void: a statementless row for a charge that never posted (migration 0040)

A card notification is sent for an authorisation; a released hold never
reaches a statement. Txn 7951 (a $1,040.41 Airbnb hold, booking cancelled
the next day) sat in the reconcile queue forever and counted as spend.
Neither reconcile (needs a statement twin) nor supersede (needs a survivor)
can express it.

- transactions.voided_at + void_reason; CHECKs: reason mandatory,
  statementless only
- EXCLUDE_RECONCILED_SOURCE excludes voided rows; reconcile queue,
  findAutoMatches, both adoption helpers, annual charges and the manual
  reconcile POST filter it; merge refuses a voided row on either side
- POST/DELETE /api/transactions/<id>/void; refuses reconciled, superseded
  and settled-split rows
- "Never posted" button on /reconcile
- wallet-ingestion deliberately still sees voided rows, so a late twin or a
  prompt_version re-post cannot rebook the hold

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

CLAUDE.md

@@ -65,7 +65,7 @@ docker logs finance -f
 npm test                  # unit — pure logic, no database
 npm run test:integration  # database-backed; resolves the DSN itself (see below)
 npm run test:all          # both
-npm run test:setup        # once, and after any migration: creates/refreshes personal_test
+npm run test:setup        # once, and after any migration: drops and rebuilds personal_test
 npm run lint              # blocking in CI since #79 — 0 errors, warnings allowed
 npx tsc --noEmit          # blocking in CI; nothing else compiles the test files
 ```
@@ -86,6 +86,13 @@ had pointed the suite at **postgres-pantry** — whose database it would have
 TRUNCATEd had the credentials happened to match. A loopback address cannot be
 reassigned to another container, which is why the published port wins.
 
+**`test:setup` really rebuilds now (2026-10-03).** It used to replay the
+schema dump over the existing database, where every `CREATE TABLE` failed with
+"already exists" and the old shape survived — so no migration ever reached
+`personal_test`, and new-column code failed every test with a misleading
+`column does not exist`. It now `DROP DATABASE … WITH (FORCE)`s first, refusing
+any name not ending `_test`, and loads with `ON_ERROR_STOP`.
+
 **Two guards, and both matter.** `helpers.ts::assertTestDatabase` refuses to
 truncate anything whose `current_database()` is not `personal_test`, and the
 suite runs `fileParallelism: false` because every file shares that one database
@@ -176,6 +183,215 @@ docker exec postgres-personal psql -U personal -d personal < prisma/migrations/<
 - `statements` — one row per billing period per bank account
 - `transactions` — line items; `statement_id` is nullable (NULL = manual entry); `reconciled_with_id` links a manual tx to its matched statement tx; `payment_method` (migration 0016, `credits` added by 0018) is `card | cash | bank_transfer | credits | other`, NULL = unknown, and the TypeScript mirror is `src/lib/payment-methods.ts`; `source` / `source_ref` / `source_account` (migration 0028) identify a row that came from an account feed rather than a statement or a hand entry — `source_ref` is the provider's own id and carries a partial unique index, which is the only thing making a re-import idempotent
 
+### The change log (migration 0037)
+
+`transaction_audit` is an **append-only** row-level record of every change to
+`transactions`, `transaction_splits`, `transaction_tags`, `transaction_overrides`
+and `expense_metadata`. Never `UPDATE` or `DELETE` from it. There is deliberately
+**no foreign key** to `transactions` — the log has to outlive the row it describes.
+
+```sql
+-- everything that ever happened to one transaction
+SELECT changed_at, table_name, op, old_data, new_data, app_name
+  FROM transaction_audit WHERE transaction_id = 6961 ORDER BY changed_at;
+
+-- one logical edit, all the rows it touched (a reconcile moves 4 children at once)
+SELECT * FROM transaction_audit WHERE txid = <txid> ORDER BY id;
+```
+
+**Why it exists.** On 2026-08-23 the n8n CC Statement Processor stamped
+`statement_id` onto seven statementless wallet/pantry rows without setting
+`reconciled_with_id`, double-counting $245.90 and hiding those rows from the
+reconcile queue (`WHERE statement_id IS NULL`). Answering *"was anything lost?"*
+took restoring the 02:30 `pg_dump` into a scratch database and diffing it column
+by column, because nothing recorded what changed.
+
+**Why it is a trigger and not application code.** The ledger has at least five
+independent writers — this app (Prisma), n8n (raw SQL in a Postgres node),
+pantry-app, the ingestion-engine runner (psycopg), and ad-hoc `docker exec psql`.
+That bug was an n8n raw-SQL `UPDATE`; an app-level audit would not have seen it.
+
+**Why not Postgres' own logs.** Checked live: `wal_level=replica` (WAL cannot be
+decoded into row changes), `wal_keep_size=0` and `archive_mode=off` (recycled
+almost immediately), `log_statement=none`, `track_commit_timestamp=off`, no
+`pgaudit`. Postgres records nothing usable and has no built-in audit or temporal
+tables.
+
+Three details worth knowing before relying on it:
+- **`old_data`/`new_data` hold only the columns that actually changed** on an
+  `UPDATE`. A whole-row copy would bury a one-column edit, and
+  `expense_metadata.line_items` alone can be kilobytes.
+- **A no-op `UPDATE` writes nothing** (the upsert paths issue plenty).
+- **Attribution is partial and that is understood.** Every client connects as the
+  same `personal` role, so `db_user` never discriminates. `app_name` is only as
+  good as what each client sets; `client_addr` separates containers and is NULL
+  for `docker exec` over the unix socket — itself a useful signal. `txid` is the
+  reliable grouping key.
+
+`TRUNCATE` does not fire row-level triggers, so the integration suite's table
+wipes produce no audit rows.
+
+### Resolving a duplicate by hand (`POST /api/transactions/<id>/merge`)
+
+`<id>` is the row that gets **hidden**; `other_id` is the one that **survives**.
+
+| mode | requires | sets | for |
+|---|---|---|---|
+| `reconcile` | loser statementless, survivor a statement row | `reconciled_with_id` | the everyday case |
+| `supersede` | nothing | `superseded_by_id` | two rows on the same footing, e.g. a statement imported twice |
+
+**There is no delete, deliberately.** Every child of `transactions` is
+`ON DELETE CASCADE`, so removing a row destroys the splits, tags and metadata
+that make it legible. Both modes keep the row and its children and merely stop
+it counting. Do not add a delete endpoint.
+
+Both modes **move curation onto the survivor first**, through the single
+`moveCurationWithin()` in `lib/reconcile.ts` — because both hide the source row
+from every figure, and a split stranded on a hidden row stops counting in
+silence. That is also why the two callers share one implementation: the file
+header warns that a second copy would eventually move a different set.
+
+Refusals (all deliberate, all tested in `merge-supersede.test.ts`): a row already
+reconciled or superseded on **either** side → 409 (re-pointing it strands the
+curation moved on the first pass); superseding under a row that is itself
+superseded → 409; a row that is not the caller's → 404. Ownership is checked with
+`COALESCE(t.owner_id, s.owner_id)` — a statement row carries `owner_id` on the
+statement, so checking `t.owner_id` alone rejects every statement row as "not
+found".
+
+**Why it had to exist.** `superseded_by_id` was read everywhere (analytics and
+auto-reconcile both exclude it) but written **nowhere** in application code — set
+exactly once, by migration 0023. The reconcile queue was the only merge path, and
+it only offers candidates inside an amount band.
+
+**The band is ±5%, not ±1% (2026-08-23).** A foreign card purchase is charged the
+notified amount **plus the card's foreign transaction fee** — 3% on the Westpac
+Altitude — so a wallet row and its statement line can never match exactly.
+`SP PEAKDESIGN` was notified at **$42.44** and charged **$43.71** (×1.03), which
+put the one row that *was* its twin outside the old band: the picker could not
+offer the only correct answer. Widening is safe because the list is capped at 5
+by the `rn` window and ordered by amount closeness, so a looser band changes
+*which* five appear, never how many, and a person confirms each. Measured on the
+live queue at the time of the change: candidate counts went **0 → 0** on all 12
+rows, i.e. no new noise. **The auto-matcher stays exact** — see `findAutoMatches()`.
+
+### A charge that never posted (`POST /api/transactions/<id>/void`, migration 0040)
+
+The third way a row stops counting, for a case with **no survivor**: a card
+notification is sent for an *authorisation*, and when the merchant releases the
+hold no statement line ever arrives. Txn 7951 — a $1,040.41 Airbnb hold on Amex,
+booking cancelled the next day — sat in the reconcile queue with nothing to
+match and counted as spend everywhere. `reconcile` needs a statement twin,
+`supersede` needs a survivor (and `transactions_no_self_supersede` blocks
+pointing a row at itself), so it gets its own columns: `voided_at` +
+`void_reason` (reason mandatory, by CHECK). Body `{reason}`; `DELETE` undoes it.
+UI: **Never posted** on every row of `/reconcile`.
+
+- **In `EXCLUDE_RECONCILED_SOURCE`**, so every "count each purchase once" query
+  drops it without being edited; the queue, `findAutoMatches`, both adoption
+  helpers, annual charges and the manual reconcile POST filter it explicitly.
+- **Statementless only** — `transactions_void_statementless` enforces it in the
+  database. A statement line is the bank's record of a posted charge; a wrong
+  one is a supersede.
+- **Refused** on a reconciled row (it posted), a superseded row, and a row with a
+  *settled* split (someone already paid their share — voiding would quietly turn
+  that into a credit). `merge` refuses a voided row on either side.
+- **The wallet lane deliberately still SEES voided rows.** `alreadyIngested()` and
+  the issuer/Wallet pairing in `wallet-ingestion.ts` do not filter `voided_at`:
+  the voided row is the sink that stops a late twin notification, or the next
+  `prompt_version` re-post, from booking the hold again. Same for the intake
+  router's idempotency lookup. Do not "fix" those.
+- **Outside this repo, nothing filters it**: Hermes' `agent_ro` views
+  (`cashflow_monthly` etc.) and ingestion-engine's `instalment_matcher` already
+  ignore `superseded_by_id`/`reconciled_with_id` too, so a voided row is no worse
+  there than any other hidden row.
+
+### Upcoming annual charges (`GET /api/annual-charges`, migration 0038)
+
+Card annual fees, loan package fees and annual subscriptions, derived from
+**payments alone** — independent of the email lane, because a bank announces an
+annual fee on the statement and nowhere else.
+
+**`transaction_type = 'fee'` is the universe, not a description match.** The
+extractor already classifies fees structurally (131 rows), and *every* card and
+loan annual fee on this ledger is typed that way. The first version of this
+filtered `transaction_type = 'debit'` and returned **none** of them. The fees are
+not debits.
+
+**Two signals, kept apart because they are not equally trustworthy:**
+
+| confidence | basis |
+|---|---|
+| `confirmed` | typed a fee **and** repeated on an annual cadence |
+| `likely` | typed a fee, seen once, and *named* annual/card/package — the kind is certain, the cadence inferred |
+| `possible` | no fee wording; a merchant whose every gap is annual **and** whose amount is stable |
+
+**A naive "two charges ~365 days apart" is useless** — measured on this ledger it
+ranks a bakery first with **202** such pairs, then a donation, Patreon, Coles.
+Frequency is the confound. Even "every gap is annual" admits a Target visited
+twice ($30, then $198). The `possible` tier therefore also requires amount
+stability (≤20%, which absorbs VicRoads' ~3%/yr CPI drift) and ≥2 observations.
+It still carries real false positives (H&M, Superdry) — that is what the label is
+for, and why it must never render like a bank fact.
+
+**`expected_not_seen` is the more interesting status.** The estimate passed and
+nothing landed, which means the account stopped being imported, or closed, or the
+charge arrived under a different label. `account_days_stale` is the
+discriminator: Westpac at 0 days stale is trustworthy; ANZ at 392 is simply
+missing data. Never report these as upcoming.
+
+**Account continuity (`account_links`, migration 0038).** One card can appear
+under two identities: Citi Prestige `5258…0253` (to 2025-11-13) became NAB MyCard
+`1015…6227` (from 2025-11-14), same 14th-to-13th cycle, no gap. Without the link
+the $700 fee reads as two dead one-offs instead of one due again ~Nov 2026. The
+table is **hand-asserted with evidence** (`source_documents` ids), not inferred —
+a contiguous-cycle heuristic would join two genuinely different accounts as
+readily as one renamed one. The resolver walks it forward recursively;
+`uq_account_links_from` forbids a fork and `account_links_no_self` a self-link, so
+it terminates. **Display the resolved current bank**, not the one that printed the
+last fee — naming the retired brand sends you to the wrong bank.
+
+Known limitation: the merchant signal fragments on label drift. VicRoads rego was
+paid 2024/2025 as `VICROADS` and 2026 as `VICROADS ONLINE PAYM`, so the series
+splits and the 2025 estimate shows as `expected_not_seen`. Not fuzzy-matched on
+purpose — see the rejected pg_trgm merchant grouping.
+
+### The wallet lane (`/api/wallet/ingest`, `lib/wallet-ingestion.ts`)
+
+iOS notifications → `capture.wallet_events` → n8n **Wallet Interpret**
+(`CAp1s5cuzJk2mDaI`, local gemma, every 15 min) → this route. A row per tap,
+statementless, `source='wallet'`, `source_ref='evt:<id>'`; the statement later
+confirms it through the reconcile queue. Fixed 2026-10-03 after a month abroad
+(ticket #266); the rules below are each the fix for something that broke.
+
+- **Storage convention differs from the order lane.** `amount` = `amount_aud` =
+  a *provisional AUD* figure; the native figure is `foreign_currency_amount`.
+  `NATIVE_CURRENCY` reads `amount_aud IS NOT NULL` as "amount is AUD" — without
+  that an INR 524 tap displayed as "INR 9.42".
+- **Provisional rate:** ledger median over *statement* rows (≥5, carries the
+  card's FX fee) → else ECB via Frankfurter (`rate_ecb`) → else `0` with
+  `no_rate_reference`. Never the native figure as AUD (672006 VND once became
+  A$672,006).
+- **One tap = two notifications** (issuer app + Wallet, 2–8 s apart), merged.
+  A pair must be **≤10 minutes apart AND from different apps**
+  (`expense_metadata.notified_at` / `notif_app`, migration 0039). The old
+  amount-within-±1-day rule merged the UNICEF $50 into an unrelated $50 charge.
+- **The whole check→pair→adopt→insert runs under one advisory lock**
+  (`hashtextextended('wallet-ingest',0)`), on the transaction's own connection.
+  n8n posts a poll's items concurrently; without it 11 taps became 22 rows.
+  Prisma cannot deserialise `void` — the lock SELECT needs `::text`.
+- **A merged half records `wallet_evt:<id>` in the survivor's flags**, so a
+  replay is idempotent. Reconciled wallet rows carry **no** `expense_metadata`
+  (reconcile moves it to the statement twin) — the flag check follows it there.
+- **Skipped, not booked:** account notices (`overdrawn`, `to avoid fees`, …) and
+  transit-card drawdowns (`Your new/current balance is` — the Octopus *load* is
+  the spend). The fixed-width descriptor parser runs on Westpac bodies only.
+- **Dated in Australia/Melbourne**, like the statements (UTC slicing filed 28 of
+  134 rows a day early).
+- **Replay a few events by relabelling their interpretations**
+  (`UPDATE capture.wallet_interpretations SET prompt_version='v1-<why>' WHERE …`),
+  not by bumping `prompt_version`, which re-posts every event ever captured.
+
 ### Cash and reconciliation
 
 `payment_method` in (`cash`, `credits`) excludes a transaction from
@@ -699,6 +915,25 @@ and the cross-repo contract are in smarthome `plans/intake-router-slice-plan.md`
   delivery order's total is items + fees − promotions ($118.94 of Coles lines
   became $85.74), so checking lines against the total would hold every
   delivery order. Tolerance $1 or 2%; a failure is `held`, nothing written.
+- **Nothing that cannot be a receipt reaches the model, and the prompt names no
+  real shop.** On 2026-08-22 an iOS Shortcut coerced a shared photo to text, so
+  a batch arrived as the five characters `Image` — and the model returned a
+  whole receipt: `Burgertory (Braybrook)`, two "The Pride" lines, $41.15. None
+  of it existed; "Burgertory" was in `ORDER_ITEM_PROMPT` as the example of how
+  to print a branch in parentheses. Only the absent date kept it out of the
+  matcher, and the invented header and items were still stored on the batch as
+  `needs_header`. Three layers now, in order: `textCouldBeReceipt()` rejects a
+  TEXT part with fewer than two words, no digit, or under 40 characters with no
+  money figure (`TOTAL $17.00` passes — length alone is *not* the test)
+  **before** any model call; the prompt's merchant examples are
+  `PROMPT_PLACEHOLDER_MERCHANTS` format placeholders and it is told to return
+  nulls for a non-receipt; and `confabulationReason()` catches what the
+  pre-check cannot see — a PDF that renders to junk — plus any answer echoing a
+  placeholder. All three file `unsupported`, never `needs_header`: a
+  confabulated header is not a partial one. Same shape as the ingestion
+  engine's 2026-08-05 empty-body defect — **a stage that can be handed nothing
+  must say so, because the model downstream always answers, and it answers out
+  of its prompt.**
 - **Names stay verbatim.** Burgertory calls its burgers "Michael" and "Joel";
   the prompt forbids canonicalising and the stored shape is the order lane's
   (`qty`, `description`, `amount`, `options`) so `order-details.tsx` renders it.

