Stop silently misreading real-world booking exports

A market-readiness audit found three parser bugs that produce WRONG risk verdicts
rather than errors, which is the worst failure mode this product has: the owner
cannot tell it happened, and the agent acts on it.

Dates. `new Date("04/09/2026")` reads American, so a UK, EU or French-Canadian
export lost every row whose day exceeded 12. Verified: an eight-visit regular on
a 28-day rhythm arrived as visitCount 1, cadence null, verdict "lost", and the
agent would then have emailed a loyal client to say she came once and never came
back. That is the same humiliation as the partial-reimport bug, reached by a
different road. The order is now inferred once from the whole column and applied
uniformly, since no single row can be disambiguated (04/09 is valid either way).

Money. Stripping every non-digit turned "165,50" into 16550, which annualised one
client to $201,358 and dominated the revenue-at-risk tile the paywall rests on.
Decimal comma and thousands comma are now told apart by what follows the last
separator.

Cancellations. The status column was read for nothing, so a no-show counted as a
visit. Verified: a client 62 days gone who no-showed 20 days ago reported "On
rhythm" and never surfaced. A churn product was hiding churn. She now reads
critical.

Also fixes a timezone bug in the same-day merge: toISOString on a local Date
rolls the day for anyone west of Greenwich.

Regression-tested end to end, including that MM/DD exports and the Aisha/Jane
demo argument still behave.

Author: JonathanSolvesProblems

src/lib/import-csv.ts

@@ -34,6 +34,10 @@ export type ImportResult = {
   clients: ClientAggregate[]
   visitsParsed: number
   rowsSkipped: number
+  /** Rows that were a cancellation or no-show, so deliberately not counted. */
+  cancelledSkipped?: number
+  /** Which way the date column was read, so the UI can say so and be corrected. */
+  dateOrder?: 'iso' | 'dmy' | 'mdy'
   /** Which CSV column was used for each field, so the UI can show its work. */
   columnsUsed: Record<string, string | null>
   error?: string
@@ -52,18 +56,88 @@ function findColumn(headers: string[], keywords: string[], exclude: string[] = [
   return null
 }
 
+/**
+ * Money, in whatever way the owner's country writes it.
+ *
+ * "165,50" is a hundred and sixty five euros fifty, not sixteen thousand. Stripping
+ * every non-digit turned it into 16550, which annualised one French-Canadian client
+ * to $201,358 and blew out the "revenue at risk" figure the whole paywall rests on.
+ */
 function parsePrice(raw: string): number {
   if (!raw) return 0
-  const cleaned = raw.replace(/[^0-9.-]/g, '')
-  const n = Number.parseFloat(cleaned)
+  let s = raw.replace(/[^0-9.,-]/g, '').trim()
+  if (!s) return 0
+
+  const lastComma = s.lastIndexOf(',')
+  const lastDot = s.lastIndexOf('.')
+
+  if (lastComma > -1 && lastDot > -1) {
+    // Both present: whichever comes last is the decimal separator.
+    s = lastComma > lastDot ? s.replace(/\./g, '').replace(',', '.') : s.replace(/,/g, '')
+  } else if (lastComma > -1) {
+    // Only commas. Exactly two trailing digits reads as a decimal separator
+    // ("165,50"); anything else is a thousands separator ("1,234", "1,234,567").
+    const tail = s.length - lastComma - 1
+    s = tail === 2 ? s.replace(',', '.') : s.replace(/,/g, '')
+  }
+
+  const n = Number.parseFloat(s)
   return Number.isFinite(n) && n >= 0 ? n : 0
 }
 
