					"status":          "success",
+					"message":         "This gift card had already been refunded and has now been cancelled.",
+					"refund_id":       refundID,
+					"amount_refunded": refundAmount,
+				}); err != nil {
+					log.Printf("Failed to encode JSON response: %v", err)
+				}
+				return
+			}
+
+			// Less than the entitlement has landed — re-issue ONLY the
+			// difference (entitlement minus what already completed at Square),
+			// still capped at min(priorAmount, currentEntitlement), under a
+			// FRESH deterministic key for THAT difference. The prior key
+			// encodes the prior amount and must NOT be reused when the amount
+			// changed: replaying the stale amount after a till spend would
+			// over-refund. When nothing has landed the key is derived exactly
+			// as the original attempt's (same amount → same key → Square dedups
+			// a prior landed refund). The row is updated to the new key+amount
+			// so a crash-retry re-reads the same state and the sweep reconciles
+			// consistently.
+			refundAmount = float64(entitlementPence-alreadyRefundedPence) / 100
+			if refundAmount > maxRefund {
+				refundAmount = maxRefund
+			}
+			if alreadyRefundedPence > 0 {
+				refundKey = paymentID + "-gccancel-diff-" + strconv.FormatInt(int64(math.Round(refundAmount*100)), 10)
+			} else {
+				refundKey = paymentID + "-gccancel-" + strconv.FormatInt(int64(math.Round(refundAmount*100)), 10)
+			}
+			if _, uErr := tx.Exec(ctx, `UPDATE refunds SET amount = $1, idempotency_key = $2 WHERE id = $3`, refundAmount, refundKey, refundID); uErr != nil {
+				log.Printf("Failed to update resumed gift-card cancel refund %s to £%.2f / key %s: %v", refundID, refundAmount, refundKey, uErr)
+				http.Error(w, "internal server error", http.StatusInternalServerError)
+				return
+			}
+			log.Printf("Gift-card purchase %s has already refunded £%.2f at Square of the £%.2f entitlement — issuing the £%.2f remainder with a fresh key", paymentID, float64(alreadyRefundedPence)/100, currentEntitlement, refundAmount)
 		}
 	}
 
@@ -3154,6 +3292,32 @@ func cancelGiftCardForUser(ctx context.Context, w http.ResponseWriter, r *http.R
 	}
 }
 
+// reconcileAllRefundsAtSquare lists EVERY refund Square has recorded for the
+// payment (pending AND completed, unfiltered by amount) so the gift-card cancel
+// resume can see an in-flight (PENDING) prior attempt that
+// reconcileRefundAtSquareExact — which only matches an exact-amount COMPLETED
+// refund — would miss. Missing it is the double-refund bug: the resume would
+// re-issue under a fresh amount-derived key while the prior refund is still in
+// flight, minting a SECOND Square refund that lands on top of the first. The
+// returned slice lets the caller classify refunds by Status (PENDING blocks a
+// re-issue; COMPLETED/APPROVED sum to the already-refunded money;
+// FAILED/REJECTED/CANCELED never moved money and count as zero). A non-nil
+// error means Square state is unknown and no re-issue decision may be made.
+func reconcileAllRefundsAtSquare(ctx context.Context, chargeID string) ([]square.RefundResult, error) {
+	refunds, err := SquareClient.ListPaymentRefunds(ctx, chargeID, time.Time{})
+	if err != nil {
+		log.Printf("Failed to reconcile refunds for charge %s against Square: %v", chargeID, err)
+		return nil, err
+	}
+	ours := make([]square.RefundResult, 0, len(refunds))
+	for _, r := range refunds {
+		if r.PaymentID == chargeID {
+			ours = append(ours, r)
+		}
+	}
+	return ours, nil
+}
+
 // giftCardSpendAtTill returns the total value of completed payments made
 // against the gift card at the till. Till spend (handlers.go) decrements
 // amount_remaining and records a payments row with payment_method='giftcard',

backend/handlers/payments/handlers.go

@@ -1297,14 +1297,68 @@ func CreateTerminalPayment(w http.ResponseWriter, r *http.Request) {
 			}
 		}
 
-		if _, upErr := recheckTx.Exec(r.Context(),
-			`UPDATE payments SET status = 'completed', square_payment_id = $1 WHERE id = $2`,
+		// R10: guard the completion flip on status='pending'. The Square
+		// payment.completed webhook (webhooks/square.go) can win the booking
+		// FOR UPDATE lock between Square returning and this flip, completing
+		// the payment row and running the booking-completion side-effects
+		// itself. Without the guard this UPDATE would blindly re-flip the
+		// already-completed row (RowsAffected=1), re-run VAT and the
+		// completion side-effects below, and double-record money. With the
+		// guard, a row the webhook/sweep already resolved is a no-op
+		// (RowsAffected=0): the payment IS completed — skip the side-effects
+		// and report success.
+		flipRes, upErr := recheckTx.Exec(r.Context(),
+			`UPDATE payments SET status = 'completed', square_payment_id = $1 WHERE id = $2 AND status = 'pending'`,
 			paymentResult.SquarePayID, paymentID,
-		); upErr != nil {
+		)
+		if upErr != nil {
 			log.Printf("CRITICAL: Square payment %s succeeded but saved-card payment %s update failed: %v — manual reconciliation required", paymentResult.SquarePayID, paymentID, upErr)
 			http.Error(w, "internal server error", http.StatusInternalServerError)
 			return
 		}
+		if flipRes.RowsAffected() == 0 {
+			// Defense-in-depth re-read (R10): the guarded flip no-opped, so a
+			// concurrent resolver already moved the row off 'pending'. Re-read
+			// the status inside THIS transaction to distinguish a webhook-
+			// completed row (skip everything — the webhook already ran the
+			// completion side-effects) from a vanished/failed row (CRITICAL).
+			var curStatus string
+			rErr := recheckTx.QueryRow(r.Context(), `SELECT status FROM payments WHERE id = $1`, paymentID).Scan(&curStatus)
+			if rErr != nil || curStatus != "completed" {
+				log.Printf("CRITICAL: Square payment %s (ID=%s) succeeded but payment row %s could not be verified as completed (re-read status=%q err=%v) — manual reconciliation required",
+					paymentResult.Status, paymentResult.SquarePayID, paymentID, curStatus, rErr)
+				http.Error(w, "internal server error", http.StatusInternalServerError)
+				return
+			}
+			log.Printf("Saved-card payment %s for booking %s was already resolved to %q by a concurrent resolver (Square webhook/sweep) before the sync completion flip — skipping completion side-effects", paymentID, bookingID, curStatus)
+			// MEDIUM-2: burn the re-issued 2FA code anyway (idempotent) — the
+			// charge reached terminal success, so a single-use code re-issued
+			// for this retry must not authorize another charge.
+			if bookingUserID.Valid && reusePendingRecord {
+				if consErr := twofa.ConsumePendingCode(r.Context(), recheckTx, bookingUserID.String); consErr != nil {
+					log.Printf("CRITICAL: Square payment %s succeeded but consuming the 2FA code for user %s failed: %v — manual reconciliation required", paymentResult.SquarePayID, bookingUserID.String, consErr)
+					http.Error(w, "internal server error", http.StatusInternalServerError)
+					return
+				}
+			}
+			if cErr := recheckTx.Commit(r.Context()); cErr != nil {
+				log.Printf("CRITICAL: Square payment %s succeeded but committing the post-charge transaction for already-completed payment %s failed: %v — manual reconciliation required",
+					paymentResult.SquarePayID, paymentID, cErr)
+				http.Error(w, "internal server error", http.StatusInternalServerError)
+				return
+			}
+			// The payment IS completed (by the webhook/sweep) — report success
+			// exactly like the normal completion path; no side-effects re-run.
+			if err := json.NewEncoder(w).Encode(map[string]any{
+				"payment_id": paymentID,
+				"status":     "COMPLETED",
+				"card_brand": paymentResult.CardBrand,
+				"card_last4": paymentResult.CardLast4,
+			}); err != nil {
+				log.Printf("Failed to encode JSON response: %v", err)
+			}
+			return
+		}
 		// MEDIUM-2 / finding 1: the charge reached its terminal SUCCESS state.
 		// For a FRESH charge the gate already consumed the code (consume=true
 		// at verify time — single-use), so nothing is left to do here. A
@@ -2226,6 +2280,39 @@ func CreateBookingPayment(w http.ResponseWriter, r *http.Request) {
 		return
 	case err != nil && !errors.Is(err, pgx.ErrNoRows):
 		log.Printf("Failed to check idempotency: %v", err)
+	case errors.Is(err, pgx.ErrNoRows):
+		// R12: no payment on THIS booking carries the key, but the key may
+		// still exist globally — payments.idempotency_key is UNIQUE — on a
+		// DIFFERENT booking. Without a guard the pending insert below would die
+		// on the constraint → 500 → the frontend's same-key retry 500s forever.
+		// Mirror the A6 cross-user collision pattern (giftcards.go:1592-1607):
+		// a foreign match is not this booking's operation, so derive a fresh
+		// deterministic key for THIS booking and proceed as a fresh charge.
+		// A deterministic no-client-key fallback already embeds this booking's
+		// id, so this branch can only fire for a client-supplied key.
+		var foreignID string
+		fErr := tx.QueryRow(r.Context(), `
+			SELECT id FROM payments
+			WHERE idempotency_key = $1 AND booking_id != $2
+		`, req.IdempotencyKey, bookingID).Scan(&foreignID)
+		if fErr == nil {
+			log.Printf("Payment idempotency key %q matched payment row %s on a different booking — deriving a fresh deterministic key for booking %s", req.IdempotencyKey, foreignID, bookingID)
+			cardPart := "new"
+			if savedCardRef != nil && *savedCardRef != "" {
+				cardPart = *savedCardRef
+			}
+			derivedKey, dErr := deriveBookingPaymentIdempotencyKey(r.Context(), tx, bookingID, req.PaymentType, req.Amount, cardPart)
+			if dErr != nil {
+				log.Printf("Failed to derive fresh booking idempotency key after cross-booking collision: %v", dErr)
+				http.Error(w, "internal server error", http.StatusInternalServerError)
+				return
+			}
+			req.IdempotencyKey = derivedKey
+		} else if !errors.Is(fErr, pgx.ErrNoRows) {
+			log.Printf("Failed to check cross-booking idempotency collision for key %q: %v", req.IdempotencyKey, fErr)
+			http.Error(w, "internal server error", http.StatusInternalServerError)
+			return
+		}
 	}
 
 	// 2FA gating (C5): persisting a card requires 2FA when the feature is
@@ -2254,7 +2341,23 @@ func CreateBookingPayment(w http.ResponseWriter, r *http.Request) {
 	// needed. The SAVE gate is skipped because the combined path never persists
 	// a card (resolveChargeSource uses the token as a one-time source, no
 	// card-on-file is created).
-	if req.SaveCard && !scaTokenizedSavedCard {
+	//
+	// R13: the charge-time SAVE gate's tokenForwardedToSquare=false variant
+	// refuses 402 verification_required in an enforced deployment — but a
+	// genuine SCA tokenize-result (cnon:sca-... / verify_mock_..., the shapes
+	// the frontend's tokenizeWithVerification mints) only exists after the
+	// buyer completed the STORE-intent SCA, so it is SCA-proven and skips the
+	// gate. This unblocks the NEW-card SCA tokenize-result save (a cnon:sca-...
+	// token with save_card=true and no saved-card reference), which
+	// resolveChargeSource persists via CreateCardOnFile. A plain card-entry
+	// nonce (cnon:... without the sca- marker — card.tokenize() with no
+	// verification) is NOT an SCA proof and keeps failing closed through the
+	// gate — 402 in an enforced deployment (auth-F1).
+	saveGateToken := ""
+	if req.NewCardToken != nil {
+		saveGateToken = *req.NewCardToken
+	}
+	if req.SaveCard && !scaTokenizedSavedCard && !isSCATokenizeResultShape(saveGateToken) {
 		if gateOK, _ := requireTwoFactorForCardAccessWithTokenValidation(w, r, service, userID, bookingVerificationToken, true, false); !gateOK {
 			return
 		}
@@ -2270,13 +2373,17 @@ func CreateBookingPayment(w http.ResponseWriter, r *http.Request) {
 	if req.PaymentType != "partial" {
 		var existingCount int
 		if err := tx.QueryRow(r.Context(), `
-			SELECT COUNT(*) FROM payments
-			WHERE booking_id = $1
-			  AND status = 'completed'
-			  AND payment_method NOT IN ('discount', 'on_the_house')
+			SELECT COUNT(*) FROM payments p
+			WHERE p.booking_id = $1
+			  AND p.status = 'completed'
+			  AND p.payment_method NOT IN ('discount', 'on_the_house')
 			  AND (
-				  payment_type = $2
-				  OR ($2 IN ('full', 'deposit') AND payment_type = 'deposit')
+				  p.payment_type = $2
+				  OR ($2 IN ('full', 'deposit') AND p.payment_type = 'deposit')
+			  )
+			  AND NOT EXISTS (
+				  SELECT 1 FROM refunds r
+				  WHERE r.payment_id = p.id AND r.status IN ('completed', 'pending')
 			  )
 		`, bookingID, req.PaymentType).Scan(&existingCount); err == nil && existingCount > 0 {
 			log.Printf("Payment rejected: booking %s already has a completed %q payment", bookingID, req.PaymentType)
@@ -2688,6 +2795,20 @@ func CreateBookingPayment(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	// R11: a nil error does not mean the payment is COMPLETED. Square's
+	// terminal payment states are COMPLETED, CANCELED, FAILED; APPROVED
+	// (authorization-only) and PENDING are NON-terminal — both can still
+	// transition to COMPLETED. Mirror the stale-pending sweep's classification
+	// (sweep.go classifyStalePendingByKey): APPROVED/PENDING stay pending for a
+	// later sweep run to re-poll (staleReconcileLeavePending), CANCELED/FAILED
+	// are marked failed (the charge never landed). A status-blind handler that
+	// records 'completed' on nil error alone would book money that was never
+	// collected.
+	if paymentResult.Status != "COMPLETED" {
+		routeNonCompletedPayment(r.Context(), w, bookingID, paymentID, paymentResult, "booking")
+		return
+	}
+
 	paymentAmount := float64(chargeAmount) / 100.0
 