prisma/migrations/0035_transaction_search_matched_txn/migration.sql

@@ -0,0 +1,73 @@
+-- transaction_search: resolve expense_metadata by BOTH links, without fanning out.
+--
+-- THE DEFECT (measured 2026-08-22): the view joined only on
+-- `em.transaction_id = t.id`, so it silently omitted every card-settled order,
+-- which carries only `matched_transaction_id`. 148 transactions are affected --
+-- Glovo orders in Milan, Uber Eats, all `source='order-bridge'` -- and for each
+-- of them the view reports NULL payment_method / order_reference / line_items
+-- while the data sits in the table.
+--
+-- That is not cosmetic. This view is the agent read surface: Hermes'
+-- postgres-personal MCP selects payment_method and order_reference from it
+-- (personal-agent-gateway/mcp-servers/postgres-personal/queries.py), and
+-- home-dashboard reads it via dash_ro. CLAUDE.md states the rule the view broke:
+-- "That lookup must match transaction_id OR matched_transaction_id: a
+-- card-settled order has only the latter."
+--
+-- NOT a fan-out fix, contrary to a first reading. The original join CANNOT fan
+-- out: expense_metadata_transaction_id_key is UNIQUE on transaction_id, so it
+-- matched at most one row by construction. Sums taken over this view were never
+-- inflated.
+--
+-- But broadening the join is exactly what INTRODUCES that risk, which is why
+-- LATERAL is used rather than a second OR'd condition. transaction_id and
+-- matched_transaction_id are enforced by two SEPARATE unique indexes
+-- (expense_metadata_transaction_id_key, uq_expense_matched_txn), so one
+-- transaction can legally be pointed at by two DIFFERENT rows -- one via each
+-- link. Zero such transactions exist today; the LATERAL means that stays a
+-- non-event rather than a silent duplicate in the agent's spending totals.
+--
+-- Tie-break matches every other reader (getTransactions, replaceLineItems,
+-- matchTransaction, /api/transactions/[id]/order), so the row this view resolves
+-- is the row the UI shows.
+--
+-- Verified before applying: row count unchanged at 4760 (= count(transactions)),
+-- sum(amount) over debits unchanged at 1514412.89 -- amount comes from
+-- transactions, never from the metadata row -- and the only difference is the
+-- 148 rows that gain the metadata they always should have had.
+--
+-- CREATE OR REPLACE VIEW DROPS GRANTS. dash_ro (home-dashboard) and agent_ro
+-- (Hermes) both read this view and losing either breaks a surface with no error
+-- at replace time, so both are re-granted below.
+
+CREATE OR REPLACE VIEW transaction_search AS
+ SELECT t.id,
+    t.transaction_date,
+    t.description,
+    COALESCE(t.merchant_normalized, t.merchant_name) AS merchant,
+    t.category,
+    t.transaction_type,
+    COALESCE(t.amount_aud, t.amount) AS amount,
+    t.foreign_currency_amount,
+    t.foreign_currency_code,
+    t.location,
+    s.bank_name,
+    s.card_name,
+    s.account_number,
+    s.statement_type,
+    s.billing_end_date,
+    em.payment_method,
+    em.order_reference,
+    em.line_items
+   FROM transactions t
+     LEFT JOIN statements s ON s.id = t.statement_id
+     LEFT JOIN LATERAL (
+       SELECT em2.payment_method, em2.order_reference, em2.line_items
+         FROM expense_metadata em2
+        WHERE em2.transaction_id = t.id OR em2.matched_transaction_id = t.id
+        ORDER BY (em2.transaction_id = t.id) DESC NULLS LAST, em2.id
+        LIMIT 1
+     ) em ON true;
+
+GRANT SELECT ON transaction_search TO dash_ro;
+GRANT SELECT ON transaction_search TO agent_ro;

prisma/migrations/0036_normalize_foreign_currency_code/migration.sql

@@ -0,0 +1,73 @@
+-- Normalise foreign_currency_code to ISO-4217, and clear two sentinel values.
+--
+-- WHY THIS MATTERS: exact-code equality is the whole basis of FX reconciliation.
+-- ingestion-engine's order_transaction_bridge matches foreign orders with
+--   AND t.foreign_currency_code = %(ccy)s
+--   AND t.foreign_currency_amount = %(total)s
+-- so a row labelled 'BAHT' can never match a THB order, and a row labelled
+-- 'null' can never match anything. 226 of 1,632 coded rows (14%) were
+-- unmatchable for this reason alone, which silently degrades a shipped job.
+--
+-- THE SENTINELS. 187 rows carry the literal 4-character STRING 'null' rather
+-- than SQL NULL, and 42 of those also carry foreign_currency_amount = -1.00
+-- exactly. Neither is data: they are "no value" markers that survived a truthy
+-- guard. Root cause is in the CC statement processors' Parse node, which does
+--   foreign_currency_code: t.foreign_currency_code || null
+-- and a non-empty string is truthy, so a model that answers with the STRING
+-- "null" (rather than JSON null) passes straight through. Same for -1. Fixed
+-- workflow-side in the same change; this migration cleans what already landed.
+-- Not purely historical -- 155 rows in 2026-03, but also 8 in May, 3 in July
+-- and 21 in August 2026.
+--
+-- MAPPINGS WERE VERIFIED AGAINST IMPLIED RATE, NOT TRUSTED FROM THE LABEL,
+-- because a row whose rate contradicts its claimed currency is a broken ledger
+-- row rather than a mislabelled one (the trap order_transaction_bridge's rate
+-- guard exists for: 2 of its first 73 matches were broken rows, not bad
+-- matches). Median amount/foreign_currency_amount per raw code vs the ISO
+-- median over statement rows (n>=5):
+--   BAHT          0.0433  vs THB 0.0470   INDIAN RUPEE 0.0183 vs INR 0.0180
+--   U.A.E. DIRHAM 0.4287  vs AED 0.4184   US DOLLAR    1.4054 vs USD 1.5513
+--   NZ DOLLAR     0.8243  vs NZD 0.9311   EURO         1.6342 vs EUR 1.6964
+-- Every label agrees with its rate. foreign_currency_amount is untouched by the
+-- relabels -- only the code was ever wrong, so this is a relabel, never a
+-- recompute.
+--
+-- EXPECTED ROW COUNTS (verified pre-flight, includes superseded rows):
+--   code 'null'    -> NULL   187      BAHT          -> THB   13
+--   amount -1.00   -> NULL    42      U.A.E. DIRHAM -> AED   10
+--   NZ DOLLAR      -> NZD      6      US DOLLAR     -> USD    4
+--   INDIAN RUPEE   -> INR      4      EURO          -> EUR    2
+-- Pre-flight also confirmed: no -1.00 amount exists outside the 'null'-code
+-- rows, and no other non-ISO code exists.
+
+BEGIN;
+
+-- Sentinels first: the -1.00 amounts live only on 'null'-code rows, so clear
+-- them before that code is nulled and the set becomes unfindable.
+UPDATE transactions SET foreign_currency_amount = NULL
+ WHERE foreign_currency_amount = -1.00 AND foreign_currency_code = 'null';
+
+UPDATE transactions SET foreign_currency_code = NULL
+ WHERE foreign_currency_code = 'null';
+
+-- Word-form relabels.
+UPDATE transactions SET foreign_currency_code = 'THB' WHERE foreign_currency_code = 'BAHT';
+UPDATE transactions SET foreign_currency_code = 'AED' WHERE foreign_currency_code = 'U.A.E. DIRHAM';
+UPDATE transactions SET foreign_currency_code = 'NZD' WHERE foreign_currency_code = 'NZ DOLLAR';
+UPDATE transactions SET foreign_currency_code = 'USD' WHERE foreign_currency_code = 'US DOLLAR';
+UPDATE transactions SET foreign_currency_code = 'INR' WHERE foreign_currency_code = 'INDIAN RUPEE';
+UPDATE transactions SET foreign_currency_code = 'EUR' WHERE foreign_currency_code = 'EURO';
+
+-- Postcondition: every remaining code is ISO-4217 shaped. Fails the whole
+-- transaction rather than leaving a partial normalisation behind.
+DO $$
+DECLARE bad int;
+BEGIN
+  SELECT count(*) INTO bad FROM transactions
+   WHERE foreign_currency_code IS NOT NULL AND foreign_currency_code !~ '^[A-Z]{3}$';
+  IF bad <> 0 THEN
+    RAISE EXCEPTION 'normalisation incomplete: % non-ISO codes remain', bad;
+  END IF;
+END $$;
+
+COMMIT;

prisma/migrations/0037_transaction_audit_log/migration.sql

@@ -0,0 +1,152 @@
+-- Append-only change log for the ledger and its curation children.
+--
+-- WHY THIS EXISTS: on 2026-08-23 the CC Statement Processor's `Run Reconciliation`
+-- node stamped `statement_id` onto seven statementless wallet/pantry rows without
+-- setting `reconciled_with_id` or superseding the statement's own extracted line.
+-- The purchase counted twice, and because finance-app's reconcile queue is
+-- `WHERE statement_id IS NULL`, the stamp also hid those rows from the only thing
+-- that could repair them. Answering "was anything lost?" required restoring the
+-- 02:30 pg_dump into a scratch database and diffing it column by column, because
+-- NOTHING in this estate records what changed. That is the gap this closes.
+--
+-- WHY AT THE DATABASE AND NOT IN THE APP: the ledger has at least five independent
+-- writers -- finance-app (Prisma), n8n (raw SQL in a Postgres node), pantry-app,
+-- the ingestion-engine runner (psycopg), and ad-hoc `docker exec psql`. The 4a bug
+-- was an n8n raw-SQL UPDATE; an application-level audit would not have seen it at
+-- all. A trigger cannot be bypassed by choice of client.
+--
+-- WHY NOT POSTGRES' OWN LOGS: checked live on this instance -- wal_level=replica
+-- (so WAL cannot be logically decoded into row changes), wal_keep_size=0 and
+-- archive_mode=off (so WAL is recycled almost immediately), log_statement=none,
+-- logging_collector=off, track_commit_timestamp=off, and pgaudit is not installed.
+-- Postgres was recording nothing usable. It has no built-in audit or temporal
+-- tables; this always has to be built.
+
+CREATE TABLE IF NOT EXISTS transaction_audit (
+  id             bigserial   PRIMARY KEY,
+  changed_at     timestamptz NOT NULL DEFAULT now(),
+  -- Groups every row touched by one logical edit. `applyReconcileMatch` moves
+  -- overrides, tags, splits and metadata in a single transaction; without this
+  -- they read as four unrelated events.
+  txid           bigint      NOT NULL DEFAULT txid_current(),
+  table_name     text        NOT NULL,
+  op             text        NOT NULL CHECK (op IN ('INSERT','UPDATE','DELETE')),
+  -- Full primary key as an object, because transaction_tags is composite
+  -- (transaction_id, tag_id) and the others are a bare `id`.
+  row_key        jsonb       NOT NULL,
+  -- Denormalised so "everything that ever happened to transaction N" is one
+  -- index scan rather than a five-way union.
+  transaction_id bigint,
+  -- On UPDATE these hold ONLY the columns that actually changed. Storing whole
+  -- rows would bury a one-column edit in forty unchanged fields, and
+  -- expense_metadata.line_items alone can be kilobytes.
+  old_data       jsonb,
+  new_data       jsonb,
+  -- Attribution is partial and that is understood: every client connects as the
+  -- same `personal` role. app_name is only as good as what each client sets;
+  -- client_addr distinguishes containers and is NULL for `docker exec` over the
+  -- unix socket, which is itself a useful signal.
+  db_user        text        NOT NULL DEFAULT current_user,
+  app_name       text,
+  client_addr    inet
+);
+
+COMMENT ON TABLE transaction_audit IS
+  'Append-only row-level change log for transactions + curation children. Never UPDATE or DELETE from this table.';
+
+CREATE INDEX IF NOT EXISTS transaction_audit_txn_time
+  ON transaction_audit (transaction_id, changed_at DESC);
+CREATE INDEX IF NOT EXISTS transaction_audit_time
+  ON transaction_audit (changed_at DESC);
+CREATE INDEX IF NOT EXISTS transaction_audit_txid
+  ON transaction_audit (txid);
+CREATE INDEX IF NOT EXISTS transaction_audit_table_time
+  ON transaction_audit (table_name, changed_at DESC);
+
+-- TG_ARGV[0]: comma-separated candidate columns for the owning transaction id;
+--             the first non-null wins (expense_metadata carries the id in
+--             transaction_id OR matched_transaction_id -- a card-settled order
+--             only ever has the latter).
+-- TG_ARGV[1]: comma-separated primary key columns.
+CREATE OR REPLACE FUNCTION log_transaction_change() RETURNS trigger
+LANGUAGE plpgsql SECURITY DEFINER
+SET search_path = public, pg_temp
+AS $fn$
+DECLARE
+  v_full jsonb;
+  v_old  jsonb;
+  v_new  jsonb;
+  v_key  jsonb := '{}'::jsonb;
+  v_txn  bigint;
+  k      text;
+BEGIN
+  IF TG_OP = 'INSERT' THEN
+    v_new  := to_jsonb(NEW);
+    v_full := v_new;
+  ELSIF TG_OP = 'DELETE' THEN
+    v_old  := to_jsonb(OLD);
+    v_full := v_old;
+  ELSE
+    v_old  := to_jsonb(OLD);
+    v_new  := to_jsonb(NEW);
+    v_full := v_new;
+    -- Reduce to the changed columns only. IS DISTINCT FROM so a NULL->NULL
+    -- comparison counts as unchanged rather than evaluating to NULL.
+    FOR k IN SELECT jsonb_object_keys(v_old) LOOP
+      IF (v_old -> k) IS NOT DISTINCT FROM (v_new -> k) THEN
+        v_old := v_old - k;
+        v_new := v_new - k;
+      END IF;
+    END LOOP;
+    -- A no-op UPDATE is noise, and finance-app's upserts issue plenty of them.
+    IF v_old = '{}'::jsonb THEN
+      RETURN NULL;
+    END IF;
+  END IF;
+
+  FOREACH k IN ARRAY string_to_array(TG_ARGV[1], ',') LOOP
+    v_key := v_key || jsonb_build_object(k, v_full -> k);
+  END LOOP;
+
+  FOREACH k IN ARRAY string_to_array(TG_ARGV[0], ',') LOOP
+    IF v_txn IS NULL AND (v_full ->> k) IS NOT NULL THEN
+      v_txn := (v_full ->> k)::bigint;
+    END IF;
+  END LOOP;
+
+  INSERT INTO transaction_audit
+    (table_name, op, row_key, transaction_id, old_data, new_data, app_name, client_addr)
+  VALUES
+    (TG_TABLE_NAME, TG_OP, v_key, v_txn, v_old, v_new,
+     NULLIF(current_setting('application_name', true), ''), inet_client_addr());
+
+  RETURN NULL;  -- AFTER trigger; return value is ignored
+END
+$fn$;
+
+-- AFTER, so a statement that raises never leaves a phantom audit row; the audit
+-- write is in the same transaction as the change and rolls back with it.
+DROP TRIGGER IF EXISTS trg_audit_transactions ON transactions;
+CREATE TRIGGER trg_audit_transactions
+  AFTER INSERT OR UPDATE OR DELETE ON transactions
+  FOR EACH ROW EXECUTE FUNCTION log_transaction_change('id', 'id');
+
+DROP TRIGGER IF EXISTS trg_audit_transaction_splits ON transaction_splits;
+CREATE TRIGGER trg_audit_transaction_splits
+  AFTER INSERT OR UPDATE OR DELETE ON transaction_splits
+  FOR EACH ROW EXECUTE FUNCTION log_transaction_change('transaction_id', 'id');
+
+DROP TRIGGER IF EXISTS trg_audit_transaction_tags ON transaction_tags;
+CREATE TRIGGER trg_audit_transaction_tags
+  AFTER INSERT OR UPDATE OR DELETE ON transaction_tags
+  FOR EACH ROW EXECUTE FUNCTION log_transaction_change('transaction_id', 'transaction_id,tag_id');
+
+DROP TRIGGER IF EXISTS trg_audit_transaction_overrides ON transaction_overrides;
+CREATE TRIGGER trg_audit_transaction_overrides
+  AFTER INSERT OR UPDATE OR DELETE ON transaction_overrides
+  FOR EACH ROW EXECUTE FUNCTION log_transaction_change('transaction_id', 'id');
+
+DROP TRIGGER IF EXISTS trg_audit_expense_metadata ON expense_metadata;
+CREATE TRIGGER trg_audit_expense_metadata
+  AFTER INSERT OR UPDATE OR DELETE ON expense_metadata
+  FOR EACH ROW EXECUTE FUNCTION log_transaction_change('transaction_id,matched_transaction_id', 'id');

prisma/migrations/0038_account_links/migration.sql

@@ -0,0 +1,76 @@
+-- Account continuity across a rebrand, reissue or issuer change.