-function parseDate(raw: string): Date | null {
+/** Rows that are not a visit. A no-show is the opposite of one. */
+const NOT_A_VISIT = /cancel|no.?show|void|refund|declin|abandon|deleted|removed/i
+
+type DateOrder = 'iso' | 'dmy' | 'mdy'
+
+/**
+ * Work out whether the whole column is DD/MM or MM/DD, ONCE, from every row.
+ *
+ * `new Date("04/09/2026")` silently reads American, so a UK, EU or French-Canadian
+ * export lost every row whose day exceeded 12: an eight-visit regular arrived as
+ * `visitCount: 1` and the agent then wrote to her saying she came once and never
+ * came back. Deciding per row is not possible (04/09 is valid either way), so the
+ * order is inferred from the one column and applied uniformly.
+ */
+function detectDateOrder(samples: string[]): DateOrder {
+  let sawDayFirst = false
+  let sawMonthFirst = false
+  for (const raw of samples) {
+    const m = raw.trim().match(/^(\d{1,4})[/.-](\d{1,2})[/.-](\d{1,4})$/)
+    if (!m) continue
+    const a = Number(m[1])
+    const b = Number(m[2])
+    if (m[1].length === 4) return 'iso'
+    if (a > 12 && b <= 12) sawDayFirst = true
+    else if (b > 12 && a <= 12) sawMonthFirst = true
+  }
+  // Ambiguous columns (every value <= 12) fall back to month-first, which is what
+  // the platforms most of these exports come from emit.
+  if (sawDayFirst && !sawMonthFirst) return 'dmy'
+  return 'mdy'
+}
+
+function parseDate(raw: string, order: DateOrder): Date | null {
   if (!raw) return null
-  const d = new Date(raw.trim())
+  const s = raw.trim()
+
+  const m = s.match(/^(\d{1,4})[/.-](\d{1,2})[/.-](\d{1,4})$/)
+  let d: Date
+  if (m && m[1].length !== 4) {
+    const first = Number(m[1])
+    const second = Number(m[2])
+    let year = Number(m[3])
+    if (year < 100) year += year < 70 ? 2000 : 1900
+    const day = order === 'dmy' ? first : second
+    const month = order === 'dmy' ? second : first
+    // Local noon, so a timezone shift can never roll the date onto another day.
+    d = new Date(year, month - 1, day, 12)
+    if (d.getMonth() !== month - 1 || d.getDate() !== day) return null
+  } else {
+    d = new Date(s)
+  }
+
   if (Number.isNaN(d.getTime())) return null
-  // Guard against nonsense far-future/far-past rows.
   const year = d.getFullYear()
   if (year < 2000 || year > 2100) return null
   return d
@@ -93,8 +167,12 @@ export function parseBookingCsv(text: string, now: Date = new Date()): ImportRes
   const nameCol = findColumn(headers, ['clientname', 'customername', 'fullname', 'name'], ['service', 'staff', 'employee', 'business'])
   const serviceCol = findColumn(headers, ['service', 'item', 'treatment', 'description', 'title'])
   const priceCol = findColumn(headers, ['price', 'amount', 'total', 'paid', 'revenue', 'value'])
+  // A cancellation is not a visit, and counting one hides the exact client this
+  // product exists to catch: someone who no-showed three weeks ago reads as "on
+  // rhythm" and never surfaces.
+  const statusCol = findColumn(headers, ['status', 'state', 'appointmentstatus'])
 
-  const columnsUsed = { email: emailCol, date: dateCol, name: nameCol, service: serviceCol, price: priceCol }
+  const columnsUsed = { email: emailCol, date: dateCol, name: nameCol, service: serviceCol, price: priceCol, status: statusCol }
 
   if (!emailCol) {
     return {
@@ -117,15 +195,23 @@ export function parseBookingCsv(text: string, now: Date = new Date()): ImportRes
 
   const visits: ParsedVisit[] = []
   let skipped = 0
+  let notAVisit = 0
+
+  // Decided once, from the whole column, before any row is read.
+  const dateOrder = detectDateOrder(rows.map((r) => r[dateCol] ?? ''))
 
   for (const row of rows) {
     const email = (row[emailCol] ?? '').trim().toLowerCase()
-    const date = parseDate(row[dateCol] ?? '')
+    const date = parseDate(row[dateCol] ?? '', dateOrder)
     // No email or no usable date means the agent could never act on it.
     if (!email || !email.includes('@') || !date || date > now) {
       skipped++
       continue
     }
+    if (statusCol && NOT_A_VISIT.test(row[statusCol] ?? '')) {
+      notAVisit++
+      continue
+    }
     visits.push({
       email,
       name: (nameCol ? row[nameCol] : '')?.trim() || email.split('@')[0],
@@ -135,10 +221,20 @@ export function parseBookingCsv(text: string, now: Date = new Date()): ImportRes
     })
   }
 
-  return { clients: aggregate(visits), visitsParsed: visits.length, rowsSkipped: skipped, columnsUsed }
+  return {
+    clients: aggregate(visits),
+    visitsParsed: visits.length,
+    rowsSkipped: skipped,
+    cancelledSkipped: notAVisit,
+    dateOrder,
+    columnsUsed,
+  }
 }
 
-const dayKey = (d: Date) => d.toISOString().slice(0, 10)
+/** Local calendar day, not UTC: toISOString on a local Date shifts the day for
+ *  anyone west of Greenwich, which merged or split same-day appointments. */
+const dayKey = (d: Date) =>
+  `${d.getFullYear()}-${String(d.getMonth() + 1).padStart(2, '0')}-${String(d.getDate()).padStart(2, '0')}`
 
 /**
  * One appointment per client per day.