 	// Step 3: Square succeeded — record the completed payment state in a NEW
@@ -2793,7 +2914,17 @@ func CreateBookingPayment(w http.ResponseWriter, r *http.Request) {
 	// fields are cleared so apply_vat_to_payment recomputes on the final amount
 	// — the pending record had VAT applied at the pre-split amount.
 	primary := records[0]
-	if _, upErr := tx2.Exec(r.Context(), `
+	// R10: guard the split-primary flip on status='pending'. The Square
+	// payment.completed webhook can win the booking FOR UPDATE lock between
+	// the Square call and this UPDATE and complete the payment row + run the
+	// completion side-effects itself. Without the guard this UPDATE would
+	// blindly re-flip the already-completed row (RowsAffected=1), OVERWRITE
+	// the primary row's amount/payment_type/fees/VAT with the split values,
+	// and re-insert duplicate split tip/balance rows — the ledger would no
+	// longer reconcile to the actual Square charge. With the guard a row the
+	// webhook/sweep already resolved is a no-op (RowsAffected=0): the payment
+	// IS completed — skip the split/booking side-effects and report success.
+	flipRes, upErr := tx2.Exec(r.Context(), `
 		UPDATE payments SET
 			status = 'completed',
 			square_payment_id = $1,
@@ -2805,13 +2936,62 @@ func CreateBookingPayment(w http.ResponseWriter, r *http.Request) {
 			vat_amount = NULL,
 			net_amount = NULL,
 			updated_at = NOW()
-		WHERE id = $5
-	`, paymentResult.SquarePayID, primary.Amount, primary.PaymentType, primary.Fees, paymentID); upErr != nil {
+		WHERE id = $5 AND status = 'pending'
+	`, paymentResult.SquarePayID, primary.Amount, primary.PaymentType, primary.Fees, paymentID)
+	if upErr != nil {
 		log.Printf("CRITICAL: Square payment %s (ID=%s) was processed but updating payment %s to completed failed: %v — manual reconciliation required",
 			paymentResult.Status, paymentResult.SquarePayID, paymentID, upErr)
 		http.Error(w, "internal server error", http.StatusInternalServerError)
 		return
 	}
+	if flipRes.RowsAffected() == 0 {
+		// Defense-in-depth re-read (R10): distinguish a webhook-completed row
+		// (skip everything — the webhook already ran the completion side-
+		// effects; the row keeps the webhook's values, not the split overwrite)
+		// from a vanished/failed row (CRITICAL — money was taken at Square).
+		var curStatus string
+		rErr := tx2.QueryRow(r.Context(), `SELECT status FROM payments WHERE id = $1`, paymentID).Scan(&curStatus)
+		if rErr != nil || curStatus != "completed" {
+			log.Printf("CRITICAL: Square payment %s (ID=%s) was processed but payment row %s could not be verified as completed (re-read status=%q err=%v) — manual reconciliation required",
+				paymentResult.Status, paymentResult.SquarePayID, paymentID, curStatus, rErr)
+			http.Error(w, "internal server error", http.StatusInternalServerError)
+			return
+		}
+		log.Printf("Payment %s for booking %s was already resolved to %q by a concurrent resolver (Square webhook/sweep) before the sync completion flip — skipping split records, VAT and booking-completion side-effects", paymentID, bookingID, curStatus)
+		// MEDIUM-2: burn the re-issued 2FA code anyway (idempotent) — the
+		// charge reached terminal success, so a single-use code re-issued for
+		// this retry must not authorize another charge.
+		if savedCardRef != nil && *savedCardRef != "" && reusePendingRecord {
+			if consErr := twofa.ConsumePendingCode(r.Context(), tx2, userID); consErr != nil {
+				log.Printf("CRITICAL: Square payment %s (ID=%s) was processed but consuming the 2FA code for user %s failed: %v — manual reconciliation required",
+					paymentResult.Status, paymentResult.SquarePayID, userID, consErr)
+				http.Error(w, "internal server error", http.StatusInternalServerError)
+				return
+			}
+		}
+		if cErr := tx2.Commit(r.Context()); cErr != nil {
+			log.Printf("CRITICAL: Square payment %s (ID=%s) was processed but committing the transaction for already-completed payment %s failed: %v — manual reconciliation required",
+				paymentResult.Status, paymentResult.SquarePayID, paymentID, cErr)
+			http.Error(w, "internal server error", http.StatusInternalServerError)
+			return
+		}
+		// The payment IS completed (by the webhook/sweep) — report success
+		// exactly like the normal completion path; no side-effects re-run.
+		if err := json.NewEncoder(w).Encode(PaymentResponse{
+			ID:          paymentID,
+			BookingID:   bookingID,
+			PaymentType: req.PaymentType,
+			Status:      "completed",
+			Amount:      chargeAmount,
+			CardBrand:   paymentResult.CardBrand,
+			CardLast4:   paymentResult.CardLast4,
+			ReceiptURL:  paymentResult.ReceiptURL,
+			CreatedAt:   clock.Now().Format(time.RFC3339),
+		}); err != nil {
+			log.Printf("Failed to encode JSON response: %v", err)
+		}
+		return
+	}
 
 	// MEDIUM-2 / finding 1: a saved-card charge reached its terminal SUCCESS
 	// state. For a FRESH charge the gate already consumed the code
@@ -2846,10 +3026,24 @@ func CreateBookingPayment(w http.ResponseWriter, r *http.Request) {
 
 	// Apply VAT to all split records if the business is VAT-registered.
 	// Must be inside the transaction so VAT updates are atomic with inserts.
+	// Tip records are EXCLUDED: vat.go's single-source policy (see
+	// ApplyVATToBookingPayment, which skips discount / on_the_house / tip rows)
+	// and every other tip path (CreateTipPayment, the terminal sweep rescue)
+	// apply VAT to the booking portion only — an overflow payment that carves a
+	// payment_type='tip' record must never have VAT applied to the tip. The
+	// records slice holds the primary (records[0], the committed pending row
+	// paymentID) followed by the additional split records (paymentIDs, in
+	// order), so the loop keys on the record's own payment_type.
 	vatCfg, vatErr := GetVATConfig(r.Context(), tx2)
 	if vatErr == nil && vatCfg.IsVATRegistered {
-		vatIDs := append([]string{paymentID}, paymentIDs...)
-		for _, pid := range vatIDs {
+		for i, rec := range records {
+			if rec.PaymentType == "tip" {
+				continue
+			}
+			pid := paymentID
+			if i > 0 {
+				pid = paymentIDs[i-1]
+			}
 			if _, execErr := tx2.Exec(r.Context(), "SELECT apply_vat_to_payment($1, $2)", pid, vatCfg.DefaultVATRate); execErr != nil {
 				log.Printf("Failed to apply VAT to payment %s: %v", pid, execErr)
 			}
@@ -2940,6 +3134,30 @@ func CreateBookingPayment(w http.ResponseWriter, r *http.Request) {
 	}
 }
 
+// routeNonCompletedPayment handles a CreatePayment result whose status is not
+// COMPLETED, mirroring the stale-pending sweep's classification (sweep.go
+// classifyStalePendingByKey): APPROVED and PENDING are NON-terminal — Square
+// may still settle the charge — so the pending row is left pending for a later
+// sweep run to re-poll (staleReconcileLeavePending); CANCELED, FAILED and any
+// unknown status are terminal non-success — the charge never landed, so the row
+// is marked failed so a same-key retry can never issue a second Square charge.
+// The HTTP response is a 5xx in both cases: the payment is NOT complete, the
+// frontend must not show success, and the sweep reconciles the row in the
+// background regardless.
+func routeNonCompletedPayment(ctx context.Context, w http.ResponseWriter, bookingID, paymentID string, paymentResult *square.PaymentResult, label string) {
+	switch paymentResult.Status {
+	case "APPROVED", "PENDING":
+		log.Printf("%s payment %s for booking %s is %q at Square (non-terminal) — leaving the row pending; the stale-pending sweep will reconcile it", label, paymentID, bookingID, paymentResult.Status)
+		http.Error(w, "Payment is still processing at the payment provider and has not yet been confirmed", http.StatusInternalServerError)
+	default:
+		log.Printf("%s payment %s for booking %s is %q at Square (terminal non-success) — marking the row failed", label, paymentID, bookingID, paymentResult.Status)
+		if _, upErr := db.Conn.Exec(ctx, `UPDATE payments SET status = 'failed' WHERE id = $1 AND status = 'pending'`, paymentID); upErr != nil {
+			log.Printf("CRITICAL: Square %s payment %s (ID=%s) is %q but marking payment %s failed errored: %v — manual reconciliation required", label, paymentID, paymentResult.SquarePayID, paymentResult.Status, paymentID, upErr)
+		}
+		http.Error(w, "Payment failed", http.StatusPaymentRequired)
+	}
+}
+
 // campaignExhaustedAtApplyError reports that a discount campaign the customer
 // was shown as eligible at preview time was exhausted (times_redeemed reached
 // max_redemptions) by the time the payment applied it (B13). lostPence is the
@@ -3325,13 +3543,18 @@ func buildTerminalSplitRecords(primary PaymentRecord, info *BookingPaymentInfo,
 // record. The terminal base key (booking + amount + Square payment ID) can be
 // long enough that appending a -split-tip suffix would exceed the
 // payments.idempotency_key VARCHAR(64) limit (e.g. the dev mock's 28-char
-// "pay_mock_<nanosecond>" payment IDs); the base is truncated so the suffix
-// always fits. Uniqueness is preserved: the truncated base still embeds the
-// booking ID and Square payment ID, and the suffix differs per split record.
+// "pay_mock_<nanosecond>" payment IDs); an over-length base is routed through
+// the deterministic truncateIdempotencyKey helper (idempotency_helpers.go) so
+// the key always fits and the suffix stays verbatim (the split role remains
+// readable and the -split-1 / -split-tip distinction is exact). Deterministic
+// hashing — NOT a raw prefix cut — preserves uniqueness: two distinct base keys
+// (e.g. Square payment IDs differing only in their tail) must never collapse
+// onto the same truncated prefix, which would 500 the second insert on the
+// UNIQUE(idempotency_key) index.
 func splitIdempotencyKey(base, suffix string) string {
 	maxBase := 64 - len(suffix)
 	if len(base) > maxBase {
-		base = base[:maxBase]
+		base = truncateIdempotencyKey("split", base)
 	}
 	return base + suffix
 }
@@ -3510,6 +3733,20 @@ func isTokenLikeSaveSource(cardToken string) bool {
 	return strings.HasPrefix(cardToken, "cnon:") || strings.HasPrefix(cardToken, "ccof:") || strings.HasPrefix(cardToken, "verify_mock_")
 }
 
+// isSCATokenizeResultShape reports whether a card token is a genuine SCA
+// tokenize-result — the shapes Square mints only AFTER the buyer completed
+// issuer verification (the frontend's tokenizeWithVerification returns
+// cnon:sca-<prefix>_<amount>_ok|_deny; verify_mock_ is the dev/mock transition
+// shape). Mirrors the dev mock's isSCATokenizeResultSource. Unlike
+// isTokenLikeSaveSource (the DEDICATED add-card surface, where every token
+// passed through a STORE-intent tokenize flow), a plain card-entry nonce
+// (cnon:... without the sca- marker) is NOT an SCA proof — it is a raw
+// card.tokenize() result — so the charge-time SAVE gate must keep refusing it
+// 402 in an enforced deployment (auth-F1).
+func isSCATokenizeResultShape(cardToken string) bool {
+	return strings.HasPrefix(cardToken, "cnon:sca-") || strings.HasPrefix(cardToken, "verify_mock_")
+}
+
 // isDefinitiveCardSaveFailure reports whether a CreatePaymentMethod error is a
 // definitive client rejection — an expired/invalid/already-used card source or
 // a declined card that can never be saved — as opposed to an ambiguous
@@ -5033,6 +5270,15 @@ func CreateTipPayment(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	// R11: a nil error does not mean the payment is COMPLETED — see
+	// CreateBookingPayment. APPROVED/PENDING stay pending for the sweep to
+	// re-poll (staleReconcileLeavePending); CANCELED/FAILED are marked failed
+	// (the tip charge never landed).
+	if paymentResult.Status != "COMPLETED" {
+		routeNonCompletedPayment(r.Context(), w, bookingID, paymentID, paymentResult, "tip")
+		return
+	}
+
 	// Step 3a: post-charge recheck (R9). A concurrent cancellation/eviction
 	// can move the booking out of a payable state between the pre-charge
 	// status check and the Square charge completing. A tip landing on a
@@ -5065,16 +5311,70 @@ func CreateTipPayment(w http.ResponseWriter, r *http.Request) {
 	}
 
 	// Step 3: Square succeeded — update the payment record.
-	if _, upErr := recheckTx.Exec(r.Context(),
-		`UPDATE payments SET status = 'completed', square_payment_id = $1 WHERE id = $2`,
+	// R10: guard the flip on status='pending'. The Square payment.completed
+	// webhook can win the booking FOR UPDATE lock between the Square call and
+	// this UPDATE and complete the tip row itself. Without the guard this
+	// would blindly re-flip the already-completed row and re-run the 2FA
+	// consume below; with it, a row the webhook/sweep already resolved is a
+	// no-op (RowsAffected=0): the tip IS completed — report success.
+	flipRes, upErr := recheckTx.Exec(r.Context(),
+		`UPDATE payments SET status = 'completed', square_payment_id = $1 WHERE id = $2 AND status = 'pending'`,
 		paymentResult.SquarePayID, paymentID,
-	); upErr != nil {
+	)
+	if upErr != nil {
 		log.Printf("Failed to update payment %s after Square success: %v (square_payment_id=%s)", paymentID, upErr, paymentResult.SquarePayID)
 		// Square charge succeeded but status update failed.
 		// Record stays 'pending' for manual reconciliation.
 		http.Error(w, "internal server error", http.StatusInternalServerError)
 		return
 	}
+	if flipRes.RowsAffected() == 0 {
+		// Defense-in-depth re-read (R10): distinguish a webhook-completed tip
+		// row (skip the side-effects — the webhook already recorded it) from
+		// a vanished/failed row (CRITICAL — money was taken at Square).
+		var curStatus string
+		rErr := recheckTx.QueryRow(r.Context(), `SELECT status FROM payments WHERE id = $1`, paymentID).Scan(&curStatus)
+		if rErr != nil || curStatus != "completed" {
+			log.Printf("CRITICAL: Square tip payment %s (ID=%s) was processed but tip row %s could not be verified as completed (re-read status=%q err=%v) — manual reconciliation required",
+				paymentResult.Status, paymentResult.SquarePayID, paymentID, curStatus, rErr)
+			http.Error(w, "internal server error", http.StatusInternalServerError)
+			return
+		}
+		log.Printf("Tip payment %s for booking %s was already resolved to %q by a concurrent resolver (Square webhook/sweep) before the sync completion flip — skipping completion side-effects", paymentID, bookingID, curStatus)
+		// MEDIUM-2: burn the re-issued 2FA code anyway (idempotent) — the
+		// charge reached terminal success, so a single-use code re-issued for
+		// this retry must not authorize another charge.
+		if req.CardID != nil && *req.CardID != "" && reusePendingRecord {
+			if consErr := twofa.ConsumePendingCode(r.Context(), recheckTx, userID); consErr != nil {
+				log.Printf("CRITICAL: Square tip payment %s (ID=%s) was processed but consuming the 2FA code for user %s failed: %v — manual reconciliation required",
+					paymentResult.Status, paymentResult.SquarePayID, userID, consErr)
+				http.Error(w, "internal server error", http.StatusInternalServerError)
+				return
+			}
+		}
+		if cErr := recheckTx.Commit(r.Context()); cErr != nil {
+			log.Printf("CRITICAL: Square tip payment %s (ID=%s) succeeded but committing the post-charge transaction for already-completed tip %s failed: %v — manual reconciliation required",
+				paymentResult.Status, paymentResult.SquarePayID, paymentID, cErr)
+			http.Error(w, "internal server error", http.StatusInternalServerError)
+			return
+		}
+		// The tip IS completed (by the webhook/sweep) — report success exactly
+		// like the normal completion path; no side-effects re-run.
+		if err := json.NewEncoder(w).Encode(PaymentResponse{
+			ID:          paymentID,
+			BookingID:   bookingID,
+			PaymentType: "tip",
+			Status:      "completed",
+			Amount:      req.Amount,
+			CardBrand:   paymentResult.CardBrand,
+			CardLast4:   paymentResult.CardLast4,
+			ReceiptURL:  paymentResult.ReceiptURL,
+			CreatedAt:   clock.Now().Format(time.RFC3339),
+		}); err != nil {
+			log.Printf("Failed to encode JSON response: %v", err)
+		}
+		return
+	}
 	// MEDIUM-2 / finding 1: a saved-card tip charge reached its terminal
 	// SUCCESS state. For a FRESH charge the gate already consumed the code
 	// (single-use at verify time); for a PENDING-REUSE retry (gate passed

backend/handlers/payments/refund_policy.go

@@ -23,7 +23,7 @@ const (
 	// deposit REQUIRED at booking time, used by bookings.go): the promotion
 	// threshold is about already-paid money, not the amount to demand up front,
 	// even though both are 20% today.
-	depositPromotionMinPct = 0.2
+	depositPromotionMinPct = RequiredDepositPct
 