+--
+-- WHY THIS EXISTS: one card can appear in `statements` under two completely
+-- different identities. The live case is the Citi Prestige card, which moved to
+-- NAB and was rebranded MyCard:
+--
+--   Citibank  5258073330340253  statements ... 2025-11-13
+--   NAB       1015001000116227  statements 2025-11-14 ...
+--
+-- Same 14th-to-13th billing cycle, no gap, one continuous product. Nothing in
+-- the schema linked them, so anything grouping by (bank_name, account) sees a
+-- dead account and a new one with no history. For the annual-fee detector that
+-- is not cosmetic: the $700 fee charged 2024-12-13 and 2025-11-13 reads as
+-- "stopped" when it is due again around November 2026 under the new name.
+--
+-- WHY A TABLE AND NOT AN INFERENCE: contiguous-cycle heuristics would join two
+-- genuinely different accounts as readily as one renamed one, and a wrong join
+-- merges two real fee series. There are only a handful of these events in a
+-- lifetime, each announced by mail, so each link is asserted once with the
+-- document that proves it. `evidence` is not decoration -- it is how a future
+-- reader checks the claim instead of trusting it.
+--
+-- NOT `account_owner_mappings`: that answers "whose account is this" and keys on
+-- the raw (bank_name, account_number). This answers "is this the same account as
+-- that one", and keys on the normalised form.
+
+CREATE TABLE IF NOT EXISTS account_links (
+  id                bigserial   PRIMARY KEY,
+  -- Normalised the same way statements.account_number_key is generated
+  -- (non-alphanumerics stripped), so these join without per-call cleanup.
+  -- Amex-style masks keep their X's: they record which digits were redacted.
+  from_bank_name    text        NOT NULL,
+  from_account_key  text        NOT NULL,
+  to_bank_name      text        NOT NULL,
+  to_account_key    text        NOT NULL,
+  -- When the successor took over. The predecessor's last statement and the
+  -- successor's first are the observable boundary; this is the announced date
+  -- where mail gives one.
+  effective_date    date,
+  reason            text        NOT NULL
+                      CHECK (reason IN ('rebrand', 'reissue', 'issuer_change', 'migration')),
+  -- Provenance for the claim, e.g. 'source_documents:32919'. Free text because
+  -- the evidence is not always a document -- sometimes it is a statement pair.
+  evidence          text,
+  notes             text,
+  created_at        timestamptz NOT NULL DEFAULT now(),
+
+  -- One successor per predecessor: a chain, never a fork. A fork would make
+  -- "the current identity of this account" ambiguous, and the resolver walks
+  -- this forward assuming it is not.
+  CONSTRAINT uq_account_links_from UNIQUE (from_bank_name, from_account_key),
+  -- A link to itself would make that walk non-terminating.
+  CONSTRAINT account_links_no_self CHECK (
+    from_bank_name IS DISTINCT FROM to_bank_name
+    OR from_account_key IS DISTINCT FROM to_account_key
+  )
+);
+
+COMMENT ON TABLE account_links IS
+  'Predecessor -> successor account identity across rebrands/reissues. Hand-asserted, each row carrying the evidence for its claim.';
+
+CREATE INDEX IF NOT EXISTS account_links_to
+  ON account_links (to_bank_name, to_account_key);
+
+-- The one link the mail record proves today.
+--   source_documents 32919 (2025-10-01) "your Citi Prestige Card account is moving to MyCard"
+--   source_documents 32618 (2025-11-10) "MyCard is coming soon"
+--   source_documents 32539 (2025-11-24) "The new MyCard App and MyCard Online are now live"
+-- Statement boundary: Citibank ends 2025-11-13, NAB begins 2025-11-14.
+INSERT INTO account_links
+  (from_bank_name, from_account_key, to_bank_name, to_account_key, effective_date, reason, evidence, notes)
+VALUES
+  ('Citibank', '5258073330340253', 'NAB', '1015001000116227', '2025-11-14', 'issuer_change',
+   'source_documents:32919,32618,32539',
+   'Citi Prestige -> MyCard Prestige. Billing cycle (14th-13th) continues unbroken across the boundary.')
+ON CONFLICT (from_bank_name, from_account_key) DO NOTHING;

prisma/migrations/0039_wallet_notification_time/migration.sql

@@ -0,0 +1,26 @@
+-- When, and from which app, a wallet notification arrived.
+--
+-- WHY: one tap emits two notifications 2-8 seconds apart -- the issuer app
+-- (card tail) and Apple Wallet (merchant, suburb) -- and wallet-ingestion.ts
+-- merges them. It used to recognise the pair by amount alone, within +/-1 DAY,
+-- which is wide enough to fold two different purchases together: on 2026-10-01
+-- an Amex "$50.00 at A MEDIUM" notification merged into the UNICEF $50 donation
+-- from 17 hours earlier and renamed it. A real pair is seconds apart and comes
+-- from two DIFFERENT apps; two notifications from one app are two charges.
+-- Neither fact was stored anywhere on the ledger side, so the matcher could not
+-- ask. These columns let it.
+--
+-- Backfilled from capture.wallet_events through source_message_id ('evt:<id>'),
+-- which every wallet-written metadata row carries.
+
+ALTER TABLE expense_metadata
+  ADD COLUMN IF NOT EXISTS notified_at timestamptz,
+  ADD COLUMN IF NOT EXISTS notif_app   text;
+
+UPDATE expense_metadata em
+   SET notified_at = e.received_at,
+       notif_app   = e.notif_app
+  FROM capture.wallet_events e
+ WHERE em.source = 'wallet'
+   AND em.source_message_id = 'evt:' || e.id
+   AND em.notified_at IS NULL;

prisma/migrations/0040_transaction_void/migration.sql

@@ -0,0 +1,43 @@
+-- A statementless row for a charge that never posted.
+--
+-- WHY: the wallet lane books every card notification as a purchase, and a card
+-- notification is sent for an AUTHORISATION, not a settlement. When the
+-- merchant releases the hold the bank never posts the charge, so no statement
+-- line ever arrives: the row sits in the reconcile queue forever and counts as
+-- spend in every figure. Txn 7951: Airbnb HMXAHAW49Y, $1,040.41 held on Amex
+-- 1000 on 2026-09-24, booking cancelled 25 Sep inside the free window, hold
+-- released ("We removed the $1,040.41 (AUD) payment hold"). Statement 219
+-- (11 Sep - 10 Oct) has no such line and reconciles to its header without it.
+--
+-- Neither existing exclusion fits. reconciled_with_id needs a statement twin
+-- and superseded_by_id needs a surviving row (and may not point at itself,
+-- transactions_no_self_supersede). There is no survivor: nothing happened.
+--
+-- Not a DELETE: every child of transactions cascades, and the row is also the
+-- sink the wallet lane checks before re-posting an event. Deleting it would let
+-- the next prompt_version bump re-create it.
+
+ALTER TABLE transactions
+  ADD COLUMN IF NOT EXISTS voided_at   timestamptz,
+  ADD COLUMN IF NOT EXISTS void_reason text;
+
+COMMENT ON COLUMN transactions.voided_at IS
+  'Set when this statementless row records a charge that never posted (e.g. a released authorisation hold). Excluded from every figure and from the reconcile queue, kept for its children and audit trail. NULL = live.';
+
+-- A reason is mandatory: a void with no explanation is indistinguishable from
+-- a mistake six months later.
+ALTER TABLE transactions
+  ADD CONSTRAINT transactions_void_has_reason
+  CHECK ((voided_at IS NULL AND void_reason IS NULL)
+      OR (voided_at IS NOT NULL AND btrim(coalesce(void_reason, '')) <> ''));
+
+-- Statement rows cannot be voided. A statement line is the bank's own record
+-- of a posted charge, and the statement's rows must keep summing to its
+-- header; a wrong statement row is a supersede, not a void.
+ALTER TABLE transactions
+  ADD CONSTRAINT transactions_void_statementless
+  CHECK (voided_at IS NULL OR statement_id IS NULL);
+
+CREATE INDEX IF NOT EXISTS idx_transactions_voided
+    ON transactions (voided_at)
+ WHERE voided_at IS NOT NULL;

prisma/schema.prisma

@@ -190,7 +190,9 @@ model transactions {
   owner_id                Int?
   reconciled_with_id      Int?
   superseded_by_id        Int?
-  principal_amount        Decimal?      @db.Decimal(12, 2)
+  voided_at               DateTime?     @db.Timestamptz(6) // a charge that never posted (migration 0040)
+  void_reason             String?
+  principal_amount       Decimal?      @db.Decimal(12, 2)
   interest_amount         Decimal?      @db.Decimal(12, 2)
   source                  String?       // feed this row came from, e.g. "frollo"; NULL = statement/manual (migration 0028)
   source_ref              String?       // the provider's own transaction id — idempotency key for re-imports
@@ -237,6 +239,9 @@ model expense_metadata {
   tender_raw             String?
   receipt_sha256         String?
   receipt_group          String?
+  // 0039 — when/where a wallet notification arrived; the pair merge needs both.
+  notified_at            DateTime?    @db.Timestamptz(6)
+  notif_app              String?
   transaction           transactions? @relation(fields: [transaction_id], references: [id], onDelete: Cascade)
 
   @@unique([source, order_reference], name: "uq_expense_source_order")

scripts/setup-test-db.sh

@@ -19,14 +19,23 @@ if ! docker inspect "$PG_CONTAINER" >/dev/null 2>&1; then
   exit 1
 fi
 
-# Create test database (ignore error if it already exists)
-docker exec "$PG_CONTAINER" psql -U "$PG_USER" -d postgres \
-  -c "CREATE DATABASE $TEST_DB;" 2>/dev/null || true
+# Drop and recreate, so the schema is rebuilt rather than replayed. Replaying
+# the dump over an existing database was a silent no-op for every migration:
+# each CREATE TABLE failed with "already exists" and the old shape survived, so
+# a new column (0039's expense_metadata.notified_at, 2026-10-03) never reached
+# the test database. The guard is because this now DROPs: it refuses anything
+# not named *_test, whatever TEST_DB was set to.
+case "$TEST_DB" in
+  *_test) ;;
+  *) echo "ERROR: refusing to drop '$TEST_DB' -- the test database must be named *_test"; exit 1 ;;
+esac
+docker exec "$PG_CONTAINER" psql -U "$PG_USER" -d postgres -q \
+  -c "DROP DATABASE IF EXISTS $TEST_DB WITH (FORCE);" \
+  -c "CREATE DATABASE $TEST_DB;"
 
-# Wipe and rebuild schema from production (schema only, no data)
 echo "Copying schema from $PG_USER to $TEST_DB..."
 docker exec "$PG_CONTAINER" pg_dump -U "$PG_USER" --schema-only "$PG_USER" \
-  | docker exec -i "$PG_CONTAINER" psql -U "$PG_USER" -d "$TEST_DB" -q
+  | docker exec -i "$PG_CONTAINER" psql -U "$PG_USER" -d "$TEST_DB" -q -v ON_ERROR_STOP=1
 
 echo "Schema ready."
 

src/__tests__/integration/analytics-sql.test.ts

@@ -105,6 +105,19 @@ describe("NATIVE_CURRENCY", () => {
     expect(unconverted).toBe(true);
   });
 
+  it("labels a statement-less wallet row AUD, because its amount already is", async () => {
+    // Wallet rows store a provisional AUD figure in amount AND amount_aud, and
+    // the native figure in foreign_currency_amount. Reading the foreign code
+    // here printed "INR 9.42" for an INR 524 tap (txn 8162).
+    const id = await scratchTxn(
+      "transaction_date, description, amount, amount_aud, transaction_type, payment_method, source, source_ref, foreign_currency_amount, foreign_currency_code",
+      "'2026-10-03','Wallet foreign fixture', 9.42, 9.42, 'debit', 'card', 'wallet', 'evt:fixture-native-ccy', 524.00, 'INR'"
+    );
+    const { ccy, unconverted } = await currencyOf(id);
+    expect(ccy).toBe("AUD");
+    expect(unconverted).toBe(false);
+  });
+
   it("prefers the statement's currency over the foreign-charge record", async () => {
     // Opposite convention: on an AUD statement, `amount` is AUD and
     // foreign_currency_code merely notes what was originally charged. Reading

src/__tests__/integration/annual-charges.test.ts

@@ -0,0 +1,170 @@
+import { describe, it, expect, beforeAll, beforeEach, vi } from "vitest";
+import { Pool } from "pg";
+import { createPool, resetDB, seedParticipants, mockDbWithPool } from "./helpers";
+
+/**
+ * Upcoming annual charges from payments alone.
+ *
+ * The exclusions carry this file. Anything can be made to find a card annual
+ * fee; the difficulty is not also reporting the monthly account-servicing fee,
+ * the per-transfer Wise charge, and the restaurant visited twice a year apart.
+ * Measured on the real ledger, a naive "two charges ~365 days apart" ranks a
+ * bakery first with 202 such pairs.
+ */
+
+let pool: Pool;
+let ownerId: number;
+let getUpcomingAnnualCharges: typeof import("../../lib/annual-charges").getUpcomingAnnualCharges;
+
+let acctSeq = 0;
+
+/** A statement plus one transaction on it, for a named account. */
+async function charge(
+  bank: string,
+  acct: string,
+  date: string,
+  description: string,
+  amount: number,
+  type = "fee"
+) {
+  const key = `${bank}:${acct}`;
+  let stId = stmtCache.get(`${key}:${date}`);
+  if (!stId) {
+    const st = await pool.query(
+      `INSERT INTO statements (filename, account_number, bank_name, billing_start_date, billing_end_date, owner_id)
+       VALUES ($4, $1, $2, $3::date - 20, $3::date + 10, $5) RETURNING id`,
+      [acct, bank, date, `s${acctSeq++}.pdf`, ownerId]
+    );
+    stId = st.rows[0].id as number;
+    stmtCache.set(`${key}:${date}`, stId);
+  }
+  await pool.query(
+    `INSERT INTO transactions (statement_id, transaction_date, description, amount, transaction_type, owner_id)
+     VALUES ($1,$2,$3,$4,$5,$6)`,
+    [stId, date, description, amount, type, ownerId]
+  );
+}
+let stmtCache = new Map<string, number>();
+
+/** A statementless merchant purchase, for the non-fee signal. */
+async function purchase(date: string, merchant: string, amount: number) {
+  await pool.query(
+    `INSERT INTO transactions (statement_id, transaction_date, description, amount, transaction_type, owner_id, merchant_normalized, row_index)
+     VALUES (NULL,$1,$2,$3,'debit',$4,$2,0)`,
+    [date, merchant, amount, ownerId]
+  );
+}
+
+const find = <T extends { label: string }>(rows: T[], label: string) =>
+  rows.find((r) => r.label === label);
+
+beforeAll(async () => {
+  pool = createPool();
+  mockDbWithPool(pool);
+  ({ getUpcomingAnnualCharges } = await import("../../lib/annual-charges"));
+});
+
+beforeEach(async () => {
+  await resetDB(pool);
+  await pool.query("DELETE FROM account_links");
+  stmtCache = new Map();
+  ({ ownerId } = await seedParticipants(pool));
+});
+
+describe("what it finds", () => {
+  it("confirms a card annual fee that repeated a year later", async () => {
+    await charge("American Express", "1000", "2025-06-10", "ANNUAL CARD FEE", 875);
+    await charge("American Express", "1000", "2026-06-10", "ANNUAL CARD FEE", 1750);
+    const row = find(await getUpcomingAnnualCharges(ownerId), "ANNUAL CARD FEE");
+    expect(row?.confidence).toBe("confirmed");
+    expect(row?.observations).toBe(2);
+    // The amount is NOT averaged: this fee doubled, and the next one to plan for
+    // is the latest, not the mean of a rise.
+    expect(Number(row?.last_amount)).toBe(1750);
+    expect(row?.next_due_estimate.startsWith("2027-06")).toBe(true);
+  });
+
+  it("accepts a fee seen only once when the bank names it annual", async () => {
+    await charge("Westpac", "8032", "2025-10-21", "CARD FEE", 295);
+    const row = find(await getUpcomingAnnualCharges(ownerId), "CARD FEE");
+    // One charge cannot show a cadence, but the kind is not in doubt.
