feat(csv-import): drop-anywhere import — dedupe, account detection, saved profiles

.gitignore

@@ -31,6 +31,7 @@ vendor/
 internal/db/db.go
 internal/db/models.go
 internal/db/*.sql.go
+internal/db/copyfrom.go
 
 # templ generated files — *.templ files are source, *_templ.go are artifacts
 # produced by `make templ` (or `templ generate`). Keep generated Go out of

Makefile

@@ -247,7 +247,7 @@ sqlc-install:
 	fi
 
 sqlc: sqlc-install
-	@rm -f internal/db/*.sql.go internal/db/models.go internal/db/db.go
+	@rm -f internal/db/*.sql.go internal/db/models.go internal/db/db.go internal/db/copyfrom.go
 	sqlc generate
 	@$(MAKE) sqlc-tag
 
@@ -256,7 +256,7 @@ sqlc: sqlc-install
 # `emit_build_tag` option (added in 1.27+), so we patch the artifacts in a
 # follow-up step. CI runs the same target.
 sqlc-tag:
-	@for f in internal/db/db.go internal/db/models.go internal/db/*.sql.go; do \
+	@for f in internal/db/db.go internal/db/models.go internal/db/copyfrom.go internal/db/*.sql.go; do \
 		if [ -f "$$f" ] && ! head -1 "$$f" | grep -q '^//go:build'; then \
 			printf '//go:build !lite\n\n' | cat - "$$f" > "$$f.tmp" && mv "$$f.tmp" "$$f"; \
 		fi; \

docs/csv-import.md

@@ -0,0 +1,123 @@
+# CSV import
+
+Breadbox imports transactions from CSV/TSV exports. There are two paths:
+
+- **v2 — drop anywhere (recommended).** Drag a CSV onto any page. A global overlay
+  appears; on drop the import flow analyzes the file, detects which account it
+  belongs to, dedupes against what's already there, shows a preview of exactly
+  what will change, and applies only on confirm.
+- **v1 — wizard page.** The legacy 4-step wizard at
+  `/admin/connections/import-csv` (upload → map → preview → import). Still works;
+  kept for now. v2 supersedes it.
+
+This doc covers v2. v1 internals live in `internal/admin/csv_import.go`.
+
+## The flow
+
+```
+drop CSV (any page)
+  → POST /-/csv/v2/sessions            analyze: parse, detect template/mapping/
+                                       currency, apply saved profile, rank accounts
+  → [account step, if not auto-chosen]
+      POST /-/csv/v2/sessions/{id}/resolve   bind/create account → classify rows
+  → preview/edit
+      GET   /-/csv/v2/sessions/{id}/rows           paginated rows + summary
+      PATCH /-/csv/v2/sessions/{id}/rows/{rowId}   inline edit a row
+      POST  /-/csv/v2/sessions/{id}/bulk           include/exclude/set-category/remap
+  → POST /-/csv/v2/sessions/{id}/apply  upsert the included set, save the profile
+```
+
+All endpoints are admin-only (session-cookie auth), so they carry no OpenAPI
+surface. The flow UI is `components.ImportModal` + `$store.csvImport` (drop
+overlay + open/close) + the `csvImport` Alpine factory
+(`static/js/admin/components/csv_import_v2.js`). See the live variant matrix at
+`/design/c/csv-import`.
+
+## Staging (durable, so preview == apply)
+
+A drop creates a `csv_import_sessions` row holding the original bytes
+(`raw_blob`), detected mapping, resolved account, and (after apply) a result
+snapshot. Each parsed row is a `csv_import_rows` row carrying its parsed values,
+classification, dedup hashes, and the user's include/exclude intent. Classifying
+or remapping replaces the row set wholesale (delete + `:copyfrom`), so up to 50k
+rows insert fast. Abandoned sessions are swept hourly past their 24h TTL
+(`SweepExpiredImportSessions`, wired in `serve.go`).
+
+Service layer: `internal/service/csv_import_v2.go` (lifecycle + apply),
+`csv_classify.go` (dedup classification), `csv_account_match.go` (account
+detection), `csv_profiles.go` (profile CRUD).
+
+## Account detection
+
+`MatchCSVAccounts` ranks existing accounts with a deterministic, explainable
+weighted score (every point carries a reason string):
+
+| Signal | Points |
+|---|---|
+| saved profile's default account | auto-preselect |
+| mask / last-4 matches the account | +40 |
+| transaction overlap (rows already present) | up to +40 |
+| institution name appears in the account name | +15 |
+| filename mentions the account | +5 |
+
+The top account is **pre-selected** only when it scores ≥70 *and* leads the
+runner-up by ≥20 (or a profile matched). Otherwise the user picks from the ranked
+list or creates a new account inline. Confidence buckets: high ≥70, medium 40–69,
+low <40.
+
+## Deduplication
+
+Per incoming row, against the **resolved account's** live transactions (any
+provider — CSV, Plaid, Teller):
+
+- **exact_dup** — the row's stable `provider_transaction_id`
+  (`SHA-256(accountID|date|amount|desc)`) already exists, or its account-
+  independent `content_hash` (`SHA-256(date|amount|desc)`) matches an existing
+  row. Default: excluded.
+- **probable_dup** — same amount, date within ±1 day, and a similar name (scored
+  by the shared `internal/textmatch` similarity used by the sync matcher).
+  Default: excluded.
+- **new** — no match. Default: included.
+- **error** — unparseable row; surfaced for inline fixing. Excluded.
+
+The classifier loads candidates with one indexed query
+(`transactions_dedup_lookup` on `(account_id, date, amount) WHERE deleted_at IS
+NULL`). Apply upserts via `UpsertTransactionV2` (returns `(xmax = 0) AS inserted`
+for reliable new-vs-updated counts) and is **idempotent** — re-dropping the same
+file into the same account upserts to no-ops. A genuine same-key duplicate the
+user force-imports gets an occurrence disambiguator so it inserts as a distinct
+row.
+
+Why re-dropping a later export "just adds the new ones": the overlapping rows hit
+exact_dup (and the saved profile auto-resolves the account), so only the genuinely
+new rows are classified `new` and imported.
+
+## Profiles
+
+A profile is a saved import recipe keyed by a **header fingerprint** (a hash of
+the file's normalized header row). On a successful apply the profile is upserted:
+created the first time a given header layout is imported, then bumped
+(`times_used` / `last_used_at`) and updated to remember the resolved account. On a
+future drop whose fingerprint matches, the flow auto-applies the saved
+mapping/date-format/sign/currency and pre-selects the remembered account —
+collapsing straight to a deduped preview.
+
+Manage profiles via `GET/PATCH/DELETE /-/csv/v2/profiles[/{id}]`
+(`internal/service/csv_profiles.go`). The user's rename survives future imports
+(the upsert deliberately doesn't overwrite `name`).
+
+## Currency
+
+Detected from the file (currency symbols in the amount column), defaulting to the
+target account's currency, confirmable in the flow. Amounts are stored as-is with
+their `iso_currency_code`; nothing is converted (honoring the never-sum-across-
+currency invariant).
+
+## Schema
+
+Additive migrations only (shared dev DB):
+
+- `transactions.content_hash` + partial index `transactions_dedup_lookup`.
+- `csv_import_profiles` (saved recipes; one per user+fingerprint).
+- `csv_import_sessions` + `csv_import_rows` (durable staging).
+- `csv_import_rows.category_id` (user-chosen Breadbox category applied at import).