 	LoyaltyStampCost       = 10
 	LoyaltyDiscountPercent = 10.0

backend/handlers/payments/refunds.go

@@ -578,6 +578,44 @@ func ProcessCancellationRefundTx(
 				}
 				break
 			}
+			// A REDEEMED card can no longer be used — the terminal path rejects
+			// redeemed cards and RedeemGiftCard refuses a second redemption — so
+			// crediting amount_remaining back onto the card would strand the
+			// refunded money forever. Route the credit to the booking user's
+			// gift-card account balance instead: the same destination the
+			// redemption moved the card's value to (RedeemGiftCard zeroes
+			// amount_remaining and credits the balance). Guests/ownerless
+			// bookings have no balance to credit — record 'failed' for admin
+			// reconciliation, mirroring the cash branch's guest handling.
+			var redeemedBy *string
+			if err := tx.QueryRow(ctx, `SELECT redeemed_by FROM gift_cards WHERE id = $1`, *giftCardID).Scan(&redeemedBy); err != nil {
+				// Fail-open like the expiry check below: a failed read cannot
+				// prove the card is redeemed, so the card-credit path proceeds
+				// (a deleted card is still caught by the M2 RowsAffected check).
+				log.Printf("Failed to check gift card %s redemption status: %v — proceeding with card credit", *giftCardID, err)
+			} else if redeemedBy != nil {
+				if isGuest || bookingUserID == "" {
+					if isGuest {
+						log.Printf("Guest giftcard refund on redeemed card %s: booking %s, payment %s, amount £%.2f — no balance credit, recorded 'failed' (admin must hand out cash at till)", *giftCardID, bookingID, paymentID, refundThisPayment)
+					} else {
+						log.Printf("Gift card %s is redeemed and payment %s has no booking user — no balance credit, recorded 'failed' (admin reconciliation)", *giftCardID, paymentID)
+					}
+					creditFailed = true
+					break
+				}
+				log.Printf("Gift card %s is redeemed — crediting £%.2f refund to user %s gift-card balance instead", *giftCardID, refundThisPayment, bookingUserID)
+				if _, balErr := tx.Exec(ctx, `
+				INSERT INTO user_giftcard_balances (user_id, balance, updated_at)
+				VALUES ($1, $2, NOW())
+				ON CONFLICT (user_id) DO UPDATE SET
+					balance = user_giftcard_balances.balance + EXCLUDED.balance,
+					updated_at = NOW()
+			`, bookingUserID, refundThisPayment); balErr != nil {
+					log.Printf("Failed to credit user %s gift-card balance for refund of booking %s: %v", bookingUserID, bookingID, balErr)
+					creditFailed = true
+				}
+				break
+			}
 			// expiry_date is maintained by EVERY gift-card write that counts as
 			// a "use" (balance check, top-up, transfer, redeem, payment, refund
 			// credit), so a non-NULL expiry_date means the rolling timer is
@@ -633,16 +671,20 @@ func ProcessCancellationRefundTx(
 		case "cash":
 			if isGuest || bookingUserID == "" {
 				if isGuest {
-					log.Printf("Guest cash refund: booking %s, payment %s, amount £%.2f — admin must process cash refund at till", bookingID, paymentID, refundThisPayment)
+					log.Printf("Guest cash refund: booking %s, payment %s, amount £%.2f — no balance credit, recorded 'failed' (admin must hand out cash at till)", bookingID, paymentID, refundThisPayment)
 				} else {
-					log.Printf("Cash refund: payment %s has no booking user — skipping balance credit", paymentID)
-				}
-				if bookingUserLookupFailed {
-					// The booking-user lookup errored, so this may be a real
-					// user whose balance credit was skipped — a 'completed'
-					// refund would claim money was returned when it wasn't.
-					creditFailed = true
+					log.Printf("Cash refund: payment %s has no booking user — no balance credit, recorded 'failed' (admin reconciliation)", paymentID)
 				}
+				// No balance can be credited for a guest/ownerless booking, so
+				// the money never moves here. Recording the row 'completed'
+				// (the pre-fix behaviour) would let the over-refund guard
+				// (GetAlreadyRefundedAmount counts completed) permanently block
+				// re-issuance if the admin never hands out the cash. Record
+				// 'failed' instead: the post-commit pre-pass in
+				// ProcessPendingSquareRefunds surfaces a 'refund_failed' admin
+				// notification so the pending payout stays visible, and the
+				// guard stays open for a re-issued attempt.
+				creditFailed = true
 			} else {
 				log.Printf("Crediting £%.2f to user %s balance for cash payment %s", refundThisPayment, bookingUserID, paymentID)
 				if _, balErr := tx.Exec(ctx, `
@@ -902,6 +944,40 @@ func ProcessPendingSquareRefunds(ctx context.Context, bookingID string, reason s
 		insertRefundFailedNotifications(ctx, failedIDs)
 	}
 
+	// (a2) Guest/ownerless cash and redeemed-gift-card cancellation refunds are
+	//     recorded 'failed' in the cancellation loop (no balance could be
+	//     credited — the money never moved) and must surface in the admin
+	//     notification centre so the cash is actually handed out / reconciled.
+	//     Runs POST-COMMIT, mirroring the Square-less pre-pass above (the rows
+	//     committed with the enclosing transaction). The (reason, booking_id)
+	//     NOT EXISTS dedup in insertRefundFailedNotifications prevents
+	//     double-notifying across re-runs.
+	cashFailedRows, err := db.Conn.Query(ctx, `
+		SELECT r.id FROM refunds r
+		JOIN payments p ON p.id = r.payment_id
+		WHERE r.booking_id = $1 AND r.status = 'failed' AND r.origin = 'cancellation'
+		  AND p.payment_method IN ('cash', 'giftcard')
+	`, bookingID)
+	if err != nil {
+		log.Printf("Failed to query failed cash/giftcard refunds for booking %s: %v", bookingID, err)
+	} else {
+		var cashFailedIDs []string
+		for cashFailedRows.Next() {
+			var id string
+			if err := cashFailedRows.Scan(&id); err == nil {
+				cashFailedIDs = append(cashFailedIDs, id)
+			}
+		}
+		if err := cashFailedRows.Err(); err != nil {
+			log.Printf("Failed to iterate failed cash/giftcard refunds for booking %s: %v", bookingID, err)
+		}
+		cashFailedRows.Close()
+		if len(cashFailedIDs) > 0 {
+			log.Printf("Surfacing %d failed cash/giftcard refund(s) for booking %s (guest/ownerless — admin must arrange payout)", len(cashFailedIDs), bookingID)
+		}
+		insertRefundFailedNotifications(ctx, cashFailedIDs)
+	}
+
 	// (b) This booking's card charges with pending refunds — one aggregated
 	//     Square refund per charge.
 	for _, chargeID := range queryChargesWithPendingRefunds(ctx, "r.booking_id = $1", bookingID) {
@@ -950,6 +1026,37 @@ func SweepPendingSquareRefunds(ctx context.Context) (int, error) {
 		insertRefundFailedNotifications(ctx, failedIDs)
 	}
 
+	// (a2) Guest/ownerless cash and redeemed-gift-card cancellation refunds
+	//     recorded 'failed' (money never moved) must surface in the admin
+	//     notification centre so the cash is actually handed out / reconciled.
+	//     Mirrors the booking-scoped pre-pass in ProcessPendingSquareRefunds;
+	//     the (reason, booking_id) NOT EXISTS dedup prevents double-notifying.
+	cashFailedRows, err := db.Conn.Query(ctx, `
+		SELECT r.id FROM refunds r
+		JOIN payments p ON p.id = r.payment_id
+		WHERE r.status = 'failed' AND r.origin = 'cancellation'
+		  AND p.payment_method IN ('cash', 'giftcard')
+	`)
+	if err != nil {
+		log.Printf("Failed to query failed cash/giftcard refunds during sweep: %v", err)
+	} else {
+		var cashFailedIDs []string
+		for cashFailedRows.Next() {
+			var id string
+			if err := cashFailedRows.Scan(&id); err == nil {
+				cashFailedIDs = append(cashFailedIDs, id)
+			}
+		}
+		if err := cashFailedRows.Err(); err != nil {
+			log.Printf("Failed to iterate failed cash/giftcard refunds during sweep: %v", err)
+		}
+		cashFailedRows.Close()
+		if len(cashFailedIDs) > 0 {
+			log.Printf("Surfacing %d failed cash/giftcard refund(s) (guest/ownerless — admin must arrange payout)", len(cashFailedIDs))
+		}
+		insertRefundFailedNotifications(ctx, cashFailedIDs)
+	}
+
 	// (b) Charges with pending refunds — one aggregated Square refund each.
 	processed := 0
 	for _, chargeID := range queryChargesWithPendingRefunds(ctx, "") {
@@ -1539,16 +1646,27 @@ func idsOf(rows []pendingChargeRow) []string {
 	return ids
 }
 
-// aggRefundKeySuffix returns a deterministic 12-hex-char suffix for a set of
+// aggRefundKeySuffix returns a deterministic hex suffix for a set of
 // identifiers. The identifiers (CHAR(12) refund ids, or an over-length
 // square_payment_id) are sorted, sha256'd and truncated, so the SAME input
 // always yields the SAME suffix. Used to compress an over-length chargeID into
 // a fixed-width prefix for the charge-level idempotency key (see chargeAggKey).
+//
+// The suffix is the full SHA-256 hex trimmed to the budget left by the
+// "-square-agg" structure: chargeAggKey appends 11 chars, so the suffix is
+// capped at maxIdempotencyKeyLength-11 = 34 chars (136 bits). The historical
+// 12-char (48-bit) truncation made a collision between two DISTINCT chargeIDs
+// plausible enough that Square's global idempotency-key dedup could silently
+// swallow the second charge's refund (lost money, no refund row); 136 bits
+// makes a cross-charge collision cryptographically negligible.
 func aggRefundKeySuffix(ids []string) string {
 	sorted := append([]string(nil), ids...)
 	sort.Strings(sorted)
 	h := sha256.Sum256([]byte(strings.Join(sorted, "")))
-	return fmt.Sprintf("%x", h)[:12]
+	// The suffix is joined with "-square-agg" by chargeAggKey; keep the whole
+	// key inside Square's 45-char idempotency-key limit.
+	const aggSuffixLen = maxIdempotencyKeyLength - len("-square-agg")
+	return fmt.Sprintf("%x", h)[:aggSuffixLen]
 }
 