+    expect(row?.confidence).toBe("likely");
+    expect(row?.next_due_estimate).toBe("2026-10-21");
+  });
+
+  it("finds an annual merchant charge with no fee wording, if the amount is stable", async () => {
+    await purchase("2025-02-03", "AMAZON PRIME", 79);
+    await purchase("2026-02-01", "AMAZON PRIME", 79);
+    const row = find(await getUpcomingAnnualCharges(ownerId), "AMAZON PRIME");
+    expect(row?.confidence).toBe("possible");
+  });
+});
+
+describe("what it must NOT report", () => {
+  it("ignores a monthly account fee", async () => {
+    for (const d of ["2026-01-31", "2026-02-28", "2026-03-31", "2026-04-30"]) {
+      await charge("ANZ", "5626", d, "ACCOUNT SERVICING FEE", 5);
+    }
+    expect(find(await getUpcomingAnnualCharges(ownerId), "ACCOUNT SERVICING FEE")).toBeUndefined();
+  });
+
+  it("ignores per-transaction fees, which never repeat under one label", async () => {
+    await charge("Wise Australia Pty Ltd.", "9001", "2026-05-20", "WISE CHARGES FOR: TRANSFER-2143509822", 30.67);
+    await charge("Wise Australia Pty Ltd.", "9001", "2026-06-04", "WISE CHARGES FOR: TRANSFER-2171403032", 19.51);
+    const rows = await getUpcomingAnnualCharges(ownerId);
+    expect(rows.filter((r) => r.label.startsWith("WISE CHARGES"))).toHaveLength(0);
+  });
+
+  it("ignores a merchant visited twice a year apart for wildly different amounts", async () => {
+    // The exact shape that defeats a pure cadence test: Target, $30 then $198.
+    await purchase("2025-07-21", "TARGET", 30);
+    await purchase("2026-07-11", "TARGET", 198);
+    expect(find(await getUpcomingAnnualCharges(ownerId), "TARGET")).toBeUndefined();
+  });
+
+  it("ignores a merchant seen only once", async () => {
+    await purchase("2026-04-03", "UBER ONE MEMBERSHIP", 96);
+    expect(find(await getUpcomingAnnualCharges(ownerId), "UBER ONE MEMBERSHIP")).toBeUndefined();
+  });
+});
+
+describe("account identity and honesty about gaps", () => {
+  it("treats a renamed account as one continuous fee series", async () => {
+    // Citi Prestige -> NAB MyCard: same card, two bank names, two numbers.
+    await charge("Citibank", "5258073330340253", "2024-12-13", "Annual Fee", 700);
+    await charge("Citibank", "5258073330340253", "2025-11-13", "Annual Fee", 700);
+    await pool.query(
+      `INSERT INTO account_links (from_bank_name, from_account_key, to_bank_name, to_account_key, effective_date, reason)
+       VALUES ('Citibank','5258073330340253','NAB','1015001000116227','2025-11-14','issuer_change')`
+    );
+    // A statement under the NEW identity, so the link has somewhere to resolve to.
+    await charge("NAB", "1015001000116227", "2026-08-13", "Purchase", 29.99, "debit");
+
+    const row = find(await getUpcomingAnnualCharges(ownerId), "ANNUAL FEE");
+    expect(row?.observations).toBe(2);
+    expect(row?.confidence).toBe("confirmed");
+    // Names the account's CURRENT bank -- the next fee comes from there, and
+    // sending someone to the retired brand is the bug this link exists to stop.
+    expect(row?.bank_name).toBe("NAB");
+  });
+
+  it("marks a passed estimate as expected_not_seen rather than upcoming", async () => {
+    await charge("HSBC", "0351", "2024-03-06", "ANNUAL CHARGE", 450);
+    await charge("HSBC", "0351", "2025-03-06", "ANNUAL CHARGE", 450);
+    const row = find(await getUpcomingAnnualCharges(ownerId), "ANNUAL CHARGE");
+    expect(row?.status).toBe("expected_not_seen");
+    // And it reports how stale the account is, which is what separates
+    // "it stopped being charged" from "we stopped importing the statements".
+    expect(row?.account_days_stale).toBeGreaterThan(300);
+  });
+
+  it("sorts upcoming ahead of expected_not_seen", async () => {
+    await charge("HSBC", "0351", "2024-03-06", "ANNUAL CHARGE", 450);
+    await charge("HSBC", "0351", "2025-03-06", "ANNUAL CHARGE", 450);
+    await charge("Westpac", "8032", "2026-08-01", "CARD FEE", 295);
+    const rows = await getUpcomingAnnualCharges(ownerId);
+    const firstMissing = rows.findIndex((r) => r.status === "expected_not_seen");
+    const lastUpcoming = rows.map((r) => r.status).lastIndexOf("upcoming");
+    expect(lastUpcoming).toBeLessThan(firstMissing);
+  });
+});

src/__tests__/integration/auto-reconcile.test.ts

@@ -0,0 +1,271 @@
+import { describe, it, expect, beforeAll, beforeEach, vi } from "vitest";
+import { Pool } from "pg";
+import { createPool, resetDB, seedParticipants, insertTransaction } from "./helpers";
+
+/**
+ * The automatic pass over the reconcile queue.
+ *
+ * This exists because the wallet lane writes a transaction for every tap
+ * (DECISIONS.md ING-11), adding ~63 rows a month to a queue already ~177 deep
+ * and worked by hand. Hiding those rows instead is explicitly ruled out --
+ * see the awaitsStatementLine() removal note in queries.ts.
+ *
+ * The cases that must NOT match matter more than the ones that must: this runs
+ * unattended, and a wrong pairing merges one purchase's curation onto another
+ * purchase's row. So most of what follows is about refusal.
+ */
+
+const MACHINE_TOKEN = "test-wallet-token";
+let pool: Pool;
+let ownerId: number;
+let autoReconcile: typeof import("../../app/api/transactions/reconcile/route").PATCH;
+
+/** Browser-session caller: no machine token, auth comes from the mocked session. */
+const userReq = (url = "http://x/api/transactions/reconcile") =>
+  ({ url, headers: { get: () => null } }) as never;
+
+/** Scheduled caller: machine token, no session, sweeps every owner. */
+const machineReq = (url = "http://x/api/transactions/reconcile") =>
+  ({ url, headers: { get: (h: string) => (h === "x-ingest-token" ? MACHINE_TOKEN : null) } }) as never;
+
+async function statementLine(
+  amount: number,
+  date: string,
+  description: string,
+  opts: { owner?: number; type?: string; fxAmount?: number; fxCode?: string } = {}
+) {
+  const owner = opts.owner ?? ownerId;
+  const st = await pool.query(
+    // Account number varies by owner: uq_statement_identity is
+    // (bank_name, account_number, billing_end_date), so two owners sharing one
+    // account collide rather than modelling two real cards.
+    `INSERT INTO statements (filename, account_number, bank_name, billing_start_date, billing_end_date, owner_id)
+     VALUES ('s.pdf', $3, 'Test Bank', $1::date - 20, $1::date + 10, $2) RETURNING id`,
+    [date, owner, `1234-5678-9012-${String(owner).padStart(4, "0")}`]
+  );
+  const t = await pool.query(
+    `INSERT INTO transactions
+       (statement_id, transaction_date, description, amount, transaction_type, owner_id,
+        foreign_currency_amount, foreign_currency_code)
+     VALUES ($1, $2, $3, $4, $5, $6, $7, $8) RETURNING id`,
+    [st.rows[0].id, date, description, amount, opts.type ?? "debit", owner,
+     opts.fxAmount ?? null, opts.fxCode ?? null]
+  );
+  return t.rows[0].id as number;
+}
+
+async function walletRow(
+  amount: number,
+  date: string,
+  opts: { fxAmount?: number; fxCode?: string; paymentMethod?: string } = {}
+) {
+  const id = await insertTransaction(pool, ownerId, {
+    amount, transaction_date: date, description: "Wallet tap",
+  });
+  await pool.query(
+    `UPDATE transactions
+        SET foreign_currency_amount = $2, foreign_currency_code = $3, payment_method = $4
+      WHERE id = $1`,
+    [id, opts.fxAmount ?? null, opts.fxCode ?? null, opts.paymentMethod ?? null]
+  );
+  return id;
+}
+
+const reconciledWith = async (id: number) =>
+  (await pool.query(`SELECT reconciled_with_id FROM transactions WHERE id = $1`, [id]))
+    .rows[0].reconciled_with_id as number | null;
+
+beforeAll(async () => {
+  process.env.WALLET_INGEST_TOKEN = MACHINE_TOKEN;
+  pool = createPool();
+  vi.doMock("@/lib/auth", () => ({
+    getCurrentUser: async () => ({ id: ownerId, name: "Alice", email: "alice@example.com" }),
+  }));
+  ({ PATCH: autoReconcile } = await import("../../app/api/transactions/reconcile/route"));
+});
+
+beforeEach(async () => {
+  await resetDB(pool);
+  await pool.query("DELETE FROM expense_metadata");
+  ({ ownerId } = await seedParticipants(pool));
+});
+
+describe("pairs it should apply", () => {
+  it("matches an exact amount two days apart and moves the link", async () => {
+    const manual = await walletRow(13.5, "2026-08-13");
+    const stmt = await statementLine(13.5, "2026-08-15", "BREADWORLD BAKERY");
+
+    const body = await (await autoReconcile(userReq())).json();
+    expect(body.reconciled).toBe(1);
+    expect(await reconciledWith(manual)).toBe(stmt);
+  });
+
+  it("matches a foreign row on the NATIVE pair, not on the AUD amount", async () => {
+    // The bank converts at its own rate, so the wallet row cannot predict the
+    // AUD figure. It can only know THB 500. Deliberately give the statement a
+    // wildly different AUD amount to prove `amount` is not what matched.
+    const manual = await walletRow(999.99, "2026-08-13", { fxAmount: 500, fxCode: "THB" });
+    const stmt = await statementLine(22.85, "2026-08-14", "7-ELEVEN BANGKOK", {
+      fxAmount: 500, fxCode: "THB",
+    });
+
+    const body = await (await autoReconcile(userReq())).json();
+    expect(body.reconciled).toBe(1);
+    expect(await reconciledWith(manual)).toBe(stmt);
+  });
+
+  it("carries a SETTLED split across without resurrecting the debt", async () => {
+    const manual = await walletRow(50, "2026-08-13");
+    const stmt = await statementLine(50, "2026-08-14", "SHARED DINNER");
+    await pool.query(
+      `INSERT INTO transaction_splits (transaction_id, participant_id, share_percent, settled, settled_at)
+       VALUES ($1, $2, 50, true, now())`,
+      [manual, ownerId]
+    );
+
+    expect((await autoReconcile(userReq())).status ?? 200).toBe(200);
+
+    const moved = await pool.query(
+      `SELECT settled, settled_at FROM transaction_splits WHERE transaction_id = $1`, [stmt]
+    );
+    expect(moved.rowCount).toBe(1);
+    // The whole point: a delete-and-recreate defaults `settled` to false, which
+    // silently marks a paid debt owed again. Nothing errors when it regresses.
+    expect(moved.rows[0].settled).toBe(true);
+    expect(moved.rows[0].settled_at).not.toBeNull();
+  });
+});
+
+describe("pairs it must refuse", () => {
+  it("refuses two identical amounts in the same window (ambiguous both ways)", async () => {
+    // Two $13.50 bakery taps in one week is the ordinary case, not an edge one.
+    const a = await walletRow(13.5, "2026-08-13");
+    const b = await walletRow(13.5, "2026-08-14");
+    await statementLine(13.5, "2026-08-14", "BREADWORLD ONE");
+    await statementLine(13.5, "2026-08-15", "BREADWORLD TWO");
+
+    const body = await (await autoReconcile(userReq())).json();
+    expect(body.reconciled).toBe(0);
+    expect(await reconciledWith(a)).toBeNull();
+    expect(await reconciledWith(b)).toBeNull();
+  });
+
+  it("refuses a near-miss amount — a person should look at that", async () => {
+    const manual = await walletRow(13.5, "2026-08-13");
+    await statementLine(13.55, "2026-08-14", "BREADWORLD BAKERY");
+
+    expect((await (await autoReconcile(userReq())).json()).reconciled).toBe(0);
+    expect(await reconciledWith(manual)).toBeNull();
+  });
+
+  it("refuses a match outside the day window", async () => {
+    const manual = await walletRow(13.5, "2026-08-13");
+    await statementLine(13.5, "2026-08-25", "BREADWORLD BAKERY");
+
+    expect((await (await autoReconcile(userReq())).json()).reconciled).toBe(0);
+    expect(await reconciledWith(manual)).toBeNull();
+  });
+
+  it("refuses a statement row already claimed by another manual row", async () => {
+    const claimed = await statementLine(40, "2026-08-14", "ALREADY TAKEN");
+    const first = await walletRow(40, "2026-08-13");
+    await pool.query(`UPDATE transactions SET reconciled_with_id = $2 WHERE id = $1`, [first, claimed]);
+    const second = await walletRow(40, "2026-08-14");
+
+    expect((await (await autoReconcile(userReq())).json()).reconciled).toBe(0);
+    expect(await reconciledWith(second)).toBeNull();
+  });
+
+  it("refuses a credits-funded row — no card leg is ever coming", async () => {
+    const manual = await walletRow(30, "2026-08-13", { paymentMethod: "credits" });
+    await statementLine(30, "2026-08-14", "COINCIDENTAL AMOUNT");
+
+    expect((await (await autoReconcile(userReq())).json()).reconciled).toBe(0);
+    expect(await reconciledWith(manual)).toBeNull();
+  });
+
+});
+
+describe("an AUD row against a foreign statement line", () => {
+  it("DOES match on the AUD figure — this is what every real pair looks like", async () => {
+    // Checked against live data: all 48 reconciliations a human has made are
+    // this exact shape -- manual row with no currency, statement row carrying
+    // EUR or CHF, paired on exact AUD. Overseas receipts and orders record what
+    // was charged in AUD; only the wallet lane records natively. An earlier
+    // version of this rule required both sides to be non-foreign, on the
+    // reasoning that a currency mismatch is ambiguous. It matched 0 of the 48.
+    const manual = await walletRow(22.85, "2026-08-13");
+    const stmt = await statementLine(22.85, "2026-08-14", "7-ELEVEN BANGKOK", {
+      fxAmount: 500, fxCode: "THB",
+    });
+
+    expect((await (await autoReconcile(userReq())).json()).reconciled).toBe(1);
+    expect(await reconciledWith(manual)).toBe(stmt);
+  });
+
+  it("still refuses when a second candidate makes it ambiguous", async () => {
+    // The coincidence this guards against -- a domestic tap whose amount equals
+    // a foreign charge's AUD figure -- is caught by unique-both-ways, not by a
+    // currency test.
+    const manual = await walletRow(22.85, "2026-08-13");
+    await statementLine(22.85, "2026-08-14", "7-ELEVEN BANGKOK", { fxAmount: 500, fxCode: "THB" });
+    await statementLine(22.85, "2026-08-13", "MELBOURNE CAFE");
+
+    expect((await (await autoReconcile(userReq())).json()).reconciled).toBe(0);
+    expect(await reconciledWith(manual)).toBeNull();
+  });
+});
+
+describe("dry run", () => {
+  it("reports what it would do and writes nothing", async () => {
+    const manual = await walletRow(13.5, "2026-08-13");
+    await statementLine(13.5, "2026-08-14", "BREADWORLD BAKERY");
+
+    const body = await (await autoReconcile(userReq("http://x/api/transactions/reconcile?dry=1"))).json();
+    expect(body.dry_run).toBe(true);
+    expect(body.would_reconcile).toBe(1);
+    expect(await reconciledWith(manual)).toBeNull();
+  });
+});
+
+describe("the scheduled sweep", () => {
+  it("runs on the machine token with no session at all", async () => {
+    const manual = await walletRow(13.5, "2026-08-13");
+    const stmt = await statementLine(13.5, "2026-08-14", "BREADWORLD BAKERY");
+
+    const body = await (await autoReconcile(machineReq())).json();
+    expect(body.reconciled).toBe(1);
+    expect(await reconciledWith(manual)).toBe(stmt);
+  });
+
+  it("covers EVERY owner, not just one", async () => {
+    // 169 of the 177 rows outstanding belong to the second participant, so a
+    // sweep scoped to a single user would leave almost all of the queue alone.
