finance-duck/internal/classification/ledger_test.go

package classification

import (
	"context"
	"fmt"
	"net/http"
	"reflect"
	"strings"
	"testing"

	"finance-duck/internal/domain"
)

// ledgerFixture mirrors a production ledger that repeatedly broke
// classification in the field: a proposed two-level taxonomy (43 expense
// leaves), tags, a merchant registry polluted with location-like names, and
// German bank rows whose payee text carries reference numbers. Personal
// names and IBANs are fabricated.
func ledgerFixture() (domain.Dataset, domain.Facts) {
	d := domain.NewDataset()
	d.Accounts = []domain.Account{{ID: "acct_kontist", DisplayName: "Business", Institution: "Kontist", Currency: "EUR", Active: true}}
	tree := map[string][]string{
		"housing":       {"rent", "utilities", "household", "maintenance"},
		"food":          {"groceries", "restaurants", "takeaway"},
		"transport":     {"fuel", "public-transport", "parking", "taxi", "vehicle-maintenance"},
		"shopping":      {"clothing", "electronics", "household-goods", "other"},
		"pets":          {"pet-food", "pet-health", "supplies"},
		"entertainment": {"games", "events", "ent-media"},
		"travel":        {"accommodation", "travel-transport", "activities"},
		"health":        {"medical", "pharmacy", "fitness"},
		"education":     {"tuition", "books", "courses"},
		"subscriptions": {"software", "sub-media", "services"},
		"insurance":     {"vehicle-insurance", "health-insurance", "other-insurance"},
		"financial":     {"bank-fees", "interest-paid", "taxes"},
		"gifts":         nil,
		"donations":     nil,
	}
	for parent, children := range tree {
		d.Categories = append(d.Categories, domain.Category{ID: "cat_" + parent, Name: parent, ParentID: "cat_expenses", Kind: "expense"})
		for _, child := range children {
			d.Categories = append(d.Categories, domain.Category{ID: "cat_" + child, Name: child, ParentID: "cat_" + parent, Kind: "expense"})
		}
	}
	for _, name := range []string{"personal", "business", "travel", "hobby", "home", "mx5", "education", "gift", "tax-deductible", "subscription", "groceries"} {
		d.Tags = append(d.Tags, domain.Tag{ID: "tag_" + name, Name: name})
	}
	// Location-like junk from a taxonomy proposal run: it must stay selectable
	// without breaking the strict schema or the alias matcher.
	for _, name := range []string{"smart steuerservice", "kranken", "Chittaway Bay", "Toronto", "bruhl", "brunico", "St. Ulrich", "Git Server", "Mobilfunk", "Swopper"} {
		d.Merchants = append(d.Merchants, domain.Merchant{ID: domain.NewID("mer"), Name: name, Aliases: []string{}, DefaultTagIDs: []string{}, UseDefaults: false})
	}
	facts := domain.Facts{
		ID: "tx_finanzamt", Source: "enablebanking", AccountID: "acct_kontist",
		BookingDate: "2026-08-30", ValueDate: "2026-08-30", Amount: "-849.45", Currency: "EUR",
		RawDescription: "0904303543105 224/5220/5869",
		Counterparty:   "Finanzamt Bruehl", CounterpartyIBAN: "DE02120300000000202051",
		Fingerprint: "f1e2d3",
	}
	d.Transactions = []domain.Transaction{{Facts: facts, Enrichment: domain.Fallback(facts)}}
	return d, facts
}

// strictKeywords is what every targeted provider accepts in strict
// structured-output mode. uniqueItems is rejected outright by OpenAI-family
// endpoints ("'uniqueItems' is not permitted"); minItems/maxItems make Gemini
// expand array item schemas per element and reject real registries with a
// bare HTTP 400. Counts and duplicates are enforced server-side instead.
var strictKeywords = map[string]bool{
	"type": true, "properties": true, "required": true, "additionalProperties": true,
	"items": true, "enum": true, "maxLength": true, "minLength": true,
}

func checkStrict(t *testing.T, path string, value any) {
	t.Helper()
	switch v := value.(type) {
	case map[string]any:
		for key, child := range v {
			if path == "" || strings.HasSuffix(path, ".properties") {
				// Property names and the schema root are not keywords.
			} else if !strictKeywords[key] {
				t.Errorf("%s uses %q, which strict structured-output mode rejects", path, key)
			}
			checkStrict(t, path+"."+key, child)
		}
	case []any:
		for _, child := range v {
			checkStrict(t, path+"[]", child)
		}
	}
}

func TestWireSchemasUseOnlyStrictModeKeywords(t *testing.T) {
	d, facts := ledgerFixture()
	set := retrieve(facts.RawDescription, "expense", d, nil, nil)
	for name, schema := range map[string]map[string]any{
		"classification": set.schema(),
		"batch":          set.batchSchema([]string{"r1", "r2", "r3"}),
		"taxonomy":       taxonomySchema(),
		"csv":            csvMappingSchema(CSVMappingRequest{Headers: []string{"Buchung", "Betrag"}}),
	} {
		checkStrict(t, name, map[string]any{"properties": schema["properties"]})
	}
}