 // chargeAggKey builds the charge-level idempotency key for an aggregated

backend/handlers/payments/sweep.go

@@ -237,6 +237,94 @@ type staleRow struct {
 	B1Attempts int
 }
 
+// acquireTillSaleSweepLock serializes the sweep's fail/clawback of a stale
+// till_sale on the SAME advisory lock a same-key retry pins while its Square
+// charge is mid-flight: "crussell:till:<idempotency_key>" (till.go — the retry
+// path holds it from before the Square round-trip until the row is completed).
+//
+// The race this closes (FIX 1, MAJOR): without the lock, the sweep can probe
+// Square for the charge under the key while the retry's charge has not landed
+// yet, read "no payment under the key", mark the sale failed and claw back the
+// funded gift card — then the retry's charge lands. The retry's completion
+// UPDATE (guarded on status='pending') matches 0 rows and logs CRITICAL: the
+// customer is charged at Square AND the funding is clawed back. By holding the
+// same lock before failing/clawing-back, the sweep either serializes after the
+// retry (the row is already 'completed' — the fail/clawback is a 0-row no-op)
+// or the lock is contended and the row is DEFERRED to a later pass (the retry
+// may still land its charge).
+//
+// Returns (release, true) when the lock is held — the caller MUST call
+// release() before resolving the row; (nil, false) when the lock could not be
+// acquired within the bounded try-lock (~3s) — a same-key retry is likely
+// mid-flight at Square, so the caller must SKIP the row this pass (it is
+// reconciled again next sweep) — or a pinned pool connection could not be
+// acquired. A row with NO idempotency key returns a no-op release with true:
+// no retry path can hold a lock on it, so there is nothing to serialize on.
+func acquireTillSaleSweepLock(ctx context.Context, r staleRow) (release func(), ok bool) {
+	if r.IdempotencyKey == "" {
+		return func() {}, true
+	}
+	pinConn, err := db.Conn.Acquire(ctx)
+	if err != nil {
+		log.Printf("Failed to acquire connection for till-sale sweep lock on stale row %s: %v", r.ID, err)
+		return nil, false
+	}
+	lockKey := "crussell:till:" + r.IdempotencyKey
+	acquired, lErr := acquireAdvisoryLock(ctx, pinConn, lockKey)
+	if lErr != nil {
+		log.Printf("Failed to acquire till-sale sweep lock %s for stale row %s: %v", lockKey, r.ID, lErr)
+		pinConn.Release()
+		return nil, false
+	}
+	if !acquired {
+		// A same-key retry holds the lock and its Square charge may be
+		// mid-flight — deferring avoids failing/clawing-back a sale whose
+		// charge can still land. The row is reconciled again next pass.
+		log.Printf("Stale pending till_sale %s: till-sale lock %s is contended within the bound — a same-key retry may be mid-flight at Square — deferring the row to a later sweep pass", r.ID, lockKey)
+		pinConn.Release()
+		return nil, false
+	}
+	return func() {
+		releasePaymentLock(pinConn, lockKey)
+		pinConn.Release()
+	}, true
+}
+
+// tillSalePendingAfterLock re-reads a till_sales row's CURRENT stored
+// idempotency_key and status AFTER the sweep acquired its advisory lock and
+// reports whether the fail/clawback may still proceed.
+//
+// This closes the residual window of the FIX-1 lock (BUG 2, MAJOR): a same-key
+// retry of a KEYLESS sale locks the derived BASE key (crussell:till:<base>)
+// while its slot scan may resolve a SUFFIXED final key (base-2) that becomes
+// the STORED key — the key the sweep locks (till.go). Until the retry side
+// pins the final key too, the sweep and the retry do NOT serialize, so the
+// retry can complete the sale (status → 'completed', its Square charge
+// landed) while the sweep waited on — or just acquired — its lock on the stale
+// stored key. A blind fail/clawback in that window would reverse the funding
+// of a sale whose charge is real. Re-reading the row under the lock detects
+// the retry's completion; a row that is no longer 'pending' — or whose stored
+// key changed while the sweep waited — is left alone and reconciled again
+// under its current key next sweep.
+func tillSalePendingAfterLock(ctx context.Context, r staleRow) bool {
+	var key, status string
+	if err := db.Conn.QueryRow(ctx, `
+		SELECT COALESCE(idempotency_key, ''), status FROM till_sales WHERE id = $1
+	`, r.ID).Scan(&key, &status); err != nil {
+		log.Printf("Failed to re-read till sale %s after acquiring its sweep lock (%v) — skipping the fail/clawback this pass; the row is reconciled next sweep", r.ID, err)
+		return false
+	}
+	if status != "pending" {
+		log.Printf("Stale pending till sale %s is now %q (a same-key retry completed or failed it while the sweep waited) — skipping the fail/clawback; the row is reconciled next sweep", r.ID, status)
+		return false
+	}
+	if key != r.IdempotencyKey {
+		log.Printf("Stale pending till sale %s's stored idempotency key changed from %q to %q while the sweep waited (a same-key retry re-scanned its slot) — skipping the fail/clawback on the stale key; the row is reconciled next sweep under its current key", r.ID, r.IdempotencyKey, key)
+		return false
+	}
+	return true
+}
+
 // sweepStaleRows resolves the stale pending rows of one table. Rows with a
 // square_payment_id are reconciled at Square first (COMPLETED → 'completed',
 // anything else → 'failed' exactly as the legacy bulk UPDATE did); rows
@@ -312,11 +400,39 @@ func sweepStaleRows(ctx context.Context, table string, cutoff time.Time) (resolv
 		// completed (NOT_FOUND / non-COMPLETED) claws back the funded gift
 		// card atomically with the failed mark (HIGH-1). The blind-fail path
 		// (no square_payment_id — the charge outcome is unknown) NEVER claws
-		// back: the money may have landed at Square.
-		if table == "till_sales" && r.SquarePaymentID != "" && r.HasGiftCard {
-			if clawbackTillSaleFunding(ctx, r) {
-				resolved++
+		// back: the money may have landed at Square. FIX 1: a till_sale is
+		// resolved under its own till-sale advisory lock so the sweep cannot
+		// race a late same-key retry whose Square charge is mid-flight.
+		if table == "till_sales" {
+			release, lockOK := acquireTillSaleSweepLock(ctx, r)
+			if !lockOK {
+				continue
 			}
+			// BUG 2: the lock may not serialize against a keyless-sale retry
+			// (its lock is the derived base key, ours is the stored suffix key)
+			// — re-read the row and refuse to fail/clawback a sale a retry
+			// already completed or re-keyed while the sweep waited.
+			if !tillSalePendingAfterLock(ctx, r) {
+				release()
+				continue
+			}
+			if r.SquarePaymentID != "" && r.HasGiftCard {
+				if clawbackTillSaleFunding(ctx, r) {
+					resolved++
+				}
+			} else if failStaleRow(ctx, table, r.ID) {
+				resolved++
+				// A5a: blind-failing a row with NO square_payment_id leaves the
+				// charge outcome unknown (the money may have landed at Square
+				// with a lost response), so surface it in the admin notification
+				// centre. Rows that reach this line WITH a square_payment_id
+				// were PROVED never-charged by the by-id reconcile and need no
+				// alert.
+				if r.SquarePaymentID == "" {
+					notifyStaleRowCritical(ctx, r)
+				}
+			}
+			release()
 			continue
 		}
 		if failStaleRow(ctx, table, r.ID) {
@@ -388,6 +504,26 @@ func sweepKeyedStaleRows(ctx context.Context, table string, cutoff time.Time) (r
 		// A row here is failed (NOT clawed back — the duplicate money may stand
 		// at Square) with a CRITICAL notification for manual reconciliation.
 		if hasFailedSweepDuplicateRefund(ctx, table, r.ID) || r.B1Attempts >= b1DuplicateRefundAttemptCap {
+			// FIX 1: a till_sale is failed under its own till-sale advisory
+			// lock so a same-key retry mid-flight at Square cannot land its
+			// charge onto a row the sweep just marked failed.
+			if table == "till_sales" {
+				release, lockOK := acquireTillSaleSweepLock(ctx, r)
+				if !lockOK {
+					continue
+				}
+				if !tillSalePendingAfterLock(ctx, r) {
+					release()
+					continue
+				}
+				if failStaleRow(ctx, table, r.ID) {
+					resolved++
+				}
+				notifyStaleRowCritical(ctx, r)
+				log.Printf("Stale pending %s row %s has a FAILED or attempt-capped B1 auto-refund of a replay-induced duplicate charge (b1_attempts=%d) — never re-replaying the expired key; marked failed — MANUAL RECONCILIATION REQUIRED: check Square for the duplicate charge and refund it", table, r.ID, r.B1Attempts)
+				release()
+				continue
+			}
 			if failStaleRow(ctx, table, r.ID) {
 				resolved++
 			}
@@ -403,6 +539,27 @@ func sweepKeyedStaleRows(ctx context.Context, table string, cutoff time.Time) (r
 			// A5a: this blind-fail can be hiding a real charge (the lost
 			// response the stored key was meant to reconcile), so the admin
 			// notification centre must surface it too.
+			// FIX 1: a till_sale is blind-failed under its own till-sale
+			// advisory lock so a same-key retry mid-flight at Square cannot
+			// land its charge onto a row the sweep just marked failed.
+			if table == "till_sales" {
+				release, lockOK := acquireTillSaleSweepLock(ctx, r)
+				if !lockOK {
+					continue
+				}
+				if !tillSalePendingAfterLock(ctx, r) {
+					release()
+					continue
+				}
+				if failStaleRow(ctx, table, r.ID) {
+					resolved++
+					unverifiable++
+					notifyStaleRowCritical(ctx, r)
+				}
+				release()
+				log.Printf("Stale pending %s row %s has a stored idempotency key but is already past Square's key retention window — marked failed without a replay reconcile (may have been charged with a lost response)", table, r.ID)
+				continue
+			}
 			if failStaleRow(ctx, table, r.ID) {
 				resolved++
 				unverifiable++
@@ -444,10 +601,27 @@ func sweepKeyedStaleRows(ctx context.Context, table string, cutoff time.Time) (r
 			// Square proved the charge never happened (key unknown / FAILED) —
 			// a till sale's funded gift card is clawed back atomically with the
 			// failed mark, unlike the blind-fail path where the outcome is unknown.
-			if table == "till_sales" && r.HasGiftCard {
-				if clawbackTillSaleFunding(ctx, r) {
+			// FIX 1: the fail/clawback runs under the row's till-sale advisory
+			// lock so the sweep cannot race a late same-key retry whose Square
+			// charge is mid-flight (its "no payment under the key" probe would
+			// be premature).
+			if table == "till_sales" {
+				release, lockOK := acquireTillSaleSweepLock(ctx, r)
+				if !lockOK {
+					continue
+				}
+				if !tillSalePendingAfterLock(ctx, r) {
+					release()
+					continue
+				}
+				if r.HasGiftCard {
+					if clawbackTillSaleFunding(ctx, r) {
+						resolved++
+					}
+				} else if failStaleRow(ctx, table, r.ID) {
 					resolved++
 				}
+				release()
 				continue
 			}
 			if failStaleRow(ctx, table, r.ID) {
@@ -460,7 +634,36 @@ func sweepKeyedStaleRows(ctx context.Context, table string, cutoff time.Time) (r
 			// till sale's funded gift card is NOT clawed back. Mark the parent
 			// row failed, raise the CRITICAL notification, and set b1_attempts
 			// to the cap so the expired key is never replayed (each replay
-			// would mint ANOTHER charge).
+			// would mint ANOTHER charge). FIX 1: a till_sale is failed under
+			// its own till-sale advisory lock so a same-key retry mid-flight at
+			// Square cannot land its charge onto a row the sweep just marked
+			// failed.
+			if table == "till_sales" {
+				release, lockOK := acquireTillSaleSweepLock(ctx, r)
+				if !lockOK {
+					continue
+				}
+				if !tillSalePendingAfterLock(ctx, r) {
+					release()
+					continue
+				}
+				if failStaleRow(ctx, table, r.ID) {
+					resolved++
+				}
+				notifyStaleRowCritical(ctx, r)
+				setB1AttemptsToCap(ctx, table, r.ID)
+				// Loop-B finding (MED): a capped-fail on a till_sale leaves its
+				// funded gift card outstanding — the duplicate charge at Square is
+				// REAL and a manual Square refund is the only reversal. Surface the
+				// outstanding funding with a gift_card_transactions trace so the
+				// operator sees it (see insertOutstandingGiftCardFundingTrace for
+				// the manual-resolution path).
+				if r.HasGiftCard {
+					insertOutstandingGiftCardFundingTrace(ctx, r)
+				}
+				release()
+				continue
+			}
 			if failStaleRow(ctx, table, r.ID) {
 				resolved++
 			}
@@ -759,22 +962,37 @@ func applyStaleRescueRecords(ctx context.Context, tx pgx.Tx, r staleRow, records
 	if len(records) == 0 {
 		return
 	}
+	// The align UPDATE clears the VAT fields exactly like the live post-charge
+	// path (handlers.go align UPDATE sets is_vat_applicable = FALSE, vat_rate =
+	// NULL, vat_amount = NULL, net_amount = NULL before re-applying VAT).
+	// WITHOUT the clearing (pre-fix bug) the pending row's VAT — computed at
+	// step-1 insert time on the FULL pre-split charge (handlers.go step-1
+	// ApplyVATToBookingPayment) — survived the align, and the re-apply below was
+	// a silent no-op because apply_vat_to_payment is guarded on vat_amount IS
+	// NULL (init-script.sql): the rescued primary row kept VAT on the wrong
+	// (larger) base. Clearing first makes the recompute effective on the split
+	// amount, and the SQL's payment_type != 'tip' guard leaves an all-tip
+	// rescue (the primary aligned to the carved tip record) VAT-free.
 	if _, upErr := tx.Exec(ctx, `
 		UPDATE payments SET
 			amount = $1,
 			payment_type = $2,
 			fees = $3,
+			is_vat_applicable = FALSE,
+			vat_rate = NULL,
+			vat_amount = NULL,
+			net_amount = NULL,
 			updated_at = NOW()
 		WHERE id = $4
 	`, records[0].Amount, records[0].PaymentType, records[0].Fees, r.ID); upErr != nil {
 		log.Printf("MEDIUM-3: failed to align rescued payment %s to its split primary (%v) — manual reconciliation recommended", r.ID, upErr)
 		return
 	}
-	// M5: the align UPDATE used to NULL the VAT fields, dropping a rescued
-	// charge out of VAT reporting. The synchronous post-charge path applies VAT
-	// per record (ApplyVATToBookingPayment); the sweep re-applies it after the
-	// align. apply_vat_to_payment is idempotent (guarded on vat_amount IS NULL)
-	// and skips discount/on_the_house/tip rows internally.
+	// M5: the synchronous post-charge path applies VAT per record
+	// (ApplyVATToBookingPayment); the sweep re-applies it after the align —
+	// now effective because the align cleared the stale VAT fields above.
+	// apply_vat_to_payment is idempotent (guarded on vat_amount IS NULL) and
+	// skips discount/on_the_house/tip rows internally.
 	ApplyVATToBookingPayment(ctx, tx, r.ID)
 	svc := NewPaymentService()
 	for _, rec := range records[1:] {
@@ -1092,6 +1310,26 @@ func parseReplayedCreatedAt(pr *square.PaymentResult) (time.Time, bool) {
 // seeding the Square payment at test time, and a retained-key dedup returns
 // that seeded payment). The gate mirrors the snapshot-decryption gate, which
 // also runs only in a non-dev/mock env.
+//
+// MONEY-DECISION PARITY (FIX 1, MINOR — pinned by
+// TestReplayRevealsNewCharge_DevMockGate_MoneyDecisionParity): the gate's
+// dev/mock outcome — "not a new charge", rescue the row — is IDENTICAL to what
+// the ungated prod classifier decides for every payment the dev mock can
+// actually return from ReplayPaymentByKey. The mock has NO key-retention /
+// key-expiry concept: a replay of a key already in its ledger dedups to the
+// ORIGINAL payment (created ~at row-creation time — inside the legitimate
+// window, so prod would also rescue it), and the only other COMPLETED outcome
+// is a fresh charge the mock creates AT the replay instant for an unknown key —
+// which lands within replayRescueUpperBoundSkew (5s) of replayAt and is
+// therefore still legitimate to prod (the same-key-retry tolerance). The gate
+// short-circuits ONLY the prod-branch scenarios the mock cannot produce — an
+// UNPARSEABLE created_at (prod: leave the row PENDING with a CRITICAL
+// notification) and a payment created beyond the upper-bound skew (prod: B1
+// auto-refund the provable duplicate) — so the dev sweep path deliberately
+// does NOT exercise those prod branches. If the mock ever gains a key-expiry
+// simulation that returns a payment created beyond the skew, the pin test
+// fails and this gate MUST be revisited (prod and dev would diverge on the
+// money decision for a provably-new expired-key charge).
 func replayRevealsNewCharge(r staleRow, replayAt time.Time, pr *square.PaymentResult) (newCharge bool, created time.Time, createdOK bool) {
 	if IsExplicitDevOrMockEnv() {
 		return false, time.Time{}, false
@@ -2681,6 +2919,12 @@ func recordUntrackedTillSalePayment(ctx context.Context, saleID string, pr *squa
 		log.Printf("Terminal till sale %s was already resolved — skipping untracked terminal completion record", saleID)
 		return false
 	}
+	// FIX 2: the synchronous till-completion path (GetTillCheckoutStatus) and
+	// the stale-pending rescue apply VAT via ApplyVATToTillSale; an untracked
+	// terminal till sale recorded here must do the same or the sale silently
+	// drops out of VAT reporting. The SQL function is idempotent (guarded on
+	// vat_amount IS NULL), so a run racing the poll's own apply is a no-op.
+	ApplyVATToTillSale(ctx, db.Conn, saleID)
 	log.Printf("CRITICAL: recorded untracked terminal till-sale charge %s (sale %s) from a stale checkout — sale marked completed (never polled by the frontend)", pr.SquarePayID, saleID)
 	return true
 }

backend/handlers/payments/till.go

@@ -342,6 +342,235 @@ func revertGiftCardFunding(ctx context.Context, action, giftCardID string, amoun
 	return RevertGiftCardFunding(ctx, action, giftCardID, amount, redeemToUserID, tillSaleID)
 }
 