+    const other = (await pool.query(
+      `INSERT INTO participants (name) VALUES ('Carol') RETURNING id`
+    )).rows[0].id as number;
+
+    const mine = await walletRow(11, "2026-08-13");
+    const theirTx = (await pool.query(
+      `INSERT INTO transactions (owner_id, statement_id, transaction_date, description,
+         amount, transaction_type, category, row_index)
+       VALUES ($1, NULL, '2026-08-13', 'Theirs', 22, 'debit', 'other', 0) RETURNING id`, [other]
+    )).rows[0].id as number;
+
+    await statementLine(11, "2026-08-14", "MINE");
+    await statementLine(22, "2026-08-14", "THEIRS", { owner: other });
+
+    const body = await (await autoReconcile(machineReq())).json();
+    expect(body.reconciled).toBe(2);
+    expect(await reconciledWith(mine)).not.toBeNull();
+    expect(await reconciledWith(theirTx)).not.toBeNull();
+  });
+
+  it("rejects a wrong machine token and falls through to session auth", async () => {
+    const bad = ({ url: "http://x/", headers: { get: () => "not-the-token" } }) as never;
+    // Session is mocked, so this still succeeds -- but as the session user only,
+    // proving the bad token did not grant the all-owner sweep.
+    const body = await (await autoReconcile(bad)).json();
+    expect(body.reconciled).toBe(0);
+  });
+});

src/__tests__/integration/helpers.ts

@@ -84,6 +84,11 @@ export async function resetDB(pool: Pool) {
   await assertTestDatabase(pool);
   await pool.query(`
     TRUNCATE
+      -- The audit triggers (migration 0037) fire on every INSERT the suite makes.
+      -- TRUNCATE does not fire row-level triggers, so nothing here is logged --
+      -- but without this line the log is never cleared either, and the table
+      -- grows by roughly a thousand rows per full run, forever.
+      transaction_audit,
       split_payments,
       transaction_splits,
       transaction_tags,

src/__tests__/integration/intake.test.ts

@@ -105,6 +105,17 @@ async function seedOrder(over: { merchant: string; date: string; amount: number;
   }
   return id;
 }
+/** A row shaped the way the wallet lane writes one: statementless, no metadata
+ *  at all, no platform, and named by the notification rather than the receipt. */
+async function seedWalletRow(over: { merchant: string; date: string; amount: number; owner?: number }) {
+  const t = await pool.query(
+    `INSERT INTO transactions (owner_id, statement_id, transaction_date, description, amount, amount_aud, transaction_type, category, merchant_name, merchant_normalized, payment_method, location, source, source_ref, row_index)
+     VALUES ($1, NULL, $2, $3, $4, $4, 'debit', 'other', $3, $3, 'card', 'Werribee, VIC', 'wallet', $5, 0) RETURNING id`,
+    [over.owner ?? ownerId, over.date, over.merchant, over.amount, `evt:${Math.floor(over.amount * 100)}`]
+  );
+  return t.rows[0].id as number;
+}
+
 /** What `components/order-details.tsx` would actually render for a transaction. */
 const orderPanel = async (txnId: number) =>
   (await (await orderGET({ headers: { get: () => null } } as unknown as NextRequest, { params: Promise.resolve({ id: String(txnId) }) })).json()) as { line_items?: unknown[] } | null;
@@ -339,6 +350,64 @@ describe("POST /api/intake — gates", () => {
     expect((await batch(id)).status).toBe("unsupported");
     expect(calls).toHaveLength(0);
   });
+
+  // 2026-08-22, live batch 3d2ffbe8…: an iOS Shortcut coerced the shared photo
+  // to text, so the whole upload was the five characters `Image`. The model was
+  // asked anyway and returned Burgertory (Braybrook), two "The Pride" lines and
+  // a $41.15 total — none of which existed. Only the absent date kept it out of
+  // the matcher, and the invented header was stored on the batch as
+  // `needs_header`. The model must never be asked.
+  it("a text part that cannot be a receipt is declined BEFORE any model call", async () => {
+    const id = uuid();
+    modelReply = { ...burgertory, merchant: "Burgertory (Braybrook)", order_date: null, total: 41.15 };
+    const res = await intakePOST(
+      multipart({ source: "other", batch_id: id }, [{ name: "shared.txt", type: "text/plain", bytes: Buffer.from("Image") }])
+    );
+    const body = await res.json();
+    expect(body).toMatchObject({ kind: "unsupported", status: "unsupported" });
+    expect(body.text).toMatch(/single word/);
+    expect(body.text).toMatch(/"Image"/);
+    // The model was never asked — the only thing that makes confabulation impossible.
+    expect(calls).toHaveLength(0);
+    const row = await batch(id);
+    expect(row.status).toBe("unsupported");
+    expect(row.kind).toBe("unsupported");
+    expect(row.header).toBeNull();
+    expect(row.items).toBeNull();
+    expect(row.transaction_id).toBeNull();
+    expect((await pool.query(`SELECT count(*)::int AS n FROM transactions`)).rows[0].n).toBe(0);
+  });
+
+  it("a junk text part beside a real receipt does not sink the batch", async () => {
+    const txn = await seedOrder({ merchant: "Burgertory (Werribee)", date: "2026-08-19", amount: 22.4 });
+    const body = await (
+      await intakePOST(
+        multipart({ source: "uber_eats" }, [
+          { name: "shared.txt", type: "text/plain", bytes: Buffer.from("Image") },
+          { name: "Receipt_19Aug2026.pdf", type: "application/pdf", bytes: BURGERTORY },
+        ])
+      )
+    ).json();
+    expect(body.status).toBe("enriched");
+    expect(await lineItemsOf(txn)).toHaveLength(2);
+  });
+
+  it("an order extracted from an unreadable PDF is unsupported, not needs_header", async () => {
+    const id = uuid();
+    // A PDF that renders to a scrap of junk still reaches the model — the
+    // pre-check only inspects text parts. `confabulationReason` is the belt.
+    vi.spyOn(await import("../../lib/intake-extract"), "extractPdfText").mockResolvedValueOnce("Image");
+    modelReply = { ...burgertory, merchant: "Burgertory (Braybrook)" };
+    const body = await (
+      await intakePOST(multipart({ source: "uber_eats", batch_id: id }, [{ name: "x.pdf", type: "application/pdf", bytes: BURGERTORY }]))
+    ).json();
+    expect(body).toMatchObject({ kind: "unsupported", status: "unsupported" });
+    expect(body.text).toMatch(/nothing shared could be read/);
+    const row = await batch(id);
+    expect(row.status).toBe("unsupported");
+    expect(row.header).toBeNull();
+    expect(row.items).toBeNull();
+  });
 });
 
 describe("POST /api/intake — food (enrich vs candidates)", () => {
@@ -483,7 +552,13 @@ describe("POST /api/intake — food (enrich vs candidates)", () => {
     const id = uuid();
     const body = await (await intakePOST(multipart({ source: "uber_eats", batch_id: id }, [pdf]))).json();
     expect(body.status).toBe("needs_header");
-    expect(body.text).toMatch(/share the top of the screen/);
+    // Asserts BOTH halves of needsHeaderCard(): that it names what is actually
+    // missing (total, not a blanket "total or date"), and that a uber_eats
+    // source gets the share-more-of-the-screen advice rather than the in-store
+    // Share-button advice. 258fe3b changed this text and updated the unit tests
+    // but not this one.
+    expect(body.text).toMatch(/I can see items but no total/);
+    expect(body.text).toMatch(/share the part of the screen that shows it/);
     expect(await lineItemsOf(txn)).toEqual([]);
     expect((await batch(id)).status).toBe("needs_header");
   });
@@ -574,6 +649,33 @@ describe("POST /api/intake — grocery → pantry", () => {
     expect(pantryCall.intake).toMatchObject({ platform: null, finance_target_transaction_id: null, finance_action: "create" });
   });
 
+  it("pantry saying the shop is already on its board is SHOWN, not swallowed", async () => {
+    // Pantry accepts a second capture of one receipt on purpose (the first may
+    // have been a bad crop) and reports it rather than refusing. Dropping that
+    // field turned "reported" into silence — and a duplicate nobody sees is a
+    // duplicate that gets confirmed, which counts the same groceries twice
+    // because pantry's addItem accumulates. Seen live on captures 66/69.
+    modelReply = coles;
+    pantryReply = { status: 200, body: { captureId: 902, items: [1, 2, 3], alreadyCaptured: { captureId: 66, confirmed: true } } };
+    const body = await (await intakePOST(multipart({ source: "coles" }, [pdf]))).json();
+    expect(body.status).toBe("forwarded_to_pantry");
+    expect(body.text).toMatch(/Already on the pantry board/);
+    expect(body.text).toMatch(/capture 66 \(already confirmed\)/);
+    expect(body.text).toMatch(/capture 902/);
+  });
+
+  it("two ledger rows for one shop are reported, not treated as no match", async () => {
+    // Ambiguity is not a miss. This is the exact state that produced the
+    // 2026-08-18 Woolworths double-book: the shop was in the ledger twice, the
+    // matcher returned nothing, and the card said finance would book it.
+    modelReply = coles;
+    await seedOrder({ merchant: "Coles (Wyndham Vale)", date: "2026-08-18", amount: 85.74, platform: null });
+    await seedOrder({ merchant: "Coles (Wyndham Vale)", date: "2026-08-18", amount: 85.74, platform: null });
+    const body = await (await intakePOST(multipart({ source: "coles" }, [pdf]))).json();
+    expect(body.status).toBe("forwarded_to_pantry");
+    expect(body.text).toMatch(/found more than one row it could be/);
+  });
+
   it("pantry 409 → duplicate", async () => {
     modelReply = coles;
     pantryReply = { status: 409, body: { error: "This receipt has already been added to the pantry.", captureId: 77 } };
@@ -647,6 +749,37 @@ describe("Slack taps (C5)", () => {
     expect(lastSlackUpdate().text).toMatch(/^Added 2 items/);
   });
 
+  it("intake_create adopts the wallet row for the same tap instead of booking it twice", async () => {
+    // The normal arrival order: the notification lands within seconds of the
+    // tap, the screenshot is shared later. matchTransaction cannot see that row
+    // — its key includes the platform and a notification has none — so before
+    // this, Create booked the purchase a second time.
+    const wallet = await seedWalletRow({ merchant: "Burgertory", date: "2026-08-19", amount: 22.4 });
+    const id = uuid();
+    const body = await (await intakePOST(multipart({ source: "uber_eats", batch_id: id }, [pdf]))).json();
+    expect(body.status).toBe("candidates");
+
+    await slackPOST(slackPress("intake_create", { batch_id: id }));
+
+    expect((await pool.query(`SELECT count(*)::int AS n FROM transactions`)).rows[0].n).toBe(1);
+    expect(await batch(id)).toMatchObject({ status: "enriched", transaction_id: wallet });
+    expect((await lineItemsOf(wallet)).map((i: { description: string }) => i.description)).toEqual(["Michael", "Joel"]);
+    // The suburb the notification supplied is the one thing nothing else knows.
+    expect((await pool.query(`SELECT location FROM transactions WHERE id = $1`, [wallet])).rows[0].location).toBe("Werribee, VIC");
+  });
+
+  it("intake_create still creates when a same-amount row already carries a receipt", async () => {
+    // A row that another source fully described is a DIFFERENT purchase that
+    // happens to share an amount — adopting it would move one order's items
+    // onto another order's row.
+    const other = await seedOrder({ merchant: "Someone Else", date: "2026-08-19", amount: 22.4, platform: null });
+    await pool.query(`UPDATE expense_metadata SET line_items = $2::jsonb WHERE transaction_id = $1`, [other, JSON.stringify([{ description: "theirs", qty: 1, amount: 22.4 }])]);
+    const id = uuid();
+    await intakePOST(multipart({ source: "uber_eats", batch_id: id }, [pdf]));
+    await slackPOST(slackPress("intake_create", { batch_id: id }));
+    expect((await pool.query(`SELECT count(*)::int AS n FROM transactions`)).rows[0].n).toBe(2);
+  });
+
   it("two pending cards for one order book it ONCE", async () => {
     const a = uuid();
     const b = uuid();

src/__tests__/integration/merge-supersede.test.ts

@@ -0,0 +1,175 @@
+import { describe, it, expect, beforeAll, beforeEach, vi } from "vitest";
+import { Pool } from "pg";
+import { createPool, resetDB, seedParticipants, insertTransaction } from "./helpers";
+
+/**
+ * Manually resolving a duplicate the automation cannot see.
+ *
+ * Until this route existed there was no way to merge or remove a transaction at
+ * all: `superseded_by_id` was read everywhere and written nowhere in
+ * application code, having been set exactly once by migration 0023. The only
+ * merge path was the reconcile queue, whose candidate list is bounded by an
+ * amount band -- so SP PEAKDESIGN (notified $42.44, charged $43.71, the card's
+ * 3% foreign transaction fee) had no offerable twin.
+ *
+ * The refusals matter more than the happy paths, exactly as in
+ * auto-reconcile.test.ts: this writes to columns that hide a row from every
+ * figure, so a wrong call does not error, it quietly stops counting money.
+ */
+
+let pool: Pool;
+let ownerId: number;
+let otherId: number;
+let merge: typeof import("../../app/api/transactions/[id]/merge/route").POST;
+
+const req = (body: unknown) => ({ json: async () => body, headers: { get: () => null } }) as never;
+const call = (id: number, body: unknown) => merge(req(body), { params: Promise.resolve({ id: String(id) }) });
+
+// uq_statement_identity is (bank_name, account_number, billing_end_date), so
+// every statement in a test needs its own account or the second insert collides.
+let acctSeq = 0;
+
+async function statementLine(amount: number, date: string, description: string) {
+  const st = await pool.query(
+    `INSERT INTO statements (filename, account_number, bank_name, billing_start_date, billing_end_date, owner_id)
+     VALUES ('s.pdf', $3, 'Test Bank', $1::date - 20, $1::date + 10, $2) RETURNING id`,
+    [date, ownerId, `4444-0000-0000-${String(1000 + acctSeq++)}`]
+  );
+  const t = await pool.query(
+    `INSERT INTO transactions (statement_id, transaction_date, description, amount, transaction_type, owner_id)
+     VALUES ($1, $2, $3, $4, 'debit', $5) RETURNING id`,
+    [st.rows[0].id, date, description, amount, ownerId]
+  );
+  return t.rows[0].id as number;
+}
+
+beforeAll(async () => {
+  pool = createPool();
+  vi.doMock("@/lib/auth", () => ({
+    getCurrentUser: async () => ({ id: ownerId, name: "Alice", email: "alice@example.com" }),
+  }));
+  ({ POST: merge } = await import("../../app/api/transactions/[id]/merge/route"));
+});
+
+beforeEach(async () => {
+  await resetDB(pool);
+  await pool.query("DELETE FROM expense_metadata");
+  ({ ownerId, otherId } = await seedParticipants(pool));
+});
+
+describe("reconcile mode", () => {
+  it("folds a wallet row into its statement twin even though the amounts differ by the FX fee", async () => {
+    const wallet = await insertTransaction(pool, ownerId, {
+      description: "SP PEAKDESIGN", amount: 42.44, transaction_date: "2026-08-16",
+    });
+    const stmt = await statementLine(43.71, "2026-08-15", "SP PEAKDESIGN SAN FRANCISCO USA");
+
+    const res = await call(wallet, { other_id: stmt, mode: "reconcile" });
+    expect(res.status ?? 200).toBe(200);
+
+    const r = await pool.query(`SELECT reconciled_with_id FROM transactions WHERE id = $1`, [wallet]);
+    expect(r.rows[0].reconciled_with_id).toBe(stmt);
+  });
+
+  it("refuses when the target is not a statement row", async () => {
+    const a = await insertTransaction(pool, ownerId, { amount: 10 });
+    const b = await insertTransaction(pool, ownerId, { amount: 10 });
+    const res = await call(a, { other_id: b, mode: "reconcile" });
+    expect(res.status).toBe(400);
+  });
+
+  it("refuses when the source is itself a statement row", async () => {
+    const s1 = await statementLine(10, "2026-08-01", "A");
+    const s2 = await statementLine(10, "2026-08-02", "B");
+    const res = await call(s1, { other_id: s2, mode: "reconcile" });
+    expect(res.status).toBe(400);
+  });
+});
+
+describe("supersede mode", () => {
+  it("carries splits onto the survivor before hiding the loser", async () => {
+    // The whole reason supersede moves curation: `superseded_by_id` removes the
+    // row from every figure, so a split left behind stops counting silently and
+    // a settled debt reappears as owed.
+    const loser = await insertTransaction(pool, ownerId, { amount: 80, description: "DUP" });
+    const survivor = await insertTransaction(pool, ownerId, { amount: 80, description: "KEEP" });
+    await pool.query(
+      `INSERT INTO transaction_splits (transaction_id, participant_id, share_percent, settled)
+       VALUES ($1,$2,50,true), ($1,$3,50,false)`,
+      [loser, ownerId, otherId]
+    );
+
+    const res = await call(loser, { other_id: survivor, mode: "supersede" });
+    expect(res.status ?? 200).toBe(200);
+
+    const onLoser = await pool.query(`SELECT count(*)::int c FROM transaction_splits WHERE transaction_id=$1`, [loser]);
+    const onSurvivor = await pool.query(
+      `SELECT participant_id, share_percent, settled FROM transaction_splits
+        WHERE transaction_id=$1 ORDER BY participant_id`, [survivor]
+    );
+    expect(onLoser.rows[0].c).toBe(0);
+    expect(onSurvivor.rows).toHaveLength(2);
+    // settled must survive the move -- a delete-and-recreate defaults it to false.
+    expect(onSurvivor.rows.map((r) => r.settled)).toContain(true);
+
+    const marked = await pool.query(`SELECT superseded_by_id FROM transactions WHERE id=$1`, [loser]);
+    expect(marked.rows[0].superseded_by_id).toBe(survivor);
+  });
+
+  it("works between two statement rows, which reconcile cannot express", async () => {
+    const s1 = await statementLine(25, "2026-08-01", "IMPORTED TWICE");
+    const s2 = await statementLine(25, "2026-08-01", "IMPORTED TWICE");
+    const res = await call(s1, { other_id: s2, mode: "supersede" });
+    expect(res.status ?? 200).toBe(200);
+    const r = await pool.query(`SELECT superseded_by_id FROM transactions WHERE id=$1`, [s1]);
+    expect(r.rows[0].superseded_by_id).toBe(s2);
+  });
+
+  it("refuses to hide a statement row behind a statementless one", async () => {
+    // The wrong direction, and silently wrong: a wallet row holds the amount
+    // BEFORE the card's foreign transaction fee, so burying the statement line
+    // under it leaves the ledger holding $42.44 for a $43.71 charge.
+    const stmt = await statementLine(43.71, "2026-08-15", "SP PEAKDESIGN SAN FRANCISCO USA");
+    const wallet = await insertTransaction(pool, ownerId, {
+      description: "SP PEAKDESIGN", amount: 42.44, transaction_date: "2026-08-16",
+    });
+    const res = await call(stmt, { other_id: wallet, mode: "supersede" });
+    expect(res.status).toBe(400);
+    // and the statement row is untouched
+    const r = await pool.query(`SELECT superseded_by_id FROM transactions WHERE id=$1`, [stmt]);
+    expect(r.rows[0].superseded_by_id).toBeNull();
+  });
+
+  it("refuses to bury a row under one that is already superseded", async () => {
+    const a = await insertTransaction(pool, ownerId, { amount: 5 });
+    const b = await insertTransaction(pool, ownerId, { amount: 5 });
+    const c = await insertTransaction(pool, ownerId, { amount: 5 });
+    await call(b, { other_id: c, mode: "supersede" });          // b -> c
+    const res = await call(a, { other_id: b, mode: "supersede" }); // a -> b, but b is buried
+    expect(res.status).toBe(409);
+  });
+});
+
+describe("refusals common to both modes", () => {
+  it("refuses a transaction merging into itself", async () => {
+    const a = await insertTransaction(pool, ownerId, { amount: 1 });
+    const res = await call(a, { other_id: a, mode: "supersede" });
+    expect(res.status).toBe(400);
+  });
+
+  it("refuses a row that is already reconciled", async () => {
+    const wallet = await insertTransaction(pool, ownerId, { amount: 12, transaction_date: "2026-08-10" });
+    const stmt = await statementLine(12, "2026-08-10", "X");
+    await call(wallet, { other_id: stmt, mode: "reconcile" });
+    const other = await insertTransaction(pool, ownerId, { amount: 12 });
+    const res = await call(wallet, { other_id: other, mode: "supersede" });
+    expect(res.status).toBe(409);
+  });
+
+  it("refuses a transaction belonging to someone else", async () => {
+    const mine = await insertTransaction(pool, ownerId, { amount: 7 });
+    const theirs = await insertTransaction(pool, otherId, { amount: 7 });
+    const res = await call(mine, { other_id: theirs, mode: "supersede" });
+    expect(res.status).toBe(404);
+  });
+});

src/__tests__/integration/receipt-adopts-wallet-row.test.ts

@@ -0,0 +1,153 @@
+import { describe, it, expect, beforeAll, beforeEach, vi } from "vitest";
+import { Pool } from "pg";
+import { createPool, resetDB, seedParticipants } from "./helpers";
+
+/**
+ * A receipt arriving after the wallet notification must ENRICH that row, not
+ * create a second one.
+ *
+ * The case this is built from is real. Woolworths $103.30 on 2026-08-18: the
+ * wallet event landed at 09:00:14 on the 18th, and nothing created a
+ * transaction until 20:25:06 on the 19th -- 35h25m later -- at which point the
+ * receipt path made its own row from scratch. One shop, two rows, and the
+ * notification's contribution (the instrument, and the suburb) thrown away.
+ *
+ * Sources for one purchase arrive in sequence, not in competition, so the rule
+ * is: the receipt read the register tape and wins on what was bought; anything
+ * only the notification knew survives untouched.
+ */
+
+const TOKEN = "test-receipt-token";
+let pool: Pool;
+let ownerId: number;
+let ingest: typeof import("../../app/api/receipts/ingest/route").POST;
+
+const ingestReq = (body: unknown) =>
+  ({ headers: { get: (h: string) => (h === "x-ingest-token" ? TOKEN : null) }, json: async () => body }) as never;
+
+/** A row shaped the way the wallet lane writes one: statementless, no line
+ *  items, carrying the suburb the notification supplied. */
+async function walletRow(amount: number, date: string, merchant: string, location: string) {
+  const t = await pool.query(
+    `INSERT INTO transactions
+       (owner_id, statement_id, transaction_date, description, amount, amount_aud,
+        transaction_type, category, payment_method, merchant_name, merchant_normalized,
+        location, source, source_ref, row_index)
+     VALUES ($1, NULL, $2, $3, $4, $4, 'debit', 'other', 'card', $3, $3, $5, 'wallet', $6, 0)
+     RETURNING id`,
+    [ownerId, date, merchant, amount, location, `evt-${amount}-${date}`]
+  );
+  const id = t.rows[0].id as number;
+  await pool.query(
+    `INSERT INTO expense_metadata
+       (transaction_id, source, merchant_normalized, amount, transaction_date,
+        currency, line_items, payment_method_detail, flags)
+     VALUES ($1, 'wallet', $2, $3, $4, 'AUD', '[]'::jsonb, 'Visa', '[]'::jsonb)`,
+    [id, merchant, amount, date]
+  );
+  return id;
+}
+
+const receipt = (amount: number, date: string) => ({
+  receipt_uid: `woolworths:3345:${amount}:2026-08-18`,
+  capture_event_id: 900,
+  merchant_name: "Woolworths",
+  store_detail: "Wyndham Vale",
+  transaction_date: date,
+  total: amount,
+  tender_legs: [
+    { leg_index: 1, amount, card_last4: "6078", card_product: "VISA PREPAID", class: "gift_card" },
+  ],
+  line_items: [
+    { description: "Milk 2L", qty: 1, unit: "ea", amount: 3.5, category: "dairy" },
+    { description: "Bread", qty: 1, unit: "ea", amount: 4.0, category: "bakery" },
+  ],
+});
+
+beforeAll(async () => {
+  process.env.RECEIPT_INGEST_TOKEN = TOKEN;
+  pool = createPool();
+  vi.doMock("@/lib/auth", () => ({
+    getCurrentUser: async () => ({ id: ownerId, name: "Alice", email: "alice@example.com" }),
+  }));
+  ({ POST: ingest } = await import("../../app/api/receipts/ingest/route"));
+});
+
+beforeEach(async () => {
+  await resetDB(pool);
+  await pool.query("DELETE FROM expense_metadata");
+  ({ ownerId } = await seedParticipants(pool));
+});
+
+const txCount = async () =>
+  Number((await pool.query(`SELECT count(*)::int AS n FROM transactions`)).rows[0].n);
+
+describe("a receipt meeting the wallet row for the same tap", () => {
+  it("adopts it instead of creating a second transaction", async () => {
+    const walletId = await walletRow(103.3, "2026-08-18", "Woolworths", "Wyndham Vale, VIC");
+    expect(await txCount()).toBe(1);
+
+    await ingest(ingestReq(receipt(103.3, "2026-08-18")));
+
+    expect(await txCount()).toBe(1);
+    const meta = await pool.query(
+      `SELECT transaction_id, jsonb_array_length(line_items) AS items, source,
+              payment_method_detail, card_last4
+         FROM expense_metadata`
+    );
+    expect(meta.rowCount).toBe(1);
+    expect(meta.rows[0].transaction_id).toBe(walletId);
+    expect(meta.rows[0].items).toBe(2);
+    expect(meta.rows[0].source).toBe("pantry");
+    // The receipt read the card off the tape; the notification only saw "Visa".
+    expect(meta.rows[0].payment_method_detail).toBe("VISA PREPAID");
+    expect(meta.rows[0].card_last4).toBe("6078");
+  });
+
+  it("keeps the suburb only the notification knew", async () => {
+    await walletRow(103.3, "2026-08-18", "Woolworths", "Wyndham Vale, VIC");
+    await ingest(ingestReq(receipt(103.3, "2026-08-18")));
+
+    const t = await pool.query(`SELECT location, merchant_normalized, category FROM transactions`);
+    // `location` is absent from the adoption UPDATE's column list on purpose:
+    // the receipt has no suburb, so it must not be able to erase one.
+    expect(t.rows[0].location).toBe("Wyndham Vale, VIC");
+    // ...while what the receipt does know wins.
+    expect(t.rows[0].merchant_normalized).toBe("Woolworths");
+    expect(t.rows[0].category).toBe("groceries");
+  });
+
+  it("creates its own row when there is no wallet row to adopt", async () => {
+    await ingest(ingestReq(receipt(58.4, "2026-08-18")));
+    expect(await txCount()).toBe(1);
+    const meta = await pool.query(`SELECT source, jsonb_array_length(line_items) AS items FROM expense_metadata`);
+    expect(meta.rows[0].source).toBe("pantry");
+    expect(meta.rows[0].items).toBe(2);
+  });
+
+  it("refuses to adopt when two candidates share the amount", async () => {
+    // Two taps of the same value in the window: adopting either would move this
+    // shop's line items onto a purchase it did not describe.
+    await walletRow(103.3, "2026-08-17", "Woolworths", "Wyndham Vale, VIC");
+    await walletRow(103.3, "2026-08-18", "Coles", "Manor Lakes, VIC");
+    expect(await txCount()).toBe(2);
+
+    await ingest(ingestReq(receipt(103.3, "2026-08-18")));
+
+    // Falls back to creating its own row rather than guessing.
+    expect(await txCount()).toBe(3);
+  });
+
+  it("refuses to adopt a row that already carries line items", async () => {
+    const other = await walletRow(103.3, "2026-08-18", "Woolworths", "Wyndham Vale, VIC");
+    await pool.query(
+      `UPDATE expense_metadata SET line_items = '[{"description":"Something else","amount":1}]'::jsonb
+        WHERE transaction_id = $1`, [other]
+    );
+
+    await ingest(ingestReq(receipt(103.3, "2026-08-18")));
+
+    // A fully-described row is a different purchase that happens to share an amount.
+    expect(await txCount()).toBe(2);
+  });
+});

src/__tests__/integration/reconcile-queue.test.ts

@@ -0,0 +1,86 @@
+import { describe, it, expect, beforeAll, beforeEach } from "vitest";
+import { Pool } from "pg";
+import { createPool, resetDB, seedParticipants, insertTransaction } from "./helpers";
+import { getPendingReconciliations } from "../../lib/queries";
+
+/**
+ * The reconcile queue must not offer what is already resolved.
+ *
+ * A superseded row is hidden from every figure (EXCLUDE_RECONCILED_SOURCE), so
+ * listing it is not merely noise: on 2026-10-03 the queue showed the duplicate
+ * wallet half 7037, superseded under its twin 7038, as a second "Digidirect"
+ * with the same statement line offered as its match. Accepting it there would
+ * have spent that line on a hidden row and left the real one unmatched forever.
+ * findAutoMatches() always had both filters; this query had neither.
+ */
+
+let pool: Pool;
+let ownerId: number;
+let acctSeq = 0;
+
+async function statementLine(amount: number, date: string, description: string) {
+  const st = await pool.query(
+    `INSERT INTO statements (filename, account_number, bank_name, billing_start_date, billing_end_date, owner_id)
+     VALUES ('q.pdf', $3, 'Test Bank', $1::date - 20, $1::date + 10, $2) RETURNING id`,
+    [date, ownerId, `5555-0000-0000-${String(1000 + acctSeq++)}`]
+  );
+  const t = await pool.query(
+    `INSERT INTO transactions (statement_id, transaction_date, description, amount, transaction_type, owner_id)
+     VALUES ($1, $2, $3, $4, 'debit', $5) RETURNING id`,
+    [st.rows[0].id, date, description, amount, ownerId]
+  );
+  return t.rows[0].id as number;
+}
+
+beforeAll(() => { pool = createPool(); });
+
+beforeEach(async () => {
+  await resetDB(pool);
+  ({ ownerId } = await seedParticipants(pool));
+});
+
+describe("getPendingReconciliations", () => {
+  it("does not list a superseded manual row", async () => {
+    const survivor = await insertTransaction(pool, ownerId, { description: "DIGIDIRECT", amount: 1180.05, transaction_date: "2026-08-25" });
+    const hidden = await insertTransaction(pool, ownerId, { description: "Digidirect", amount: 1180.05, transaction_date: "2026-08-25" });
+    await pool.query(`UPDATE transactions SET superseded_by_id = $1 WHERE id = $2`, [survivor, hidden]);
+    const line = await statementLine(1180.05, "2026-08-25", "SP DIGIDIRECT STRATHFIELD");
+
+    const queue = await getPendingReconciliations(ownerId);
+    const ids = queue.map((r) => r.id);
+    expect(ids).toContain(survivor);
+    expect(ids).not.toContain(hidden);
+    // The survivor is still offered its line.
+    const row = queue.find((r) => r.id === survivor)!;
+    expect(row.matches.map((m) => m.id)).toContain(line);
+  });
+
+  it("does not offer a superseded statement row as a match", async () => {
+    const manual = await insertTransaction(pool, ownerId, { description: "Grocer", amount: 42.0, transaction_date: "2026-08-10" });
+    const live = await statementLine(42.0, "2026-08-10", "GROCER");
+    const dupe = await statementLine(42.0, "2026-08-10", "GROCER");
+    await pool.query(`UPDATE transactions SET superseded_by_id = $1 WHERE id = $2`, [live, dupe]);
+
+    const [row] = (await getPendingReconciliations(ownerId)).filter((r) => r.id === manual);
+    const offered = row.matches.map((m) => m.id);
+    expect(offered).toContain(live);
+    expect(offered).not.toContain(dupe);
+  });
+
+  it("offers the exact native twin of a foreign row even when its provisional AUD is far off", async () => {
+    // Live shape, 2026-10-09: wallet INR 112 at a provisional A$2.01; CommBank
+    // charged A$1.68 -- 16% under, outside the +-5% band.
+    const manual = await insertTransaction(pool, ownerId, { description: "SWIGGY LIMITED", amount: 2.01, transaction_date: "2026-10-04" });
+    await pool.query(`UPDATE transactions SET foreign_currency_code = 'INR', foreign_currency_amount = 112 WHERE id = $1`, [manual]);
+    const twin = await statementLine(1.68, "2026-10-04", "Swiggy Limited Bengaluru");
+    const otherRupees = await statementLine(1.68, "2026-10-04", "Swiggy Limited Bengaluru");
+    await pool.query(`UPDATE transactions SET foreign_currency_code = 'INR', foreign_currency_amount = 112 WHERE id = $1`, [twin]);
+    await pool.query(`UPDATE transactions SET foreign_currency_code = 'INR', foreign_currency_amount = 113 WHERE id = $1`, [otherRupees]);
+
+    const [row] = (await getPendingReconciliations(ownerId)).filter((r) => r.id === manual);
+    const offered = row.matches.map((m) => m.id);
+    expect(offered).toContain(twin);
+    // A different native figure at the same AUD is not admitted by this path.
+    expect(offered).not.toContain(otherRupees);
+  });
+});

src/__tests__/integration/void.test.ts

@@ -0,0 +1,167 @@
+import { describe, it, expect, beforeAll, beforeEach, vi } from "vitest";
+import { Pool } from "pg";
+import { createPool, resetDB, seedParticipants, insertTransaction } from "./helpers";
+import { EXCLUDE_RECONCILED_SOURCE } from "@/lib/analytics-sql";
+
+/**
+ * A charge that never posted (migration 0040).
+ *
+ * Txn 7951: a $1,040.41 Airbnb authorisation hold on Amex, booking cancelled
+ * the next day, hold released. The wallet lane booked the notification as a
+ * purchase and no statement line ever arrived, so it sat in the reconcile queue
+ * forever and counted as spend everywhere. A void has no survivor, which is why
+ * neither reconcile nor supersede could express it.
+ *
+ * As in merge-supersede.test.ts, the refusals matter as much as the happy path:
+ * a void quietly stops money counting.
+ */
+
+let pool: Pool;
+let ownerId: number;
+let otherId: number;
+let voidRoute: typeof import("../../app/api/transactions/[id]/void/route");
+let merge: typeof import("../../app/api/transactions/[id]/merge/route").POST;
+
+const req = (body: unknown) => ({ json: async () => body, headers: { get: () => null } }) as never;
+const params = (id: number) => ({ params: Promise.resolve({ id: String(id) }) });
+const doVoid = (id: number, reason: unknown = "hold released, booking cancelled") =>
+  voidRoute.POST(req({ reason }), params(id));
+
+let acctSeq = 0;
+async function statementLine(amount: number, date: string, description: string) {
+  const st = await pool.query(
+    `INSERT INTO statements (filename, account_number, bank_name, billing_start_date, billing_end_date, owner_id)
+     VALUES ('s.pdf', $3, 'Test Bank', $1::date - 20, $1::date + 10, $2) RETURNING id`,
+    [date, ownerId, `5555-0000-0000-${String(1000 + acctSeq++)}`]
+  );
+  const t = await pool.query(
+    `INSERT INTO transactions (statement_id, transaction_date, description, amount, transaction_type, owner_id)
+     VALUES ($1, $2, $3, $4, 'debit', $5) RETURNING id`,
+    [st.rows[0].id, date, description, amount, ownerId]
+  );
+  return t.rows[0].id as number;
+}
+
+/** What every "count each purchase once" query sees. */
+async function countedIds(): Promise<number[]> {
+  const r = await pool.query(
+    `SELECT t.id FROM transactions t WHERE ${EXCLUDE_RECONCILED_SOURCE} ORDER BY t.id`
+  );
+  return r.rows.map((x) => x.id);
+}
+
+beforeAll(async () => {
+  pool = createPool();
+  vi.doMock("@/lib/auth", () => ({
+    getCurrentUser: async () => ({ id: ownerId, name: "Alice", email: "alice@example.com" }),
+  }));
+  voidRoute = await import("../../app/api/transactions/[id]/void/route");
+  ({ POST: merge } = await import("../../app/api/transactions/[id]/merge/route"));
+});
+
+beforeEach(async () => {
+  await resetDB(pool);
+  await pool.query("DELETE FROM expense_metadata");
+  ({ ownerId, otherId } = await seedParticipants(pool));
+});
+
+describe("voiding a charge that never posted", () => {
+  it("drops the row out of every figure and the reconcile queue, but keeps it", async () => {
+    const { getPendingReconciliations } = await import("@/lib/queries");
+    const hold = await insertTransaction(pool, ownerId, { description: "AIRBNB", amount: 1040.41 });
+    const real = await insertTransaction(pool, ownerId, { description: "AIRBNB", amount: 529.05 });
+
+    expect((await getPendingReconciliations(ownerId)).map((t) => t.id)).toContain(hold);
+
+    const res = await doVoid(hold);
+    expect(res.status ?? 200).toBe(200);
+
+    expect(await countedIds()).toEqual([real]);
+    expect((await getPendingReconciliations(ownerId)).map((t) => t.id)).not.toContain(hold);
+    const r = await pool.query(`SELECT voided_at, void_reason FROM transactions WHERE id = $1`, [hold]);
+    expect(r.rows[0].voided_at).not.toBeNull();
+    expect(r.rows[0].void_reason).toBe("hold released, booking cancelled");
+  });
+
+  it("can be undone", async () => {
+    const hold = await insertTransaction(pool, ownerId, { amount: 1040.41 });
+    await doVoid(hold);
+    const res = await voidRoute.DELETE(req({}), params(hold));
+    expect(res.status ?? 200).toBe(200);
+    expect(await countedIds()).toEqual([hold]);
+  });
+
+  it("requires a reason", async () => {
+    const hold = await insertTransaction(pool, ownerId, { amount: 10 });
+    expect((await doVoid(hold, "   ")).status).toBe(400);
+    expect((await doVoid(hold, null)).status).toBe(400);
+    expect(await countedIds()).toEqual([hold]);
+  });
+
+  it("refuses a statement line -- that charge posted", async () => {
+    const stmt = await statementLine(50, "2026-09-10", "SOMETHING");
+    expect((await doVoid(stmt)).status).toBe(409);
+    expect(await countedIds()).toEqual([stmt]);
+  });
+
+  it("refuses a reconciled row -- it found its statement line, so it posted", async () => {
+    const wallet = await insertTransaction(pool, ownerId, { amount: 50, transaction_date: "2026-09-10" });
+    const stmt = await statementLine(50, "2026-09-10", "SOMETHING");
+    await pool.query(`UPDATE transactions SET reconciled_with_id = $1 WHERE id = $2`, [stmt, wallet]);
+    expect((await doVoid(wallet)).status).toBe(409);
+  });
+
+  it("refuses a row with a settled split -- someone already paid their share", async () => {
+    const hold = await insertTransaction(pool, ownerId, { amount: 80 });
+    await pool.query(
+      `INSERT INTO transaction_splits (transaction_id, participant_id, share_percent, settled)
+       VALUES ($1, $2, 50, true)`,
+      [hold, otherId]
+    );
+    expect((await doVoid(hold)).status).toBe(409);
+    expect(await countedIds()).toEqual([hold]);
+  });
+
+  it("is enforced by the database, not only the route", async () => {
+    const stmt = await statementLine(50, "2026-09-10", "SOMETHING");
+    await expect(
+      pool.query(`UPDATE transactions SET voided_at = now(), void_reason = 'x' WHERE id = $1`, [stmt])
+    ).rejects.toThrow(/transactions_void_statementless/);
+
+    const manual = await insertTransaction(pool, ownerId, { amount: 5 });
+    await expect(
+      pool.query(`UPDATE transactions SET voided_at = now() WHERE id = $1`, [manual])
+    ).rejects.toThrow(/transactions_void_has_reason/);
+  });
+});
+
+describe("the matchers never use a voided row", () => {
+  it("auto-reconcile does not pair a voided row with a statement line of the same amount", async () => {
+    const { findAutoMatches } = await import("@/lib/auto-reconcile");
+    const hold = await insertTransaction(pool, ownerId, { amount: 120, transaction_date: "2026-09-10" });
+    await statementLine(120, "2026-09-11", "SOMETHING");
+    expect((await findAutoMatches(ownerId)).map((m) => m.manual_id)).toContain(hold);
+
+    await doVoid(hold);
+    expect(await findAutoMatches(ownerId)).toEqual([]);
+  });
+
+  it("a receipt for a rebooking at the same amount adopts the live row, not the released hold", async () => {
+    const { findAdoptableTransaction } = await import("@/lib/auto-reconcile");
+    const hold = await insertTransaction(pool, ownerId, { amount: 300, transaction_date: "2026-09-10" });
+    const rebooked = await insertTransaction(pool, ownerId, { amount: 300, transaction_date: "2026-09-11" });
+    // Two candidates: the adopter refuses rather than guess.
+    expect(await findAdoptableTransaction(ownerId, 300, "2026-09-11")).toBeNull();
+
+    await doVoid(hold);
+    expect(await findAdoptableTransaction(ownerId, 300, "2026-09-11")).toBe(rebooked);
+  });
+
+  it("merge refuses a voided row on either side", async () => {
+    const hold = await insertTransaction(pool, ownerId, { amount: 70, transaction_date: "2026-09-10" });
+    const stmt = await statementLine(70, "2026-09-10", "SOMETHING");
+    await doVoid(hold);
+    const res = await merge(req({ other_id: stmt, mode: "reconcile" }), params(hold));
+    expect(res.status).toBe(409);
+  });
+});

src/__tests__/integration/wallet-adopts-receipt-row.test.ts

@@ -0,0 +1,176 @@
+import { describe, it, expect, beforeAll, beforeEach, vi } from "vitest";
+import { Pool } from "pg";
+import { createPool, resetDB, seedParticipants } from "./helpers";
+
+/**
+ * The mirror of receipt-adopts-wallet-row: a wallet notification arriving AFTER
+ * something else already booked the purchase must fold into that row.
+ *
+ * Built from the failure, not from the design. The 2026-08-18 Woolworths shop
+ * ($103.30, Visa gift card) reached the ledger twice: txn 6858 from the pantry
+ * receipt on the 19th, and txn 6927 from the wallet lane on the 22nd, when the
+ * first 21 wallet events were backfilled days after their taps. The sequencing
+ * the lane is designed around -- wallet first, receipt second -- is a tendency,
+ * and every writer has to survive it not holding.
+ *
+ * Two rules are load-bearing here and are tested separately below:
+ *   - the merchant must agree, because losing the "no line items" guard that
+ *     protects the other direction means amount+date alone could swallow a tap
+ *     into an unrelated purchase of the same value
+ *   - a gift-card tap is `credits`, because `card` means "a statement line is
+ *     coming" and the reconcile queue filters on exactly that
+ */
+
+const TOKEN = "test-wallet-token";
+let pool: Pool;
+let ownerId: number;
+let ingest: typeof import("../../app/api/wallet/ingest/route").POST;
+
+const req = (body: unknown) =>
+  ({ headers: { get: (h: string) => (h === "x-ingest-token" ? TOKEN : null) }, json: async () => body }) as never;
+
+/** A row shaped the way the pantry receipt lane writes one: statementless,
+ *  itemised, no location (nothing but a notification supplies a suburb). */
+async function receiptRow(amount: number, date: string, merchant: string, opts: { items?: number; meta?: boolean } = {}) {
+  const { items = 3, meta = true } = opts;
+  const t = await pool.query(
+    `INSERT INTO transactions
+       (owner_id, statement_id, transaction_date, description, amount, amount_aud,
+        transaction_type, category, payment_method, merchant_name, merchant_normalized, row_index)
+     VALUES ($1, NULL, $2, $3, $4, $4, 'debit', 'groceries', 'credits', $3, $3, 0)
+     RETURNING id`,
+    [ownerId, date, merchant, amount]
+  );
+  const id = t.rows[0].id as number;
+  if (meta) {
+    const lineItems = Array.from({ length: items }, (_, i) => ({ description: `item ${i}`, qty: 1, amount: 1 }));
+    await pool.query(
+      `INSERT INTO expense_metadata
+         (transaction_id, source, merchant_normalized, amount, transaction_date, currency, line_items, flags)
+       VALUES ($1, 'pantry', $2, $3, $4, 'AUD', $5::jsonb, '[]'::jsonb)`,
+      [id, merchant, amount, date, JSON.stringify(lineItems)]
+    );
+  }
+  return id;
+}
+
+/** The real event 80: the Apple Wallet half of a Visa gift-card tap. */
+const woolworthsGiftCard = {
+  event_id: 80, kind: "transaction", direction: "unknown", amount: 103.3,
+  amount_text: "$103.30", currency: "$?", merchant: "Woolworths", issuer: "Visa",
+  notif_app: "Wallet", notif_title: "Visa Gift Card",
+  notif_subtitle: "Woolworths, Wyndham Vale, VIC",
+  occurred_at: "2026-08-18T09:00:14+10:00",
+};
+
+const rows = async () =>
+  (await pool.query(
+    `SELECT id, amount, merchant_normalized, location, source, source_ref, payment_method
+       FROM transactions ORDER BY id`
+  )).rows;
+
+beforeAll(async () => {
+  process.env.WALLET_INGEST_TOKEN = TOKEN;
+  pool = createPool();
+  vi.doMock("@/lib/auth", () => ({
+    getCurrentUser: async () => ({ id: ownerId, name: "Alice", email: "alice@example.com" }),
+  }));
+  ({ POST: ingest } = await import("../../app/api/wallet/ingest/route"));
+});
+
+beforeEach(async () => {
+  await resetDB(pool);
+  await pool.query("DELETE FROM expense_metadata");
+  ({ ownerId } = await seedParticipants(pool));
+});
+
+describe("a wallet notification for a purchase the ledger already holds", () => {
+  it("adopts the receipt's row instead of creating a second one", async () => {
+    const existing = await receiptRow(103.3, "2026-08-18", "WOOLWORTHS 3345");
+
+    const body = await (await ingest(req(woolworthsGiftCard))).json();
+
+    expect(body.skipped).toBe("adopted");
+    expect(body.transactionId).toBe(existing);
+    expect(await rows()).toHaveLength(1);
+  });
+
+  it("gives the adopted row the suburb, and leaves what the receipt read alone", async () => {
+    const existing = await receiptRow(103.3, "2026-08-18", "WOOLWORTHS 3345");
+    await ingest(req(woolworthsGiftCard));
+
+    const [t] = await rows();
+    // Wallet's own field: nothing else on the estate writes a suburb.
+    expect(t.location).toBe("Wyndham Vale, VIC");
+    // The receipt read the register tape. The notification does not get to
+    // rename the merchant to its own shorter string.
+    expect(t.merchant_normalized).toBe("WOOLWORTHS 3345");
+    expect(t.id).toBe(existing);
+
+    const [m] = (await pool.query(`SELECT * FROM expense_metadata ORDER BY id`)).rows;
+    expect(m.source).toBe("pantry");
+    expect(m.line_items).toHaveLength(3);
+    expect(m.payment_method_detail).toBe("Visa");
+    expect(m.flags).toContain("wallet_evt:80");
+  });
+
+  it("is idempotent when the interpreter replays every event", async () => {
+    // Bumping prompt_version re-interprets and re-posts the whole capture
+    // history on purpose -- that is the replay mechanism. An adopted row keeps
+    // no wallet source_ref of its own, so the flag is what makes replay safe.
+    await ingest(req(woolworthsGiftCard));
+    const second = await (await ingest(req(woolworthsGiftCard))).json();
+    expect(second.skipped).toBe("already_ingested");
+    expect(await rows()).toHaveLength(1);
+  });
+
+  it("writes its own row when the merchant disagrees", async () => {
+    // Same money, same day, different shop. Folding into it would make the tap
+    // vanish -- worse than a duplicate, which is at least visible in the queue.
+    await receiptRow(103.3, "2026-08-18", "Bunnings Werribee");
+    const body = await (await ingest(req(woolworthsGiftCard))).json();
+
+    expect(body.skipped).toBeUndefined();
+    expect(await rows()).toHaveLength(2);
+  });
+
+  it("writes its own row when two candidates share the amount", async () => {
+    await receiptRow(103.3, "2026-08-18", "Woolworths Wyndham Vale");
+    await receiptRow(103.3, "2026-08-18", "Woolworths Point Cook");
+    const body = await (await ingest(req(woolworthsGiftCard))).json();
+
+    expect(body.skipped).toBeUndefined();
+    expect(await rows()).toHaveLength(3);
+  });
+
+  it("creates the metadata row when the adopted transaction has none", async () => {
+    const existing = await receiptRow(103.3, "2026-08-18", "WOOLWORTHS 3345", { meta: false });
+    await ingest(req(woolworthsGiftCard));
+
+    const meta = (await pool.query(`SELECT * FROM expense_metadata`)).rows;
+    expect(meta).toHaveLength(1);
+    expect(meta[0].transaction_id).toBe(existing);
+    expect(meta[0].card_last4).toBeNull();
+    expect(meta[0].flags).toContain("wallet_evt:80");
+  });
+});
+
+describe("a stored-value instrument", () => {
+  it("is booked as credits, so it never enters the reconcile queue", async () => {
+    const body = await (await ingest(req(woolworthsGiftCard))).json();
+    const [t] = await rows();
+
+    expect(body.transactionId).toBeTruthy();
+    // 'card' would park it forever: findAutoMatches and the queue both exclude
+    // 'credits' and 'cash' precisely because no statement line is coming.
+    expect(t.payment_method).toBe("credits");
+    expect(body.flags).toContain("stored_value_instrument");
+  });
+
+  it("leaves an ordinary card tap as card", async () => {
+    const body = await (await ingest(req({ ...woolworthsGiftCard, notif_title: "Payment", issuer: "Mastercard", notif_app: "Westpac" }))).json();
+    const [t] = await rows();
+    expect(t.payment_method).toBe("card");
+    expect(body.flags).not.toContain("stored_value_instrument");
+  });
+});

src/__tests__/integration/wallet-ingest.test.ts

@@ -0,0 +1,403 @@
+import { describe, it, expect, beforeAll, beforeEach, vi } from "vitest";
+import { Pool } from "pg";
+import { createPool, resetDB, seedParticipants } from "./helpers";
+
+/**
+ * Wallet notifications becoming transactions.