// The exact answer a live gpt-5.6-luna-pro returned for this row over a
// zero-data-retention route must land as reviewable enrichment: taxes
// category, a new public merchant seeded with the counterparty alias, no
// tags, recorded confidence.
func TestLedgerRowClassifiesThroughStrictSchema(t *testing.T) {
	d, facts := ledgerFixture()
	taxes := ""
	for _, c := range d.Categories {
		if c.Name == "taxes" {
			taxes = c.ID
		}
	}
	c := mockClient(t, func(w http.ResponseWriter, r *http.Request) {
		reply(w, `{"merchant_id":null,"new_merchant":"Finanzamt Bruehl","category_id":"`+taxes+`","tag_ids":[],"confidence":"high"}`)
	})
	p, err := c.Classify(context.Background(), facts, d, true)
	if err != nil {
		t.Fatal(err)
	}
	if p.Enrichment.CategoryID != taxes || p.Enrichment.Classification.Confidence != "high" {
		t.Fatalf("classification lost: %+v", p.Enrichment)
	}
	if p.NewMerchant == nil || p.NewMerchant.Name != "Finanzamt Bruehl" ||
		!reflect.DeepEqual(p.NewMerchant.Aliases, []string{"Finanzamt Bruehl"}) {
		t.Fatalf("merchant proposal lost: %+v", p.NewMerchant)
	}
	if len(p.Enrichment.TagIDs) != 0 {
		t.Fatalf("unexpected tags: %+v", p.Enrichment.TagIDs)
	}
}

// Identifier redaction must not eat ordinary 8- and 11-letter payee words,
// which blinded the model to the merchant it was asked to classify
// ("WWW.RACETRACKER.DE" became "WWW. .DE"). A bare bank-code-shaped token is
// vocabulary; real BICs still die labeled or trailing their IBAN.
func TestBICRedactionKeepsPayeeVocabulary(t *testing.T) {
	d, facts := ledgerFixture()
	clean := redactor(d, facts, nil)
	for _, keep := range []string{"Openbank", "OPENBANK", "Baumarkt", "BAUMARKT", "RACETRACKER", "toom Baumarkt"} {
		if got := clean(keep); got != normalize(keep) {
			t.Errorf("payee word %q was redacted to %q", keep, got)
		}
	}
	for name, text := range map[string]string{
		"labeled iban":  "IBAN DE89370400440532013000 COBADEFFXXX invoice",
		"trailing bic":  "pay DE89370400440532013000 COBADEFFXXX today",
		"labeled bic":   "BIC DEUTDEDBFRA",
		"labeled swift": "SWIFT GENODED1SPO",
	} {
		got := clean(text)
		if strings.Contains(got, "de8937") || strings.Contains(got, "cobadeff") || strings.Contains(got, "deutdedb") || strings.Contains(got, "genoded1") {
			t.Errorf("%s: identifier survived redaction: %q", name, got)
		}
	}
}

// One manual correction must outrank any number of the model's own past
// answers for the same payee: without source ranking, precedent feeds the
// model its uncorrected output as majority evidence and corrections never
// stick.
func TestManualCorrectionsOutrankAIPrecedent(t *testing.T) {
	d, _ := ledgerFixture()
	groceries, events := "", ""
	for _, c := range d.Categories {
		if c.Name == "groceries" {
			groceries = c.ID
		}
		if c.Name == "events" {
			events = c.ID
		}
	}
	add := func(id, date, category, source string) {
		f := domain.Facts{ID: id, Source: "test", AccountID: "acct_kontist", BookingDate: date,
			Amount: "-13.00", Currency: "EUR", Counterparty: "LVR Landesmuseum Bonn", Fingerprint: id}
		d.Transactions = append(d.Transactions, domain.Transaction{Facts: f, Enrichment: domain.Enrichment{
			Kind: "expense", CategoryID: category, TagIDs: []string{},
			Classification: domain.Provenance{Source: source},
		}})
	}
	// Many uncorrected AI answers, one older manual correction.
	for i := range 30 {
		add(fmt.Sprintf("tx_ai_%02d", i), "2026-08-20", groceries, "openrouter")
	}
	add("tx_corrected", "2026-08-01", events, "manual")
	target := domain.Facts{ID: "tx_new", AccountID: "acct_kontist", BookingDate: "2026-08-30",
		Amount: "-13.00", Currency: "EUR", Counterparty: "LVR Landesmuseum Bonn"}
	rows := history(target, d, func(s string) string { return normalize(s) }, 20)
	if len(rows) == 0 || rows[0].Source != "user" || rows[0].CategoryID != events {
		t.Fatalf("manual correction did not lead precedent: %+v", rows[0])
	}
	// The correction keeps its slot even in a window the AI rows could fill.
	users := 0
	for _, row := range rows {
		if row.Source == "user" {
			users++
		}
	}
	if users == 0 {
		t.Fatal("correction crowded out of the history window")
	}
}