+// recreditTillSaleFundingAfterMidFlightResolve restores a till sale's gift-card
+// funding after a concurrent clawback reverted it while a same-key retry's
+// Square charge was mid-flight (FIX 1). The retry's charge HAS landed, so the
+// customer paid for a card that was clawed back — this re-credits the funding
+// so the money is not lost, turning "charged AND clawed back" into "charged
+// AND funded". It runs BEFORE the handler fails the response; the sale row
+// itself stays 'failed' (completing it would race the writer that resolved it)
+// and is left for manual reconciliation.
+//
+// Re-credit is strictly conditional so it can never DOUBLE-fund a card:
+//   - the charge must be CONFIRMED at Square (re-query GetPayment) — a charge
+//     that did not land has nothing to fund;
+//   - the funding must actually have been CLAWED BACK — a create's card was
+//     deleted (row gone) and a topup's amount was subtracted with its
+//     transaction removed. If the funding is still intact (the sale was
+//     resolved without a clawback, e.g. a blind-fail) there is nothing to
+//     restore and re-crediting would double the value.
+//
+// Returns true when the funding was re-credited; false when the charge could
+// not be confirmed, the funding was never clawed back, or the re-credit
+// failed (CRITICAL logged for manual reconciliation).
+func recreditTillSaleFundingAfterMidFlightResolve(ctx context.Context, req TillSaleRequest, tillSaleID, giftCardID string, pr *square.PaymentResult) bool {
+	if pr == nil || pr.SquarePayID == "" {
+		return false
+	}
+	// Confirm the charge landed at Square (the retry just created it, but the
+	// re-query makes the decision authoritative rather than trusting the
+	// response alone).
+	confirm, cErr := SquareClient.GetPayment(ctx, pr.SquarePayID)
+	if cErr != nil || confirm == nil || confirm.Status != "COMPLETED" {
+		log.Printf("CRITICAL: till sale %s charge %s could not be confirmed at Square (re-query: %v) — gift-card funding NOT re-credited — MANUAL RECONCILIATION REQUIRED", tillSaleID, pr.SquarePayID, cErr)
+		return false
+	}
+	if req.Action == "create" {
+		// A create's clawback DELETED the card. If it still exists the funding
+		// was never reverted — nothing to restore (re-creating would double it).
+		var cardExists bool
+		if err := db.Conn.QueryRow(ctx, `SELECT EXISTS(SELECT 1 FROM gift_cards WHERE id = $1)`, giftCardID).Scan(&cardExists); err != nil {
+			log.Printf("Failed to check gift card %s for mid-flight funding re-credit: %v", giftCardID, err)
+			return false
+		}
+		if cardExists {
+			return false
+		}
+		return recreditClawedBackCreateFunding(ctx, req, tillSaleID, giftCardID)
+	}
+	// Topup: the clawback removed THIS sale's top-up transaction. If it still
+	// exists the funding was never reverted — nothing to restore.
+	var txExists bool
+	if err := db.Conn.QueryRow(ctx, `
+		SELECT EXISTS(SELECT 1 FROM gift_card_transactions WHERE gift_card_id = $1 AND reference_type = 'till_sale' AND reference_id = $2)
+	`, giftCardID, tillSaleID).Scan(&txExists); err != nil {
+		log.Printf("Failed to check top-up transaction for mid-flight funding re-credit: %v", err)
+		return false
+	}
+	if txExists {
+		return false
+	}
+	return recreditClawedBackTopupFunding(ctx, req, tillSaleID, giftCardID)
+}
+
+// recreditClawedBackCreateFunding re-creates the gift card a create-sale
+// clawback deleted (the charge landed after the clawback), mirroring the
+// create funding block of CreateTillSale: a fresh card for the amount, the
+// purchase transaction, the redeem-to-account credit, and the till_sales
+// item_id re-pointed at the new card so the reference stays valid. The new
+// card gets a fresh auto-generated code — the original code is gone with the
+// deleted row. All writes commit in ONE transaction so a crash cannot leave a
+// funded card with no sale linkage. Returns true on success.
+func recreditClawedBackCreateFunding(ctx context.Context, req TillSaleRequest, tillSaleID, giftCardID string) bool {
+	expiryMonths, expiryErr := GetGiftCardExpiryMonths(ctx, db.Conn)
+	if expiryErr != nil {
+		log.Printf("Failed to query gift card expiry months (using default %d): %v", defaultGiftCardExpiryMonths, expiryErr)
+		expiryMonths = defaultGiftCardExpiryMonths
+	}
+	purchaseVoucherType := "SPV"
+	if err := db.Conn.QueryRow(ctx, `SELECT COALESCE(voucher_type, 'SPV') FROM business_settings LIMIT 1`).Scan(&purchaseVoucherType); err != nil {
+		log.Printf("Failed to query voucher type: %v", err)
+	}
+	if purchaseVoucherType == "" {
+		purchaseVoucherType = "SPV"
+	}
+	purchaseVoucherType = effectiveVoucherTypeForPurchase(purchaseVoucherType)
+
+	var createdBy string
+	if err := db.Conn.QueryRow(ctx, `SELECT created_by FROM till_sales WHERE id = $1`, tillSaleID).Scan(&createdBy); err != nil {
+		log.Printf("Failed to read till sale %s created_by for funding re-credit: %v", tillSaleID, err)
+		return false
+	}
+
+	tx, err := db.Conn.Begin(ctx)
+	if err != nil {
+		log.Printf("Failed to begin funding re-credit transaction for till sale %s: %v", tillSaleID, err)
+		return false
+	}
+	defer func() {
+		if rErr := tx.Rollback(ctx); rErr != nil && !errors.Is(rErr, pgx.ErrTxClosed) {
+			log.Printf("Failed to rollback funding re-credit transaction for till sale %s: %v", tillSaleID, rErr)
+		}
+	}()
+
+	var newCardID string
+	if err := tx.QueryRow(ctx, `
+		INSERT INTO gift_cards (total_funds_added, amount_remaining, created_by, is_inventory, last_used_at, expiry_date, voucher_type_at_purchase)
+		VALUES ($1, $1, $2, FALSE, NOW(), NOW() + ($4 * INTERVAL '1 month'), $3)
+		RETURNING id
+	`, req.Amount, createdBy, purchaseVoucherType, expiryMonths).Scan(&newCardID); err != nil {
+		log.Printf("CRITICAL: failed to re-create gift card for mid-flight create clawback of till sale %s: %v — MANUAL RECONCILIATION REQUIRED", tillSaleID, err)
+		return false
+	}
+	if _, err := tx.Exec(ctx, `
+		INSERT INTO gift_card_transactions (gift_card_id, transaction_type, amount, reference_type, reference_id, user_id, notes)
+		VALUES ($1, 'purchase', $2, 'till_sale', $3, $4, NULL)
+	`, newCardID, req.Amount, tillSaleID, req.UserID); err != nil {
+		log.Printf("CRITICAL: failed to record the re-credited gift-card purchase for till sale %s: %v — MANUAL RECONCILIATION REQUIRED", tillSaleID, err)
+		return false
+	}
+	if req.RedeemToUserID != nil && *req.RedeemToUserID != "" {
+		if _, err := tx.Exec(ctx, `
+			UPDATE gift_cards
+			SET amount_remaining = 0,
+			    redeemed_at = NOW(),
+			    redeemed_by = $1,
+			    last_used_at = NOW(),
+			    expiry_date = NOW() + ($3 * INTERVAL '1 month')
+			WHERE id = $2
+		`, *req.RedeemToUserID, newCardID, expiryMonths); err != nil {
+			log.Printf("CRITICAL: failed to re-apply the redeem for re-credited gift card %s (till sale %s): %v — MANUAL RECONCILIATION REQUIRED", newCardID, tillSaleID, err)
+			return false
+		}
+		if _, err := tx.Exec(ctx, `
+			INSERT INTO user_giftcard_balances (user_id, balance, updated_at)
+			VALUES ($1, $2, NOW())
+			ON CONFLICT (user_id) DO UPDATE SET
+				balance = user_giftcard_balances.balance + EXCLUDED.balance,
+				updated_at = NOW()
+		`, *req.RedeemToUserID, req.Amount); err != nil {
+			log.Printf("CRITICAL: failed to re-credit the redeemed balance for re-credited gift card %s (till sale %s): %v — MANUAL RECONCILIATION REQUIRED", newCardID, tillSaleID, err)
+			return false
+		}
+	}
+	if _, err := tx.Exec(ctx, `UPDATE till_sales SET item_id = $1 WHERE id = $2`, newCardID, tillSaleID); err != nil {
+		log.Printf("CRITICAL: failed to re-point till sale %s at the re-credited gift card %s: %v — MANUAL RECONCILIATION REQUIRED", tillSaleID, newCardID, err)
+		return false
+	}
+	if err := tx.Commit(ctx); err != nil {
+		log.Printf("CRITICAL: failed to commit the gift-card funding re-credit for till sale %s: %v — MANUAL RECONCILIATION REQUIRED", tillSaleID, err)
+		return false
+	}
+	log.Printf("Re-credited gift-card funding for till sale %s: re-created card %s (create %s after a mid-flight clawback — original card %s was deleted)", tillSaleID, newCardID, req.Action, giftCardID)
+	return true
+}
+
+// recreditClawedBackTopupFunding re-adds the amount a topup-sale clawback
+// subtracted, mirroring the top-up funding block of CreateTillSale: the amount
+// is added back to the card's funds and this sale's top-up transaction is
+// re-inserted. All writes commit in ONE transaction. Returns true on success.
+func recreditClawedBackTopupFunding(ctx context.Context, req TillSaleRequest, tillSaleID, giftCardID string) bool {
+	// A concurrent gift-card cancellation can have removed the card entirely;
+	// then the funding cannot be re-credited.
+	var cardExists bool
+	if err := db.Conn.QueryRow(ctx, `SELECT EXISTS(SELECT 1 FROM gift_cards WHERE id = $1)`, giftCardID).Scan(&cardExists); err != nil {
+		log.Printf("Failed to check gift card %s for mid-flight funding re-credit: %v", giftCardID, err)
+		return false
+	}
+	if !cardExists {
+		log.Printf("CRITICAL: till sale %s top-up funding cannot be re-credited — gift card %s no longer exists — MANUAL RECONCILIATION REQUIRED", tillSaleID, giftCardID)
+		return false
+	}
+
+	expiryMonths, expiryErr := GetGiftCardExpiryMonths(ctx, db.Conn)
+	if expiryErr != nil {
+		log.Printf("Failed to query gift card expiry months (using default %d): %v", defaultGiftCardExpiryMonths, expiryErr)
+		expiryMonths = defaultGiftCardExpiryMonths
+	}
+
+	tx, err := db.Conn.Begin(ctx)
+	if err != nil {
+		log.Printf("Failed to begin funding re-credit transaction for till sale %s: %v", tillSaleID, err)
+		return false
+	}
+	defer func() {
+		if rErr := tx.Rollback(ctx); rErr != nil && !errors.Is(rErr, pgx.ErrTxClosed) {
+			log.Printf("Failed to rollback funding re-credit transaction for till sale %s: %v", tillSaleID, rErr)
+		}
+	}()
+
+	var isInventory bool
+	var previousTotal float64
+	if err := tx.QueryRow(ctx, `SELECT is_inventory, total_funds_added FROM gift_cards WHERE id = $1`, giftCardID).Scan(&isInventory, &previousTotal); err != nil {
+		log.Printf("CRITICAL: failed to read gift card %s for the top-up funding re-credit (till sale %s): %v — MANUAL RECONCILIATION REQUIRED", giftCardID, tillSaleID, err)
+		return false
+	}
+
+	if _, err := tx.Exec(ctx, `
+		UPDATE gift_cards
+		SET total_funds_added = total_funds_added + $1,
+		    amount_remaining = amount_remaining + $1,
+		    last_used_at = NOW(),
+		    expiry_date = NOW() + ($3 * INTERVAL '1 month')
+		WHERE id = $2
+	`, req.Amount, giftCardID, expiryMonths); err != nil {
+		log.Printf("CRITICAL: failed to re-credit the top-up amount on gift card %s (till sale %s): %v — MANUAL RECONCILIATION REQUIRED", giftCardID, tillSaleID, err)
+		return false
+	}
+
+	transactionType := "topup"
+	var notes *string
+	if isInventory && previousTotal == 0 {
+		transactionType = "purchase"
+		n := "first top-up on inventory card"
+		notes = &n
+	}
+	if _, err := tx.Exec(ctx, `
+		INSERT INTO gift_card_transactions (gift_card_id, transaction_type, amount, reference_type, reference_id, user_id, notes)
+		VALUES ($1, $2, $3, 'till_sale', $4, $5, $6)
+	`, giftCardID, transactionType, req.Amount, tillSaleID, req.UserID, notes); err != nil {
+		log.Printf("CRITICAL: failed to re-insert the top-up transaction for till sale %s (gift card %s): %v — MANUAL RECONCILIATION REQUIRED", tillSaleID, giftCardID, err)
+		return false
+	}
+
+	if err := tx.Commit(ctx); err != nil {
+		log.Printf("CRITICAL: failed to commit the gift-card funding re-credit for till sale %s: %v — MANUAL RECONCILIATION REQUIRED", tillSaleID, err)
+		return false
+	}
+	log.Printf("Re-credited gift-card funding for till sale %s: added back £%.2f on card %s (topup after a mid-flight clawback)", tillSaleID, req.Amount, giftCardID)
+	return true
+}
+
 func CreateTillSale(w http.ResponseWriter, r *http.Request) {
 	ctx := r.Context()
 	// Defense-in-depth admin check (S-1) — a till sale moves money (charges a
@@ -623,6 +852,51 @@ func CreateTillSale(w http.ResponseWriter, r *http.Request) {
 		}
 	}
 
+	// FIX 2 (F6): the sweep serializes its fail/claw-back of a stale pending
+	// till_sale on "crussell:till:<STORED idempotency_key>"
+	// (acquireTillSaleSweepLock, sweep.go) — but the advisory lock acquired
+	// above is keyed on the REQUEST's base key. A keyless sale's slot scan can
+	// advance the final stored key past the base (e.g. "till-<hash>-1" when a
+	// COMPLETED/FAILED sale occupies the base slot), and the pending-reuse
+	// resolution above can adopt a row's STORED key that differs from the
+	// request key. When they diverge, a retry would hold the base lock while
+	// the sweep holds the stored-key lock — two DIFFERENT locks that do not
+	// serialize — so the sweep could fail the sale + claw back the funded gift
+	// card while this retry's Square charge is mid-flight: customer charged AND
+	// funding clawed back. req.IdempotencyKey at this point IS the key the
+	// till_sales INSERT below stores (the FINAL key), so it must ALSO be locked
+	// before the money-critical section.
+	//
+	// The base lock is deliberately kept held: it makes the slot scan-and-insert
+	// atomic against a concurrent identical keyless create (releasing it right
+	// after the scan would let the second sale derive the same free slot, insert
+	// a duplicate and die on the idempotency_key UNIQUE constraint). The final
+	// lock is acquired in the SAME "crussell:till:" family on the SAME pinned
+	// conn, in the consistent order base → final. The sweep holds only the ONE
+	// stored-key lock and never acquires a base key, and no other code path
+	// acquires "crussell:till:" keys in reverse order, so no lock-ordering cycle
+	// is possible. When the final key equals the base key (no suffix, no
+	// adoption), the base lock already IS the final lock and nothing more is
+	// needed. A bounded try-lock matches the base lock's pool-exhaustion
+	// defence: if the sweep holds the stored key across its fail/claw-back we
+	// surface 409 "try again" instead of pinning a pool conn for its duration.
+	finalLockKey := req.IdempotencyKey
+	if finalLockKey != lockKey {
+		finalLockOK, fErr := acquireAdvisoryLock(ctx, pinConn, "crussell:till:"+finalLockKey)
+		if fErr != nil {
+			log.Printf("Failed to acquire till-sale final-key serialization lock for %s: %v", finalLockKey, fErr)
+			http.Error(w, "internal server error", http.StatusInternalServerError)
+			return
+		}
+		if !finalLockOK {
+			log.Printf("Till-sale final-key serialization lock for %s not acquired within bound — a sweep or same-key retry is in progress", finalLockKey)
+			http.Error(w, "Sale in progress, try again", http.StatusConflict)
+			return
+		}
+		defer releasePaymentLock(pinConn, "crussell:till:"+finalLockKey)
+		log.Printf("Till-sale final key %s differs from base key %s — holding both till locks across the Square round-trip so the sweep's stored-key lock serializes", finalLockKey, lockKey)
+	}
+
 	service := NewPaymentService()
 
