feat(bank-integration): fuzzy purpose matching in BRT Create & Reconcile

_find_mapping_for_txn now tries exact-substring purpose matches first
(unchanged), then falls back to a partial-ratio fuzzy match against
similarity_threshold_purpose — so a short/imprecise purpose keyword can
match somewhere inside a long bank-statement description. Order:
purpose+counterparty (exact) > purpose-only (exact) > purpose+
counterparty (fuzzy) > purpose-only (fuzzy) > counterparty-only >
fallback. Set the threshold to 100 to keep the old substring-only
behaviour. This makes similarity_threshold_purpose actually do something
in the BRT path.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

jey_erp/bank_integration/create_reconcile.py

@@ -9,6 +9,7 @@ Bank Integration source. Three modes: "Both" / "Mappings Only" /
 """
 
 import json
+from difflib import SequenceMatcher
 
 import frappe
 from frappe.utils import cint, flt
@@ -16,6 +17,27 @@ from frappe.utils import cint, flt
 PURPOSE_DOCTYPE = "Bank Integration Purpose"
 
 
+def _partial_ratio(needle, haystack):
+	"""Best fuzzy ratio of `needle` against any equal-length window of `haystack`
+	(like fuzzywuzzy.partial_ratio). Lets a short keyword fuzzily match somewhere
+	inside a long bank-statement description. Inputs should be lowercased/stripped."""
+	needle = (needle or "").strip()
+	haystack = (haystack or "").strip()
+	if not needle or not haystack:
+		return 0.0
+	if len(needle) >= len(haystack):
+		return SequenceMatcher(None, needle, haystack).ratio()
+	best = 0.0
+	span = len(needle)
+	for i in range(0, len(haystack) - span + 1):
+		r = SequenceMatcher(None, needle, haystack[i:i + span]).ratio()
+		if r > best:
+			best = r
+			if best >= 0.9999:
+				break
+	return best
+
+
 def _extract_messages(message_log):
 	parts = []
 	for m in message_log:
@@ -150,6 +172,7 @@ def create_purpose_mappings(transactions, bank_integration, paid_from=None, paid
 				frappe.db.commit()
 
 		if do_documents:
+			purpose_threshold = flt(settings.similarity_threshold_purpose or 70) / 100.0
 			if mode == "Documents & Reconcile":
 				purpose_text_cache = {}
 				for row in settings.transaction_mappings:
@@ -170,7 +193,7 @@ def create_purpose_mappings(transactions, bank_integration, paid_from=None, paid
 					continue
 
 				if mode == "Documents & Reconcile":
-					mapping_row = _find_mapping_for_txn(txn, txn_mappings, purpose_text_cache)
+					mapping_row = _find_mapping_for_txn(txn, txn_mappings, purpose_text_cache, purpose_threshold)
 					if not mapping_row:
 						errors.append({
 							"reference_number": txn.get("reference_number") or "",
@@ -272,10 +295,13 @@ def create_purpose_mappings(transactions, bank_integration, paid_from=None, paid
 		return {"success": False, "message": str(e)}
 
 
-def _find_mapping_for_txn(txn, transaction_mappings, purpose_text_cache):
+def _find_mapping_for_txn(txn, transaction_mappings, purpose_text_cache, purpose_threshold):
 	"""Find the best matching transaction mapping row for a BRT txn.
 
-	Priority: rows matching both purpose+counterparty > purpose-only > counterparty-only > fallback.
+	Priority: purpose+counterparty (exact) > purpose-only (exact) >
+	purpose+counterparty (fuzzy) > purpose-only (fuzzy) > counterparty-only >
+	fallback by counterparty_type. "Exact" = the keyword is a substring of the
+	transaction's description; "fuzzy" = partial-ratio ≥ purpose_threshold.
 	"""
 	payment_type = "Pay" if txn.get("drcr") == "D" else "Receive"
 	txn_purpose = (txn.get("purpose") or "").lower()
@@ -310,25 +336,53 @@ def _find_mapping_for_txn(txn, transaction_mappings, purpose_text_cache):
 		elif row.counterparty_type:
 			fallback_rules.append(row)
 
-	def _purpose_matches(row):
-		keyword_text = purpose_text_cache.get(row.purpose_keyword, "")
-		return keyword_text and keyword_text in txn_purpose
+	def _kw(row):
+		return purpose_text_cache.get(row.purpose_keyword, "")
+
+	def _purpose_exact(row):
+		kw = _kw(row)
+		return bool(kw) and kw in txn_purpose
+
+	def _purpose_fuzzy(row):
+		kw = _kw(row)
+		return _partial_ratio(kw, txn_purpose) if kw else 0.0
 
 	def _counterparty_matches(row):
 		return row.counterparty.lower() == txn_party
 
+	def _best_fuzzy(rules, check_cp):
+		best = None
+		best_score = 0.0
+		for row in rules:
+			if check_cp and not _counterparty_matches(row):
+				continue
+			score = _purpose_fuzzy(row)
+			if score >= purpose_threshold and score > best_score:
+				best_score = score
+				best = row
+		return best
+
+	# 1. purpose + counterparty — exact
 	for row in both_rules:
-		if _purpose_matches(row) and _counterparty_matches(row):
+		if _purpose_exact(row) and _counterparty_matches(row):
 			return row
-
+	# 2. purpose-only — exact
 	for row in purpose_only_rules:
-		if _purpose_matches(row):
+		if _purpose_exact(row):
 			return row
-
+	# 3. purpose + counterparty — fuzzy
+	hit = _best_fuzzy(both_rules, check_cp=True)
+	if hit:
+		return hit
+	# 4. purpose-only — fuzzy
+	hit = _best_fuzzy(purpose_only_rules, check_cp=False)
+	if hit:
+		return hit
+	# 5. counterparty-only
 	for row in counterparty_only_rules:
 		if _counterparty_matches(row):
 			return row
-
+	# 6. fallback by counterparty_type
 	inferred_cp_type = "Customer" if payment_type == "Receive" else "Supplier"
 	for row in fallback_rules:
 		if row.counterparty_type == inferred_cp_type: