Add ExpectedAsset CSV import + bot search/pick for pre-known inventory

Lets the operator upload a CSV of already-known assets (asset tag, model,
serial, old area/assignee) before an office-move walk, via a new
session-authed /inventory-scans/import page. The Telegram bot gains a
"tim trong list" search step at the asset-tag prompt: picking a match
pre-fills tag/serial and offers one-tap "giu nguyen" buttons for the
old area/assignee, then marks that row matched once the scan is created.

Follows the same-day decision to turn off local OCR (qwen2.5vl via Ollama)
for this feature -- see CLAUDE.md and the ITAM/Ollama vault notes for the
accuracy/speed numbers behind that call.

Deviates from repo convention same as the InventoryScan work above: no
Vitest coverage (manual smoke test only, via simulated webhook calls),
hand-written migration, not through the ba->pm->feature-implementer->
test-runner->code-reviewer pipeline. tsc/lint are clean.

prisma/schema.prisma

@@ -565,6 +565,40 @@ model InventoryScan {
   telegramChatId     String?        @map("telegram_chat_id")
   telegramMessageId  Int?           @map("telegram_message_id")
   createdAt          DateTime       @default(now()) @map("created_at") @db.Timestamptz(6)
+  // Back-reference for ExpectedAsset.matchedScanId -- set once someone
+  // picks this scan's pre-uploaded row via the bot's "🔍 Tìm trong list"
+  // step. No @relation attribute needed here; the FK lives on the other
+  // side.
+  expectedAsset      ExpectedAsset?
 
   @@map("inventory_scans")
 }
+
+// A row from a pre-move Excel/CSV export of known assets (asset tag,
+// model, old area/owner), uploaded via POST /api/expected-assets/import
+// before an office-move walk starts (2026-08-21 decision -- team decided
+// plain manual entry was faster/more reliable than local OCR for ~300-400
+// assets, see docs/asset-license-management.md, but typing in a KNOWN
+// asset tag from scratch during the walk is still wasted effort when it's
+// sitting right here). Purely a lookup/staging table for the bot's "🔍 Tìm
+// trong list" picker -- modelText/oldAreaText/oldAssigneeText are free
+// text straight from the import file, deliberately NOT FK'd into
+// AssetModel/Area/Employee: import-file text won't reliably match this
+// app's catalog IDs/spelling, and forcing that match at import time would
+// block the whole import on one typo instead of just being a soft
+// "here's the old value" suggestion during the walk. Marked matched (via
+// matchedScanId) once someone actually picks it while creating a scan, so
+// it stops showing up in future searches.
+model ExpectedAsset {
+  id              String         @id @default(uuid()) @db.Uuid
+  assetTag        String?        @map("asset_tag")
+  modelText       String         @map("model_text")
+  serialNumber    String?        @map("serial_number")
+  oldAreaText     String?        @map("old_area_text")
+  oldAssigneeText String?        @map("old_assignee_text")
+  matchedScanId   String?        @unique @map("matched_scan_id") @db.Uuid
+  matchedScan     InventoryScan? @relation(fields: [matchedScanId], references: [id])
+  createdAt       DateTime       @default(now()) @map("created_at") @db.Timestamptz(6)
+
+  @@map("expected_assets")
+}