 	// MEDIUM-5: the shared £5,000/day admin gift-card cap (giftcard_limits.go
@@ -1377,11 +1651,19 @@ func CreateTillSale(w http.ResponseWriter, r *http.Request) {
 		}
 		if tillTag.RowsAffected() == 0 {
 			// The stale-pending sweep (or a clawback) resolved the sale while
-			// the Square charge was in flight: the customer WAS charged and the
-			// gift card WAS funded, but the row no longer says 'pending'.
-			// Mirroring the non-Square pending-reuse path below, never report
-			// success for a row the DB doesn't agree on.
-			log.Printf("CRITICAL: Square payment %s succeeded but till_sale %s was already resolved (0 rows updated) — charge taken and card funded; MANUAL RECONCILIATION REQUIRED", paymentResult.SquarePayID, tillSaleID)
+			// the Square charge was in flight: the customer WAS charged, but
+			// the row no longer says 'pending' and a concurrent clawback may
+			// have reverted the funded gift card. Mirroring the non-Square
+			// pending-reuse path below, never report success for a row the DB
+			// doesn't agree on. FIX 1: this retry's charge just landed, so
+			// re-credit any clawed-back gift-card funding BEFORE failing
+			// loudly — the customer ends up "charged AND funded" instead of
+			// "charged AND clawed back".
+			if recreditTillSaleFundingAfterMidFlightResolve(ctx, req, tillSaleID, giftCardID, paymentResult) {
+				log.Printf("CRITICAL: Square payment %s succeeded but till_sale %s was already resolved (0 rows updated) — gift-card funding re-credited after the concurrent clawback; sale left failed — MANUAL RECONCILIATION REQUIRED to complete the sale record", paymentResult.SquarePayID, tillSaleID)
+			} else {
+				log.Printf("CRITICAL: Square payment %s succeeded but till_sale %s was already resolved (0 rows updated) — charge taken and gift-card funding NOT re-credited; MANUAL RECONCILIATION REQUIRED", paymentResult.SquarePayID, tillSaleID)
+			}
 			http.Error(w, "internal server error", http.StatusInternalServerError)
 			return
 		}

backend/handlers/webhooks/square.go

@@ -12,12 +12,15 @@ import (
 	"fmt"
 	"io"
 	"log"
+	"math"
 	"net/http"
 	"os"
+	"strconv"
 	"strings"
 	"sync"
 	"time"
 
+	"crussell/clock"
 	"crussell/db"
 	"crussell/handlers/payments"
 	"crussell/internal/adminnotify"
@@ -795,37 +798,58 @@ func markPaymentFailed(ctx context.Context, paymentID string) error {
 	return nil
 }
 
-// squarePaymentKnown reports whether ANY local payments row carries the Square
-// payment id, whatever its status. The pending-only UPDATE in
-// handlePaymentUpdated matches zero rows both when no row exists at all and
-// when the row already settled (e.g. a completed webhook replay of an
-// already-'completed' charge). This existence check distinguishes those two:
-// a row existing means the event is a plain no-op replay of a known charge,
-// while no row at all is the signature of an ORPHANED sweep-minted duplicate.
-// Unlike findPaymentBySquareID (which swallows errors), a DB failure here
-// propagates so the caller rejects the event and Square retries.
-func squarePaymentKnown(ctx context.Context, squarePaymentID string) (bool, error) {
-	var one int
-	err := db.Conn.QueryRow(ctx,
-		`SELECT 1 FROM payments WHERE square_payment_id = $1 LIMIT 1`, squarePaymentID).Scan(&one)
-	if errors.Is(err, pgx.ErrNoRows) {
-		return false, nil
-	}
+// squareChargeKnown reports whether ANY local row — a payments row OR a
+// till_sales row — carries the Square payment id, whatever its status. The
+// pending-only reconcile in reconcileCompletedPayments matches zero payments
+// rows both when no row exists at all and when the row already settled (e.g. a
+// completed webhook replay of an already-'completed' charge); a till_sales row
+// counts as known too because a gift-card/retail till charge has no payments
+// row at all and its completion event must still be acked (the till_sales
+// reconcile runs after). This existence check distinguishes "known — a plain
+// no-op replay" from "no row at all — the signature of an unresolved unknown
+// charge". A DB failure here propagates so the caller rejects the event and
+// Square retries.
+func squareChargeKnown(ctx context.Context, squarePaymentID string) (bool, error) {
+	var known bool
+	err := db.Conn.QueryRow(ctx, `
+		SELECT EXISTS(SELECT 1 FROM payments WHERE square_payment_id = $1)
+		    OR EXISTS(SELECT 1 FROM till_sales WHERE square_payment_id = $1)
+	`, squarePaymentID).Scan(&known)
 	if err != nil {
 		return false, err
 	}
-	return true, nil
+	return known, nil
+}
+
+// squareRefundKnown reports whether ANY local refunds row carries the Square
+// refund id, whatever its status. Used by handleRefundUpdated to distinguish a
+// plain no-op replay (row exists, already in the target status) from a refund
+// event arriving before the local row was created — the latter must NOT be
+// acked (the refund row can be created in the same transaction as the charge,
+// and acking would drop the terminal settlement events forever). A DB failure
+// propagates so the caller rejects the event and Square retries.
+func squareRefundKnown(ctx context.Context, squareRefundID string) (bool, error) {
+	var known bool
+	err := db.Conn.QueryRow(ctx, `SELECT EXISTS(SELECT 1 FROM refunds WHERE square_refund_id = $1)`, squareRefundID).Scan(&known)
+	if err != nil {
+		return false, err
+	}
+	return known, nil
 }
 
 // findPendingByOrphanKeys locates the pending ORIGIN row of a likely
-// sweep-minted duplicate charge: the pending row that shares the replayed
-// charge's idempotency key (primary — exact, because Square returns the key on
-// the Payment object and payments.idempotency_key is UNIQUE) or, failing that,
+// sweep-minted duplicate charge: the row that shares the replayed charge's
+// idempotency key (primary — exact, because Square returns the key on the
+// Payment object and payments.idempotency_key is UNIQUE) or, failing that,
 // the pending row whose booking/gift-card reference AND amount match the
 // payload's reference_id/amount_money (the sweep replays the stored snapshot
-// verbatim, preserving both). Only 'pending' rows WITHOUT a square_payment_id
-// are candidates — that is exactly the population the keyed stale-pending sweep
-// replays (sweep.go), so a match is the sweep-minted duplicate's origin.
+// verbatim, preserving both). 'pending' rows WITHOUT a square_payment_id are
+// the population the keyed stale-pending sweep replays (sweep.go), so a match
+// is the sweep-minted duplicate's origin. 'failed' rows are ALSO matched on
+// the idempotency-key path: a prior webhook delivery (or the sweep's B1
+// blind-fail) already marked the origin failed, and a re-delivery of the same
+// orphan event must find it again to ack idempotently instead of treating the
+// resolved orphan as a fresh unknown charge (round-8 fix 3).
 //
 // AGE GATE (webhook-vs-response race): the candidates are additionally limited
 // to rows old enough to have actually been replayed by the sweep
@@ -837,13 +861,13 @@ func squarePaymentKnown(ctx context.Context, squarePaymentID string) (bool, erro
 // payment. When the gate blocks a match the caller leaves the row pending
 // (critical-log), never fails it — the sweep will reconcile it later.
 func findPendingByOrphanKeys(ctx context.Context, payment squarePaymentPayload) (paymentID, bookingID string, found bool, err error) {
-	replayEligibleSince := time.Now().Add(-payments.SweepKeyedReplayAge())
+	replayEligibleSince := clock.Now().Add(-payments.SweepKeyedReplayAge())
 	if payment.IdempotencyKey != "" {
 		var pid string
 		var bid *string
 		err := db.Conn.QueryRow(ctx, `
 			SELECT id, booking_id FROM payments
-			WHERE status = 'pending' AND idempotency_key = $1 AND square_payment_id IS NULL
+			WHERE status IN ('pending', 'failed') AND idempotency_key = $1 AND square_payment_id IS NULL
 			  AND created_at <= $2
 			ORDER BY created_at DESC, id DESC
 			LIMIT 1
@@ -967,8 +991,16 @@ func detectOrphanedReplayCharge(ctx context.Context, payment squarePaymentPayloa
 		return err
 	}
 	if !found {
-		log.Printf("[SQUARE-WEBHOOK] payment.updated: COMPLETED square payment %s matches no local row and no pending origin row by idempotency key/reference_id — acknowledging (not a sweep-minted duplicate)", payment.ID)
-		return nil
+		// Round-8 fix 3: a COMPLETED payment that matches NO local row and NO
+		// pending origin row by idempotency key/reference_id is a genuinely
+		// unknown charge this server never created — possibly a duplicate minted
+		// outside the sweep's replay, or a charge whose local row was erased.
+		// Acking it (200 + dedup row) would drop the event forever, hiding money
+		// the merchant holds. Return a retryable error so the caller responds
+		// 503: Square re-delivers with backoff, and its retry budget bounds the
+		// retries (the stale-pending sweeps remain the eventual backstop).
+		log.Printf("[SQUARE-WEBHOOK] payment.updated: COMPLETED square payment %s matches no local row and no pending origin row by idempotency key/reference_id — returning 503 so Square retries (unresolved unknown money event)", payment.ID)
+		return fmt.Errorf("COMPLETED square payment %s matches no local row and no pending origin row — unresolved unknown money event, rejecting so Square retries", payment.ID)
 	}
 	// B1-EVIDENCE GATE: only treat the origin as a sweep-minted duplicate when
 	// the sweep actually minted+attempted to refund it (b1_attempts > 0 or a
@@ -1046,50 +1078,52 @@ func handlePaymentUpdated(data json.RawMessage) error {
 		log.Printf("[SQUARE-WEBHOOK] payment.updated: square payment %s status %q is non-terminal — no local state change", payment.ID, payment.Status)
 		return nil
 	}
-	// Only 'pending' rows are candidates for a terminal transition — the same
-	// conservative rule the stale-pending sweeps use. A webhook for an already
-	// settled row (Square fires payment.updated for ANY field change, e.g. fee
-	// recalculation on a fully-refunded charge) must never revert a terminal
-	// status like 'refunded' back to 'completed'.
-	tag, err := db.Conn.Exec(ctx,
-		`UPDATE payments SET status = $1, updated_at = NOW() WHERE square_payment_id = $2 AND status = 'pending'`,
-		localStatus, payment.ID)
-	if err != nil {
-		log.Printf("[SQUARE-WEBHOOK] Failed to update payment %s to status %s: %v", payment.ID, localStatus, err)
-		return err
-	}
-	if tag.RowsAffected() > 0 {
-		log.Printf("[SQUARE-WEBHOOK] payment.updated: square payment %s → local status %s", payment.ID, localStatus)
-	} else if localStatus == "completed" {
-		// B1-support: a COMPLETED payment.updated that matched NO pending
-		// 'payments' row may be an ORPHANED SWEEP-MINTED duplicate — the
-		// stale-pending sweep replayed a stored idempotency key against an
-		// expired key, Square landed a NEW charge whose id matches nothing
-		// locally, and this is that new charge's completion event. If NO local
-		// row carries this square_payment_id at all (not even a settled one),
-		// hunt for the pending origin row and settle it so the sweep never
-		// blind-fails or double-rescues it. A settled row existing means this
-		// is a plain no-op replay of a known charge and is left untouched.
-		known, kErr := squarePaymentKnown(ctx, payment.ID)
-		if kErr != nil {
-			log.Printf("[SQUARE-WEBHOOK] Failed to check whether square payment %s is known: %v", payment.ID, kErr)
-			return kErr
+	if localStatus == "completed" {
+		// CRITICAL (round-8): a COMPLETED charge must be gated against the
+		// booking BEFORE any pending row is promoted. reconcileCompletedPayments
+		// re-reads every pending payments row carrying this square payment id
+		// inside a transaction and, per row, re-checks the booking status FOR
+		// UPDATE: a cancelled/lapsed/no-show booking refuses the completion (row
+		// marked failed + automatic cancellation refund row + critical
+		// notification), a payable booking is completed with the split/VAT/
+		// deposit-promotion/fully-paid-completion side-effects the sweep rescue
+		// applies, and a booking-less row (gift-card purchase) is left pending
+		// for the same-key retry that delivers the card (C6). A COMPLETED event
+		// matching no local row at all is the unresolved-unknown case: the
+		// orphaned-replay detection runs (B1) and a genuinely unknown charge is
+		// NOT acked — the returned error makes the caller respond 503 so Square
+		// retries instead of the event being dropped forever.
+		if err := reconcileCompletedPayments(ctx, payment); err != nil {
+			return err
 		}
-		if !known {
-			if dErr := detectOrphanedReplayCharge(ctx, payment); dErr != nil {
-				return dErr
-			}
+	} else {
+		// A Square FAILED/CANCELED event on a still-pending row. Only 'pending'
+		// rows are candidates for a terminal transition — the same conservative
+		// rule the stale-pending sweeps use. A webhook for an already settled
+		// row (Square fires payment.updated for ANY field change, e.g. fee
+		// recalculation on a fully-refunded charge) must never revert a terminal
+		// status like 'refunded' back to 'completed'.
+		tag, err := db.Conn.Exec(ctx,
+			`UPDATE payments SET status = $1, updated_at = NOW() WHERE square_payment_id = $2 AND status = 'pending'`,
+			localStatus, payment.ID)
+		if err != nil {
+			log.Printf("[SQUARE-WEBHOOK] Failed to update payment %s to status %s: %v", payment.ID, localStatus, err)
+			return err
+		}
+		if tag.RowsAffected() > 0 {
+			log.Printf("[SQUARE-WEBHOOK] payment.updated: square payment %s → local status %s", payment.ID, localStatus)
+		}
+		// A definitively failed charge (Square FAILED/CANCELED) claws back the
+		// gift-card funding those pending sales added, exactly like the
+		// stale-pending sweep (handlers/payments/sweep.go); an ambiguous status
+		// never reaches here.
+		if localStatus == "failed" {
+			return clawbackFailedTillSales(ctx, payment.ID)
 		}
 	}
 	// A Square charge can also map to a till_sales row (online gift-card
 	// purchase, retail at the till) — reconcile those too. Same pending-only
-	// guard: never revert a terminal till-sale status. A definitively failed
-	// charge (Square FAILED/CANCELED) claws back the gift-card funding those
-	// pending sales added, exactly like the stale-pending sweep
-	// (handlers/payments/sweep.go); an ambiguous status never reaches here.
-	if localStatus == "failed" {
-		return clawbackFailedTillSales(ctx, payment.ID)
-	}
+	// guard: never revert a terminal till-sale status.
 	tsTag, err := db.Conn.Exec(ctx,
 		`UPDATE till_sales SET status = $1, updated_at = NOW() WHERE square_payment_id = $2 AND status = 'pending'`,
 		localStatus, payment.ID)
@@ -1103,6 +1137,565 @@ func handlePaymentUpdated(data json.RawMessage) error {
 	return nil
 }
 
+// webhookPendingCharge is a local pending payments row awaiting a COMPLETED
+// Square reconcile.
+type webhookPendingCharge struct {
+	id            string
+	bookingID     *string
+	amountPence   int64
+	paymentType   string
+	paymentMethod string
+	idemKey       sql.NullString
+	createdBy     *string
+}
+
+// reconcileCompletedPayments applies a COMPLETED Square payment to every local
+// pending payments row carrying that square_payment_id. It mirrors the
+// stale-pending sweep's gateStalePaymentRescueOnBooking +
+// rescueStaleRowCompletedTx (handlers/payments/sweep.go): each pending row's
+// booking is re-read FOR UPDATE inside one transaction, and
+//
+//   - a cancelled/lapsed/no-show booking refuses the completion: the row is
+//     marked FAILED (never completed — the cancellation refund path computes
+//     refunds from completed payments and would miss it, charging a customer
+//     with NO automatic refund, F3), an M2 auto-refund row for the full
+//     stranded charge is inserted (same shape and origin 'cancellation' as
+//     ProcessCancellationRefundTx / the sweep's gate) so the pending-refund
+//     sweep issues the Square refund, and a critical admin notification is
+//     raised;
+//   - a payable booking is completed exactly like the sweep rescue / live path:
+//     the row is promoted to 'completed', re-split (any overflow carved as a
+//     tip), VAT applied, a deposit-paid pending_release booking promoted to
+//     'confirmed', and a fully-paid booking completed with the
+//     loyalty/campaign side-effects (ApplyBookingCompletionSideEffects);
+//   - a booking-less row (gift-card purchase) is LEFT pending — the same-key
+//     retry delivers the card, never the webhook (C6).
+//
+// A COMPLETED event matching no local pending row at all is the unknown-event
+// case: a settled row existing (payments or till_sales) means a plain no-op
+// replay and is acked, otherwise the orphaned-replay detection runs (B1) and a
+// genuinely unknown charge returns a retryable error (the caller responds 503,
+// no dedup row, so Square re-delivers instead of the event being dropped
+// forever). A non-nil error always means the caller must NOT commit the dedup
+// row. Idempotent: every UPDATE is status='pending'-guarded and the refund row
+// uses the deterministic paymentID+"-square-"+pence key.
+func reconcileCompletedPayments(ctx context.Context, payment squarePaymentPayload) error {
+	tx, err := db.Conn.Begin(ctx)
+	if err != nil {
+		return err
+	}
+	defer func() {
+		if err := tx.Rollback(ctx); err != nil && !errors.Is(err, pgx.ErrTxClosed) {
+			log.Printf("[SQUARE-WEBHOOK] Failed to roll back completed-payment reconcile tx for square payment %s: %v", payment.ID, err)
+		}
+	}()
+
+	rows, err := tx.Query(ctx, `
+		SELECT id, booking_id, amount, payment_type, payment_method, idempotency_key, created_by
+		FROM payments
+		WHERE square_payment_id = $1 AND status = 'pending'
+		ORDER BY created_at DESC, id DESC
+	`, payment.ID)
+	if err != nil {
+		return err
+	}
+	// refusedBookings collects the bookings whose stranded charges were refused
+	// (row failed + auto-refund inserted) so their critical admin notifications
+	// can be raised AFTER the tx commits — an in-tx notification's
+	// admin_notifications FK check on the still-locked booking row would wait on
+	// our own uncommitted xmax (the sweep's gate notifies after commit for the
+	// same reason).
+	var refusedBookings []string
+	var pending []webhookPendingCharge
+	for rows.Next() {
+		var pr webhookPendingCharge
+		var amount float64
+		if err := rows.Scan(&pr.id, &pr.bookingID, &amount, &pr.paymentType, &pr.paymentMethod, &pr.idemKey, &pr.createdBy); err != nil {
+			rows.Close()
+			return err
+		}
+		pr.amountPence = int64(math.Round(amount * 100))
+		pending = append(pending, pr)
+	}
+	rows.Close()
+	if err := rows.Err(); err != nil {
+		return err
+	}
+
+	if len(pending) == 0 {
+		// No local pending payments row carries this square payment id.
+		// Commit the read-only tx before the out-of-tx existence checks.
+		if err := tx.Commit(ctx); err != nil {
+			return err
+		}
+		known, kErr := squareChargeKnown(ctx, payment.ID)
+		if kErr != nil {
+			log.Printf("[SQUARE-WEBHOOK] Failed to check whether square payment %s is known: %v", payment.ID, kErr)
+			return kErr
+		}
+		if known {
+			// A settled row (payments or till_sales) exists — a plain no-op
+			// replay of a known charge, left untouched.
+			return nil
+		}
+		// B1-support: hunt for an ORPHANED SWEEP-MINTED duplicate (the
+		// stale-pending sweep replayed a stored idempotency key against an
+		// expired key and Square landed a NEW charge with no local row). A
+		// resolved replay (origin found + B1 evidence) returns nil; a
+		// genuinely unknown payment returns a retryable error.
+		return detectOrphanedReplayCharge(ctx, payment)
+	}
+
+	for i := range pending {
+		pr := pending[i]
+		if pr.bookingID == nil {
+			// C6: a booking-less row is a gift-card purchase. The same-key
+			// retry delivers the card through the synchronous purchase path —
+			// never auto-complete it here (the charge would be recorded while
+			// the card was never minted). Leave pending; the dedup row still
+			// commits and the sync retry flips the row when it delivers.
+			log.Printf("[SQUARE-WEBHOOK] payment.updated: square payment %s COMPLETED for booking-less row %s (gift-card purchase) — leaving pending for the same-key retry to deliver the card (C6)", payment.ID, pr.id)
+			continue
+		}
+		// Re-read the booking status FOR UPDATE so the decision serializes
+		// against a concurrent cancellation (mirrors gateStalePaymentRescueOnBooking).
+		var status string
+		if err := tx.QueryRow(ctx, `SELECT status FROM bookings WHERE id = $1 FOR UPDATE`, *pr.bookingID).Scan(&status); err != nil {
+			// Booking gone or unreadable: the money state is unknown. Never
+			// complete on an unknown state — a completed payment on a vanished
+			// booking would strand the charge outside the refund system.
+			log.Printf("[SQUARE-WEBHOOK] CRITICAL: Square reports payment %s COMPLETED but re-reading booking %s status failed (%v) — leaving row %s pending — MANUAL RECONCILIATION REQUIRED", payment.ID, *pr.bookingID, err, pr.id)
+			continue
+		}
+		if !webhookBookingStatusAllowsCompleted(status) {
+			// Cancelled / lapsed / no-show booking: completing the payment
+			// would charge a customer for a booking the cancellation flow
+			// already closed, with NO automatic refund (F3). Mark the row
+			// FAILED so the same-key retry can never reuse it, and auto-create
+			// the M2 refund row for the full stranded charge. The critical
+			// notification is raised AFTER the tx commits (it cannot run inside
+			// the tx: its admin_notifications FK check on the locked booking row
+			// would wait on our own uncommitted xmax — the sweep's gate notifies
+			// after commit for the same reason, sweep.go).
+			if webhookFailStrandedCharge(ctx, tx, pr, status, payment.ID) {
+				log.Printf("[SQUARE-WEBHOOK] CRITICAL: square payment %s is COMPLETED but booking %s is %q — payment %s marked FAILED instead of completed; an automatic refund row was created for the stranded charge — verify the Square refund settles", payment.ID, *pr.bookingID, status, pr.id)
+				refusedBookings = append(refusedBookings, *pr.bookingID)
+			}
+			continue
+		}
+		// Payable booking — promote the pending row to completed and run the
+		// split/VAT/completion side-effects exactly like the sweep rescue.
+		//
+		// R9 ordering guarantee (webhook vs. the synchronous saved-card path,
+		// handlers.go CreateBookingPayment): both paths contend on the SAME
+		// booking row — this reconcile re-reads the booking FOR UPDATE above,
+		// the sync path's postChargeRecheck takes the identical lock — and the
+		// payment row's status='pending' is the single mutual-exclusion point.
+		// Exactly one of them can win the guarded flip:
+		//
+		//   - sync path commits first: this SELECT (status='pending' filter
+		//     above) finds no row at all and the event is acked as a plain
+		//     known-charge replay (squareChargeKnown); even if the row were
+		//     somehow still selected, this guarded UPDATE returns
+		//     RowsAffected()==0 and the side-effects are skipped;
+		//   - webhook wins the flip: the guarded UPDATE succeeds and the
+		//     side-effects run HERE, inside this tx; the sync path's own
+		//     status='pending'-guarded completion (added in the same fix) then
+		//     no-ops and never re-runs its side-effects.
+		//
+		// The side-effects therefore run at most once, by whichever path won the
+		// flip — never twice — so the split rows (UNIQUE idempotency_key, e.g.
+		// "<key>-split-1") can never be minted by both paths.
+		tag, err := tx.Exec(ctx, `
+			UPDATE payments SET status = 'completed', updated_at = NOW()
+			WHERE id = $1 AND status = 'pending'
+		`, pr.id)
+		if err != nil {
+			return err
+		}
+		if tag.RowsAffected() == 0 {
+			// Already resolved concurrently (the sync path won the flip) —
+			// nothing to do.
+			continue
+		}
+		log.Printf("[SQUARE-WEBHOOK] payment.updated: square payment %s → local status completed (row %s)", payment.ID, pr.id)
+		webhookApplyCompletedPaymentRecords(ctx, tx, pr, payment.ID)
+		// The webhook won the flip, so it owns the side-effects: verify they
+		// actually landed inside this tx instead of silently claiming a
+		// completion whose bookkeeping was skipped.
+		webhookVerifyCompletionSideEffects(ctx, tx, pr)
+	}
+
+	if err := tx.Commit(ctx); err != nil {
+		log.Printf("[SQUARE-WEBHOOK] Failed to commit completed-payment reconcile for square payment %s: %v", payment.ID, err)
+		return err
+	}
+	// Raise the refused-booking critical notifications AFTER the commit — the
+	// booking row lock is released now, so the notification's FK check on it
+	// cannot wait on our own transaction.
+	for _, bookingID := range refusedBookings {
+		insertCriticalPaymentNotification(ctx, bookingID, "")
+	}
+	return nil
+}
+
+// webhookBookingStatusAllowsCompleted reports whether a charge that already
+// went through Square can still be recorded as a completed payment. It mirrors
+// the payments package's bookingStatusAllowsCompletedPayment (handlers.go) —
+// the exact predicate the sweep's gate uses — replicated here because the
+// payments package is mid-edit by another agent and must not be modified. A
+// booking that legitimately completed ('completed') must still accept the
+// recorded payment; a cancelled, lapsed, or no-show booking must NOT — the
+// money would bypass the cancellation refund system.
+func webhookBookingStatusAllowsCompleted(status string) bool {
+	switch status {
+	case "confirmed", "pending", "pending_release", "in_progress", "completed":
+		return true
+	default:
+		return false
+	}
+}
+
+// webhookFailStrandedCharge marks a pending payment row FAILED on a
+// cancelled/lapsed/no-show booking whose charge already completed at Square and
+// auto-creates the M2 cancellation refund row for the full stranded charge.
+// Mirrors gateStalePaymentRescueOnBooking's refused branch (sweep.go). Returns
+// true when the row was actually refused (still pending); false when it was
+// already resolved concurrently. The critical admin notification is NOT raised
+// here — the caller raises it after the tx commits (an in-tx notification's
+// admin_notifications FK check on the locked booking row would wait on our own
+// uncommitted xmax).
+func webhookFailStrandedCharge(ctx context.Context, tx pgx.Tx, pr webhookPendingCharge, bookingStatus, squarePaymentID string) bool {
+	tag, err := tx.Exec(ctx, `
+		UPDATE payments SET status = 'failed', updated_at = NOW()
+		WHERE id = $1 AND status = 'pending'
+	`, pr.id)
+	if err != nil {
+		log.Printf("[SQUARE-WEBHOOK] CRITICAL: Square payment %s is COMPLETED but booking %s is %q — marking the row failed errored (%v) — MANUAL RECONCILIATION REQUIRED", squarePaymentID, *pr.bookingID, bookingStatus, err)
+		return false
+	}
+	if int(tag.RowsAffected()) == 0 {
+		// Already resolved concurrently — nothing left to refuse.
+		return false
+	}
+	if pr.amountPence > 0 {
+		// M2: a refund row for the full stranded charge. Same shape and origin
+		// ('cancellation') ProcessCancellationRefundTx records, so the pending
+		// Square refund sweep aggregates and issues it at Square. The
+		// deterministic key mirrors the cancellation path (paymentID + "-square-"
+		// + amount pence) and the UNIQUE conflict guard makes a re-run
+		// idempotent.
+		if _, rfErr := tx.Exec(ctx, `
+			INSERT INTO refunds (payment_id, booking_id, amount, status, reason, idempotency_key, created_by, created_at, origin)
+			VALUES ($1, $2, $3, 'pending', $4, $5, $6, NOW(), 'cancellation')
+			ON CONFLICT (idempotency_key) DO NOTHING
+		`, pr.id, *pr.bookingID, float64(pr.amountPence)/100.0,
+			"Stranded charge on cancelled booking "+bookingStatus+" — auto-refunded by webhook",
+			pr.id+"-square-"+strconv.FormatInt(pr.amountPence, 10), pr.createdBy); rfErr != nil {
+			log.Printf("[SQUARE-WEBHOOK] CRITICAL: square payment %s (row %s) is COMPLETED but creating its auto-refund row errored (%v) — MANUAL RECONCILIATION REQUIRED: refund the charge manually", squarePaymentID, pr.id, rfErr)
+		}
+	}
+	return true
+}
+
+// webhookApplyCompletedPaymentRecords runs the post-completion bookkeeping for
+// a payable-booking payment the webhook just promoted to 'completed': re-split
+// the charge (any overflow carved as its own tip record), apply VAT, promote a
+// deposit-paid pending_release booking to 'confirmed', and complete a fully-paid
+// booking with the loyalty/campaign side-effects. It mirrors the sweep rescue's
+// buildStaleRescueRecords + applyStaleRescueRecords (sweep.go) and the live
+// post-charge path (handlers.go). All work runs inside the caller's reconcile
+// transaction; a bookkeeping failure is logged and never aborts the money
+// mutation (the charge already completed at Square — never worse than the
+// pre-fix minimal flip).
+func webhookApplyCompletedPaymentRecords(ctx context.Context, tx pgx.Tx, pr webhookPendingCharge, squarePaymentID string) {
+	// A tip-type row is never re-split (a tip can never fully pay a booking).
+	if pr.paymentType == "tip" {
+		return
+	}
+	info, err := payments.NewPaymentService().GetBookingPaymentInfo(ctx, *pr.bookingID)
+	if err != nil || info == nil {
+		log.Printf("[SQUARE-WEBHOOK] CRITICAL: failed to load booking info for completion of payment %s (booking %s): %v — row completed un-split; manual reconciliation recommended", pr.id, *pr.bookingID, err)
+		return
+	}
+	amount := float64(pr.amountPence) / 100.0
+	spID := squarePaymentID
+	primary := payments.PaymentRecord{
+		BookingID:       *pr.bookingID,
+		PaymentType:     pr.paymentType,
+		PaymentMethod:   pr.paymentMethod,
+		Status:          "completed",
+		Amount:          amount,
+		SquarePaymentID: &spID,
+		CreatedBy:       pr.createdBy,
+		CreatedAt:       clock.Now(),
+		UpdatedAt:       clock.Now(),
+	}
+	if pr.idemKey.Valid {
+		k := pr.idemKey.String
+		primary.IdempotencyKey = &k
+	}
+	records, splitErr := webhookBuildSplitRecords(primary, pr.paymentType, info, amount)
+	if splitErr != nil {
+		log.Printf("[SQUARE-WEBHOOK] CRITICAL: buildSplitRecords rejected the completion split for payment %s (booking %s): %v — row completed un-split; manual reconciliation recommended", pr.id, *pr.bookingID, splitErr)
+		return
+	}
+	if len(records) == 0 {
+		return
+	}
+	// Align the primary row to records[0] (deposit/balance/full portion).
+	if _, upErr := tx.Exec(ctx, `
+		UPDATE payments SET
+			amount = $1,
+			payment_type = $2,
+			fees = $3,
+			updated_at = NOW()
+		WHERE id = $4
+	`, records[0].Amount, records[0].PaymentType, records[0].Fees, pr.id); upErr != nil {
+		log.Printf("[SQUARE-WEBHOOK] CRITICAL: failed to align completed payment %s to its split primary (%v) — manual reconciliation recommended", pr.id, upErr)
+		return
+	}
+	// apply_vat_to_payment is idempotent (guarded on vat_amount IS NULL) and
+	// skips discount/on_the_house/tip rows internally.
+	payments.ApplyVATToBookingPayment(ctx, tx, pr.id)
+	svc := payments.NewPaymentService()
+	for _, rec := range records[1:] {
+		pid, cErr := svc.CreatePaymentRecordTx(ctx, tx, rec, nil)
+		if cErr != nil {
+			log.Printf("[SQUARE-WEBHOOK] CRITICAL: failed to insert completion split record for payment %s (%v) — manual reconciliation recommended", pr.id, cErr)
+			return
+		}
+		payments.ApplyVATToBookingPayment(ctx, tx, pid)
+	}
+	webhookPromoteAndCompleteBooking(ctx, tx, *pr.bookingID)
+}
+
+// webhookVerifyCompletionSideEffects is the R9 belt-and-braces backstop: once
+// the webhook wins the guarded flip it OWNS the post-completion bookkeeping, so
+// before the reconcile commits it re-checks, inside the same tx, that a
+// fully-paid booking actually ended up 'completed'. Every failure inside
+// webhookApplyCompletedPaymentRecords already raises its own CRITICAL log, but
+// this cross-check means a skipped side-effect chain can never be silently
+// claimed as a completion. Log-only and never aborting: the charge already
+// completed at Square, so the money mutation stands either way and the operator
+// is told to reconcile.
+func webhookVerifyCompletionSideEffects(ctx context.Context, tx pgx.Tx, pr webhookPendingCharge) {
+	var status string
+	if err := tx.QueryRow(ctx, `SELECT status FROM bookings WHERE id = $1`, *pr.bookingID).Scan(&status); err != nil {
+		log.Printf("[SQUARE-WEBHOOK] CRITICAL: payment %s completed on booking %s but verifying the booking state failed (%v) — MANUAL RECONCILIATION REQUIRED", pr.id, *pr.bookingID, err)
+		return
+	}
+	if status != "completed" && webhookBookingIsFullyPaid(ctx, tx, *pr.bookingID) {
+		log.Printf("[SQUARE-WEBHOOK] CRITICAL: payment %s made booking %s fully paid but the booking is %q, not 'completed' — the completion side-effects were skipped — MANUAL RECONCILIATION REQUIRED", pr.id, *pr.bookingID, status)
+	}
+}
+
+// webhookBuildSplitRecords partitions a completed webhook-reconciled charge
+// into its deposit/balance/tip payment records, mirroring the payments
+// package's buildSplitRecords (handlers.go) — the exact split the sweep rescue
+// and the live post-charge path apply. The records always partition
+// paymentAmount exactly (booking portion + any tip). Kept in sync by hand: the
+// payments package is mid-edit by another agent and must not be modified.
+func webhookBuildSplitRecords(primary payments.PaymentRecord, reqPaymentType string, info *payments.BookingPaymentInfo, paymentAmount float64) ([]payments.PaymentRecord, error) {
+	// After the booking starts there is no deposit protection window, but an
+	// overpayment beyond the remaining booking value is still gratuity and must
+	// be carved out as its own payment_type='tip' record (F3) — mirroring
+	// buildTerminalSplitRecords' post-start carve.
+	if clock.Now().After(info.StartTime) {
+		remaining := math.Max(0, info.TotalAmount-info.TotalPaid)
+		bookingPortion := math.Min(paymentAmount, remaining)
+		bookingPortion = math.Round(bookingPortion*100) / 100
+		tipPortion := math.Round((paymentAmount-bookingPortion)*100) / 100
+		if tipPortion > 0.004 { // payments.roundingEpsilon (errors.go)
+			// Belt-and-braces: the online tip bound (payments.maxOnlineTipPence,
+			// £250) applies to every tip row. Returning an error (never
+			// clamping) preserves the partition invariant.
+			if tipPence := int64(math.Round(tipPortion * 100)); tipPence > 25000 {
+				return nil, fmt.Errorf("webhookBuildSplitRecords: post-start carve for booking %s would mint a tip of %d pence, exceeding the £250 online tip cap (charge %.2f)", primary.BookingID, tipPence, paymentAmount)
+			}
+			records := []payments.PaymentRecord{primary}
+			records[0].Amount = bookingPortion
+			tip := primary
+			tip.PaymentType = "tip"
+			tip.Amount = tipPortion
+			tip.Fees = 0
+			if primary.IdempotencyKey != nil {
+				k := *primary.IdempotencyKey + "-split-tip"
+				tip.IdempotencyKey = &k
+			}
+			return append(records, tip), nil
+		}
+		return []payments.PaymentRecord{primary}, nil
+	}
+
+	// Deposit portion: up to 50% of total, minus what's already been paid.
+	maxDeposit := info.TotalAmount * payments.ProtectedDepositMaxPct
+	remainingDepositRoom := math.Max(0, maxDeposit-info.TotalPaid)
+	depositAmount := math.Min(paymentAmount, remainingDepositRoom)
+	depositAmount = math.Round(depositAmount*100) / 100
+
+	// Balance portion: covers whatever is still owed on the booking.
+	remainingAfterDeposit := math.Round((paymentAmount-depositAmount)*100) / 100
+	bookingRemaining := math.Max(0, info.TotalAmount-info.TotalPaid-depositAmount)
+	balancePortion := math.Min(remainingAfterDeposit, bookingRemaining)
+	balancePortion = math.Round(balancePortion*100) / 100
+
+	// Tip: anything beyond the booking total.
+	tipPortion := math.Round((remainingAfterDeposit-balancePortion)*100) / 100
+	if tipPortion > 0.004 {
+		if tipPence := int64(math.Round(tipPortion * 100)); tipPence > 25000 {
+			return nil, fmt.Errorf("webhookBuildSplitRecords: pre-start carve for booking %s would mint a tip of %d pence, exceeding the £250 online tip cap (charge %.2f)", primary.BookingID, tipPence, paymentAmount)
+		}
+	}
+
+	var records []payments.PaymentRecord
+	splitIdx := 0
+
+	if depositAmount > 0.004 {
+		dep := primary
+		dep.PaymentType = "deposit"
+		dep.Amount = depositAmount
+		records = append(records, dep)
+		splitIdx++
+	}
+
+	if balancePortion > 0.004 {
+		bal := primary
+		bal.Amount = balancePortion
+		bal.Fees = 0
+		if primary.IdempotencyKey != nil {
+			k := *primary.IdempotencyKey + fmt.Sprintf("-split-%d", splitIdx)
+			bal.IdempotencyKey = &k
+		}
+		totalPaidAfterBalance := info.TotalPaid + depositAmount + balancePortion
+		switch {
+		case totalPaidAfterBalance >= info.TotalAmount && totalPaidAfterBalance-balancePortion > 0:
+			bal.PaymentType = "balance"
+		case totalPaidAfterBalance >= info.TotalAmount:
+			bal.PaymentType = "full"
+		default:
+			bal.PaymentType = "partial"
+		}
+		records = append(records, bal)
+		splitIdx++
+	}
+
+	if tipPortion > 0.004 {
+		tip := primary
+		tip.PaymentType = "tip"
+		tip.Amount = tipPortion
+		tip.Fees = 0
+		splitIdx++
+		if primary.IdempotencyKey != nil {
+			k := *primary.IdempotencyKey + fmt.Sprintf("-split-%d", splitIdx)
+			tip.IdempotencyKey = &k
+		}
+		records = append(records, tip)
+	}
+
+	// Defensive fallback: nothing was appended (deposit, balance, AND tip all
+	// zero — impossible given paymentAmount is validated > 0 upstream, so this
+	// is a pure safety net).
+	if len(records) == 0 {
+		primary.Fees = 0
+		records = append(records, primary)
+	}
+	return records, nil
+}
+
+// webhookPromoteAndCompleteBooking runs the booking-state transitions the live
+// post-charge path applies after a payment is recorded (handlers.go): a
+// deposit-paid pending_release booking is promoted to 'confirmed' (payment
+// covers >= 20% of the total), and a fully-paid booking is completed with the
+// loyalty/campaign/name_history side-effects (completeActiveBookingFromPayment +
+// ApplyBookingCompletionSideEffects). Runs inside the caller's reconcile tx so
+// the state changes and the payment flip commit atomically.
+func webhookPromoteAndCompleteBooking(ctx context.Context, tx pgx.Tx, bookingID string) {
+	// Deposit promotion: paid >= 20% of the booking total promotes a
+	// pending_release booking to confirmed (handlers.go depositPromotionMinPct).
+	var depositMet bool
+	if err := tx.QueryRow(ctx, fmt.Sprintf(`
+		WITH booking_total AS (
+			SELECT total_amount * 100 AS total_pence FROM bookings WHERE id = $1
+		),
+		paid_total AS (
+			SELECT COALESCE(SUM(amount), 0) * 100 AS paid_pence
+			FROM payments
+			WHERE booking_id = $1 AND status = 'completed'
+			  AND payment_type != 'tip'
+			  AND payment_method NOT IN ('discount', 'on_the_house')
+		)
+		SELECT pt.paid_pence >= ROUND(bt.total_pence * %f)
+		FROM booking_total bt, paid_total pt
+	`, 0.2), bookingID).Scan(&depositMet); err != nil {
+		log.Printf("[SQUARE-WEBHOOK] Failed to check deposit threshold for booking %s: %v", bookingID, err)
+	}
+	if depositMet {
+		if _, err := tx.Exec(ctx, `
+			UPDATE bookings SET status = 'confirmed', updated_at = NOW()
+			WHERE id = $1 AND status = 'pending_release'
+		`, bookingID); err != nil {
+			log.Printf("[SQUARE-WEBHOOK] ALERT: payment completed but failed to promote booking %s from pending_release: %v", bookingID, err)
+		}
+	}
+
+	if webhookBookingIsFullyPaid(ctx, tx, bookingID) {
+		webhookCompleteActiveBooking(ctx, tx, bookingID)
+	}
+}
+
+// webhookBookingIsFullyPaid reports whether completed payments toward the
+// booking (excluding tips and on-the-house rows, but INCLUDING discount rows)
+// cover 100% of the booking total. Mirrors the payments package's
+// bookingIsFullyPaid (completion.go).
+func webhookBookingIsFullyPaid(ctx context.Context, q db.Querier, bookingID string) bool {
+	var fullyPaid bool
+	if err := q.QueryRow(ctx, `
+		WITH booking_total AS (
+			SELECT total_amount * 100 AS total_pence FROM bookings WHERE id = $1
+		),
+		paid_total AS (
+			SELECT COALESCE(SUM(amount), 0) * 100 AS paid_pence
+			FROM payments
+			WHERE booking_id = $1 AND status = 'completed'
+			  AND payment_type != 'tip'
+			  AND payment_method NOT IN ('on_the_house')
+		)
+		SELECT pt.paid_pence >= bt.total_pence AND bt.total_pence > 0
+		FROM booking_total bt, paid_total pt
+	`, bookingID).Scan(&fullyPaid); err != nil {
+		log.Printf("[SQUARE-WEBHOOK] Failed to check full-payment threshold for booking %s: %v", bookingID, err)
+	}
+	return fullyPaid
+}
+
+// webhookCompleteActiveBooking transitions an active booking to 'completed' and
+// runs the completion side-effects, all within tx. It is a no-op if the booking
+// is not in an active (completable) status, so cancelled, no-show and
+// deposit-lapsed bookings are never auto-completed — and once completed it can
+// never re-fire, because the status filter no longer matches. Mirrors
+// completeActiveBookingFromPayment (completion.go).
+func webhookCompleteActiveBooking(ctx context.Context, tx pgx.Tx, bookingID string) {
+	var completedID string
+	err := tx.QueryRow(ctx, `
+		UPDATE bookings SET status = 'completed', updated_at = NOW()
+		WHERE id = $1 AND status IN ('pending', 'confirmed', 'in_progress', 'pending_release')
+		RETURNING id
+	`, bookingID).Scan(&completedID)
+	if err != nil {
+		if !errors.Is(err, pgx.ErrNoRows) {
+			log.Printf("[SQUARE-WEBHOOK] ALERT: failed to complete fully-paid booking %s: %v", bookingID, err)
+		}
+		return
+	}
+	var userID string
+	if uErr := tx.QueryRow(ctx, `SELECT user_id FROM bookings WHERE id = $1`, bookingID).Scan(&userID); uErr != nil {
+		log.Printf("[SQUARE-WEBHOOK] ALERT: booking %s completed by payment but failed to load user for side-effects: %v", bookingID, uErr)
+		return
+	}
+	payments.ApplyBookingCompletionSideEffects(ctx, tx, bookingID, userID)
+}
+
 // clawbackFailedTillSales reverts the gift-card funding of every still-pending
 // till sale funded by a Square charge that is DEFINITIVELY failed (Square
 // FAILED/CANCELED — never ambiguous). It mirrors the stale-pending sweep's
@@ -1233,6 +1826,21 @@ func handleRefundUpdated(data json.RawMessage) error {
 	// money that already moved). Leave the row 'pending'; the Square status is
 	// logged for the audit trail.
 	if refund.Status == "APPROVED" {
+		// Round-8 fix 4: an APPROVED refund arriving BEFORE the local refund row
+		// exists must not be acked-and-dropped — the refund row can be created
+		// in the same transaction as the charge in some paths, and acking here
+		// would lose the terminal COMPLETED/FAILED settlement events forever (no
+		// sweep fallback re-discovers an APPROVED-only trail). Return 503 so
+		// Square re-delivers once the row exists.
+		known, kErr := squareRefundKnown(ctx, refund.ID)
+		if kErr != nil {
+			log.Printf("[SQUARE-WEBHOOK] Failed to check whether square refund %s is known: %v", refund.ID, kErr)
+			return kErr
+		}
+		if !known {
+			log.Printf("[SQUARE-WEBHOOK] refund.updated: square refund %s status APPROVED arrived with NO local refund row yet — returning 503 so Square retries (the refund row may be inserted in the same transaction as the charge)", refund.ID)
+			return fmt.Errorf("refund.updated APPROVED for square refund %s with no local refund row yet — rejecting so Square retries", refund.ID)
+		}
 		log.Printf("[SQUARE-WEBHOOK] refund.updated: square refund %s status APPROVED is non-terminal for the webhook (Square may still settle COMPLETED or FAILED) — leaving the local row pending", refund.ID)
 		return nil
 	}
@@ -1255,6 +1863,22 @@ func handleRefundUpdated(data json.RawMessage) error {
 		}
 		if tag.RowsAffected() > 0 {
 			log.Printf("[SQUARE-WEBHOOK] refund.updated: square refund %s → local status %s", refund.ID, localStatus)
+		} else {
+			// Zero rows: the refund is already 'completed' (a plain no-op
+			// replay) OR the local row does not exist yet. Round-8 fix 4: a
+			// genuinely absent row must not be acked — the settlement would be
+			// dropped forever. Return 503 so Square re-delivers once the row
+			// appears (the refund row can be created in the same transaction as
+			// the charge in some paths).
+			known, kErr := squareRefundKnown(ctx, refund.ID)
+			if kErr != nil {
+				log.Printf("[SQUARE-WEBHOOK] Failed to check whether square refund %s is known: %v", refund.ID, kErr)
+				return kErr
+			}
+			if !known {
+				log.Printf("[SQUARE-WEBHOOK] refund.updated: square refund %s status COMPLETED arrived with NO local refund row yet — returning 503 so Square retries (the refund row may be inserted in the same transaction as the charge)", refund.ID)
+				return fmt.Errorf("refund.updated COMPLETED for square refund %s with no local refund row yet — rejecting so Square retries", refund.ID)
+			}
 		}
 		// B1 sweep-dup refunds: a COMPLETED promotion must ALSO resolve the
 		// parent payment/till_sale row. The B1 re-poll pass (sweepPendingB1Refunds,