+ *
+ * Every fixture is a real captured event from capture.wallet_events
+ * (2026-08-13 .. 08-22). The rules being tested are the ones the design turned
+ * on, so each has a reason rather than being a shape check:
+ *
+ *   - a transaction for EVERY tap, because overseas is where live spend matters
+ *   - the pair is MERGED, never deduped: the issuer half has the card tail, the
+ *     Wallet half has the merchant and suburb, neither alone is complete
+ *   - the NATIVE amount is what gets stored for a foreign charge, because the
+ *     AUD figure is the bank's own conversion and cannot be predicted
+ *   - a bare "$" is recorded as unknown, never assumed AUD
+ */
+
+const TOKEN = "test-wallet-token";
+let pool: Pool;
+let ownerId: number;
+let ingest: typeof import("../../app/api/wallet/ingest/route").POST;
+
+const req = (body: unknown) =>
+  ({ headers: { get: (h: string) => (h === "x-ingest-token" ? TOKEN : null) }, json: async () => body }) as never;
+
+const noAuth = (body: unknown) =>
+  ({ headers: { get: () => null }, json: async () => body }) as never;
+
+// --- real captured events -------------------------------------------------
+const breadworldBank = {
+  event_id: 64, kind: "transaction", direction: "debit", amount: 13.5,
+  amount_text: "$13.50", currency: "$?", merchant: "Breadworld Bakery Cafe",
+  card_tail: "8032", issuer: "Mastercard", notif_app: "Westpac",
+  notif_body: "$13.50 paid from Mastercard ending...8032. Breadworld Bakery Cafe     WYNDHAM VAL AU",
+  occurred_at: "2026-08-13T08:01:28+10:00",
+};
+const breadworldWallet = {
+  event_id: 65, kind: "transaction", direction: "unknown", amount: 13.5,
+  amount_text: "$13.50", currency: "$?", merchant: "Breadworld Bakery Cafe",
+  notif_app: "Wallet", notif_subtitle: "Breadworld Bakery Cafe, Wyndham Vale, VIC",
+  occurred_at: "2026-08-13T08:01:35+10:00",
+};
+const amexRewardPay = {
+  event_id: 87, kind: "transaction", direction: "debit", amount: 5103.49,
+  amount_text: "A$5,103.49", currency: "AUD", merchant: "REWARDPAY",
+  card_tail: "01000", issuer: "American Express", notif_app: "Amex AU",
+  notif_body: "You have made a A$5,103.49 charge on your American Express® Card ending in 01000 at REWARDPAY.",
+  occurred_at: "2026-08-20T11:54:43+10:00",
+};
+const ebayOrderPlaced = {
+  event_id: 66, kind: "not_transaction", direction: "unknown", amount: null,
+  notif_app: "Wallet", notif_body: "eBay has received your order.",
+  occurred_at: "2026-08-13T09:45:55+10:00",
+};
+
+/** Only the rows this lane wrote -- some tests seed statement history for the
+ *  rate reference, and those must not be mistaken for wallet output. */
+const txns = async () =>
+  (await pool.query(
+    `SELECT id, amount, amount_aud, merchant_normalized, location, source, source_ref,
+            foreign_currency_amount, foreign_currency_code, payment_method, statement_id
+       FROM transactions WHERE source = 'wallet' ORDER BY id`
+  )).rows;
+
+const metas = async () =>
+  (await pool.query(`SELECT * FROM expense_metadata ORDER BY id`)).rows;
+
+beforeAll(async () => {
+  process.env.WALLET_INGEST_TOKEN = TOKEN;
+  pool = createPool();
+  vi.doMock("@/lib/auth", () => ({
+    getCurrentUser: async () => ({ id: ownerId, name: "Alice", email: "alice@example.com" }),
+  }));
+  ({ POST: ingest } = await import("../../app/api/wallet/ingest/route"));
+});
+
+beforeEach(async () => {
+  // No network in tests. The ECB fallback treats any fetch failure as "no
+  // rate", which is what every pre-existing currency test assumes; the one test
+  // about the fallback stubs a rate of its own.
+  vi.stubGlobal("fetch", async () => { throw new Error("no network in tests"); });
+  await resetDB(pool);
+  await pool.query("DELETE FROM expense_metadata");
+  ({ ownerId } = await seedParticipants(pool));
+});
+
+describe("auth", () => {
+  it("fails closed without the token", async () => {
+    expect((await ingest(noAuth(breadworldBank))).status).toBe(401);
+    expect(await txns()).toHaveLength(0);
+  });
+});
+
+describe("what reaches the ledger", () => {
+  it("creates a statementless card transaction for a tap", async () => {
+    const body = await (await ingest(req(breadworldBank))).json();
+    expect(body.transactionId).toBeTruthy();
+
+    const [t] = await txns();
+    expect(Number(t.amount)).toBe(13.5);
+    expect(t.source).toBe("wallet");
+    expect(t.source_ref).toBe("evt:64");
+    // Statementless + card is exactly what puts it in the reconcile queue, with
+    // no extra wiring: that is how the statement eventually confirms it.
+    expect(t.statement_id).toBeNull();
+    expect(t.payment_method).toBe("card");
+  });
+
+  it("ignores an eBay order-lifecycle notification", async () => {
+    // 10 of the 13 non-transactions in the capture window are these.
+    const body = await (await ingest(req(ebayOrderPlaced))).json();
+    expect(body.skipped).toBe("not_a_transaction");
+    expect(await txns()).toHaveLength(0);
+  });
+
+  it("is idempotent on the event id", async () => {
+    await ingest(req(breadworldBank));
+    const second = await (await ingest(req(breadworldBank))).json();
+    expect(second.skipped).toBe("already_ingested");
+    expect(await txns()).toHaveLength(1);
+  });
+});
+
+describe("the notification pair", () => {
+  it("merges the two halves into ONE transaction, keeping both contributions", async () => {
+    await ingest(req(breadworldBank));    // has the card tail
+    await ingest(req(breadworldWallet));  // has the suburb
+
+    const rows = await txns();
+    expect(rows).toHaveLength(1);
+    // The Wallet half supplies the suburb; the issuer half never had one.
+    expect(rows[0].location).toBe("Wyndham Vale, VIC");
+
+    const [m] = await metas();
+    // ...and the issuer half's card tail survives the merge.
+    expect(m.card_last4).toBe("8032");
+    expect(m.payment_method_detail).toBe("Mastercard");
+  });
+
+  it("merges in either arrival order", async () => {
+    await ingest(req(breadworldWallet));
+    await ingest(req(breadworldBank));
+
+    expect(await txns()).toHaveLength(1);
+    const [m] = await metas();
+    expect(m.card_last4).toBe("8032");
+    const [t] = await txns();
+    expect(t.location).toBe("Wyndham Vale, VIC");
+  });
+});
+
+describe("the notification pair is recognised by time and app, not amount alone", () => {
+  it("makes ONE row when both halves are posted at the same moment", async () => {
+    // n8n posts a poll's items concurrently, and both halves of a tap usually
+    // land in the same poll. Before the lock, each ran its pair lookup before
+    // the other inserted: 11 taps became 22 rows, identical created_at on each.
+    await Promise.all([ingest(req(breadworldBank)), ingest(req(breadworldWallet))]);
+    expect(await txns()).toHaveLength(1);
+    const [m] = await metas();
+    expect(m.card_last4).toBe("8032");
+  });
+
+  it("keeps two same-amount charges hours apart as two purchases", async () => {
+    // Event 275 ($50 at A MEDIUM) merged into the UNICEF $50 donation from 17
+    // hours earlier and renamed it, because the old window was +/-1 day.
+    await ingest(req({
+      ...amexRewardPay, event_id: 267, amount: 50, amount_text: "A$50.00", merchant: "UNICEF",
+      notif_body: "You have made a A$50.00 charge on your American Express® Card ending in 01000 at UNICEF.",
+      occurred_at: "2026-10-01T11:02:40+10:00",
+    }));
+    await ingest(req({
+      ...amexRewardPay, event_id: 275, amount: 50, amount_text: "$50.00", merchant: "A MEDIUM",
+      notif_body: "You have made a $50.00 charge on your American Express® Card ending in 01000 at A MEDIUM.",
+      occurred_at: "2026-10-02T04:06:49+10:00",
+    }));
+    const rows = await txns();
+    expect(rows).toHaveLength(2);
+    expect(rows.map((r) => r.merchant_normalized).sort()).toEqual(["A MEDIUM", "UNICEF"]);
+  });
+
+  it("never pairs two notifications from the SAME app, however close", async () => {
+    // A pair is the issuer half plus the Wallet half. Two from one app are two
+    // charges -- two $20 coffees bought together, say.
+    await ingest(req(breadworldBank));
+    await ingest(req({ ...breadworldBank, event_id: 640, occurred_at: "2026-08-13T08:01:40+10:00" }));
+    expect(await txns()).toHaveLength(2);
+  });
+
+  it("records a merged half, so a replay cannot merge it again", async () => {
+    await ingest(req(breadworldBank));
+    await ingest(req(breadworldWallet));
+    const again = await (await ingest(req(breadworldWallet))).json();
+    expect(again.skipped).toBe("already_ingested");
+    const [m] = await metas();
+    expect(m.flags).toContain("wallet_evt:65");
+  });
+});
+
+describe("notifications that name an amount but are not purchases", () => {
+  it("skips an overdraft notice", async () => {
+    // Event 235 became a $21.21 purchase named "To avoid fees and interest,".
+    const body = await (await ingest(req({
+      event_id: 235, kind: "transaction", direction: "debit", amount: 21.21,
+      amount_text: "$21.21", currency: "$?", merchant: null, notif_app: "CommBank",
+      notif_body: "Your Smart Access 1910 is overdrawn by $21.21. To avoid fees and interest, bring your account to positive by midnight 21 Sep (Syd/Melb time).",
+      occurred_at: "2026-09-22T09:00:18+10:00",
+    }))).json();
+    expect(body.skipped).toBe("account_notice");
+    expect(await txns()).toHaveLength(0);
+  });
+
+  it("skips a transit-card fare, because the card's load is the spend", async () => {
+    // Event 200: the Octopus load is on the bank statement; each fare draws it
+    // down. 11 of 13 fares were already not_transaction; this makes the other 2 agree.
+    const body = await (await ingest(req({
+      event_id: 200, kind: "transaction", direction: "debit", amount: 105,
+      amount_text: "HKD 105.00", currency: "HKD", merchant: "Octopus", notif_app: "Wallet",
+      notif_title: "Octopus", notif_subtitle: "HKD 105.00 for Train",
+      notif_body: "Your new balance is -HKD 12.60.",
+      occurred_at: "2026-09-14T02:00:37+10:00",
+    }))).json();
+    expect(body.skipped).toBe("stored_value_drawdown");
+    expect(await txns()).toHaveLength(0);
+  });
+
+  it("does not read a Westpac descriptor out of another bank's message", async () => {
+    await ingest(req({
+      event_id: 2350, kind: "transaction", direction: "debit", amount: 9.5,
+      amount_text: "$9.50", currency: "$?", merchant: null, notif_app: "CommBank",
+      notif_body: "Payment of $9.50 made. Thanks for banking with us.",
+      occurred_at: "2026-09-22T09:00:18+10:00",
+    }));
+    const [t] = await txns();
+    expect(t.merchant_normalized).toBe("Card purchase");
+  });
+});
+
+describe("dating", () => {
+  it("dates a tap in Melbourne local time, as the statements do", async () => {
+    // 21:30 UTC on 30 Sep is 07:30 on 1 Oct in Melbourne. Slicing the ISO
+    // string filed it a day early.
+    await ingest(req({ ...breadworldBank, occurred_at: "2026-09-30T21:30:00Z" }));
+    const { rows } = await pool.query(
+      `SELECT transaction_date::text AS d FROM transactions WHERE source = 'wallet'`
+    );
+    expect(rows[0].d).toBe("2026-10-01");
+  });
+});
+
+describe("currency", () => {
+  it("falls back to the ECB rate when the ledger has no history", async () => {
+    // Hong Kong: no HKD statement row existed, so every tap stored A$0.00.
+    const calls: string[] = [];
+    vi.stubGlobal("fetch", async (url: string) => {
+      calls.push(String(url));
+      return { ok: true, json: async () => ({ rates: { AUD: 0.19 } }) };
+    });
+    await ingest(req({
+      event_id: 167, kind: "transaction", direction: "debit", amount: 310.1,
+      amount_text: "HKD 310.10", currency: "HKD", merchant: "Wonder Taxi", notif_app: "Wallet",
+      notif_subtitle: "Wonder Taxi, Tsim Sha Tsui East, Kowloon",
+      occurred_at: "2026-09-10T08:54:00+10:00",
+    }));
+    const [t] = await txns();
+    expect(Number(t.amount)).toBeCloseTo(58.92, 2);
+    expect(Number(t.foreign_currency_amount)).toBe(310.1);
+    expect(calls[0]).toContain("/2026-09-10?from=HKD&to=AUD");
+    const [m] = await metas();
+    // Flagged as the mid-rate, not the bank's: it lacks the card's FX fee.
+    expect(m.flags).toEqual(expect.arrayContaining(["aud_provisional", "rate_ecb"]));
+    expect(m.flags).not.toContain("no_rate_reference");
+  });
+
+  it("records AUD when Amex states it explicitly", async () => {
+    await ingest(req(amexRewardPay));
+    const [t] = await txns();
+    expect(t.foreign_currency_code).toBeNull();
+    expect(Number(t.amount)).toBe(5103.49);
+    const [m] = await metas();
+    expect(m.currency).toBe("AUD");
+    expect(m.flags).not.toContain("currency_unverified");
+  });
+
+  it("flags a bare dollar sign rather than assuming AUD", async () => {
+    // 14 of 21 captured transaction rows are a bare "$". Assuming AUD is
+    // harmless at home and a 23x error on a THB tap abroad.
+    await ingest(req(breadworldBank));
+    const [m] = await metas();
+    expect(m.currency).toBeNull();
+    expect(m.flags).toContain("currency_unverified");
+  });
+
+  it("stores a foreign charge on the NATIVE pair, with a provisional AUD", async () => {
+    // Seed enough history for the median rate to exist (n >= 5).
+    const st = await pool.query(
+      `INSERT INTO statements (filename, account_number, bank_name, billing_start_date, billing_end_date, owner_id)
+       VALUES ('s.pdf','1','Test Bank','2026-01-01','2026-01-31',$1) RETURNING id`, [ownerId]
+    );
+    for (let i = 0; i < 6; i++) {
+      await pool.query(
+        `INSERT INTO transactions (statement_id, transaction_date, description, amount,
+           transaction_type, owner_id, foreign_currency_amount, foreign_currency_code)
+         VALUES ($1,'2026-01-10','BANGKOK',4.57,'debit',$2,100,'THB')`, [st.rows[0].id, ownerId]
+      );
+    }
+
+    await ingest(req({
+      event_id: 999, kind: "transaction", direction: "debit", amount: 500,
+      amount_text: "฿500", merchant: "7-Eleven", notif_app: "Wallet",
+      notif_subtitle: "7-Eleven, Bangkok", occurred_at: "2026-08-20T12:00:00+10:00",
+    }));
+
+    const [t] = await txns();
+    expect(Number(t.foreign_currency_amount)).toBe(500);
+    expect(t.foreign_currency_code).toBe("THB");
+    // The native figure is exact on both sides, so it is what the statement
+    // will match on. amount is only ever provisional here.
+    expect(Number(t.amount)).toBeCloseTo(22.85, 1);
+    const [m] = await metas();
+    expect(m.flags).toContain("aud_provisional");
+  });
+
+  it("still records a foreign charge when there is no rate history", async () => {
+    await ingest(req({
+      event_id: 998, kind: "transaction", direction: "debit", amount: 42,
+      amount_text: "€42.00", merchant: "Somewhere", notif_app: "Wallet",
+      occurred_at: "2026-08-20T12:00:00+10:00",
+    }));
+    const [t] = await txns();
+    expect(t.foreign_currency_code).toBe("EUR");
+    const [m] = await metas();
+    // Says so, rather than inventing a rate.
+    expect(m.flags).toContain("no_rate_reference");
+    // And never stores the native figure as AUD (672006 VND became A$672,006).
+    expect(Number(t.amount)).toBe(0);
+    expect(Number(t.foreign_currency_amount)).toBe(42);
+  });
+
+  it("derives the rate from statement rows only, never from wallet rows", async () => {
+    // Six statementless rows at a nonsense 1:1 rate must not count as history.
+    for (let i = 0; i < 6; i++) {
+      await pool.query(
+        `INSERT INTO transactions (transaction_date, description, amount, transaction_type,
+           owner_id, foreign_currency_amount, foreign_currency_code, source, source_ref)
+         VALUES ('2026-01-10','VN',1000,'debit',$1,1000,'VND','wallet',$2)`,
+        [ownerId, `evt:seed${i}`]
+      );
+    }
+    await ingest(req({
+      event_id: 997, kind: "transaction", direction: "debit", amount: 672006,
+      amount_text: "672006 VND", currency: "VND", merchant: "EVISA VIETNAM", notif_app: "CommBank",
+      occurred_at: "2026-09-05T11:22:00+10:00",
+    }));
+    const t = (await txns()).find((r) => r.source_ref === "evt:997")!;
+    expect(Number(t.amount)).toBe(0);
+    const [m] = (await metas()).filter((r) => r.source_message_id === "evt:997");
+    expect(m.flags).toContain("no_rate_reference");
+  });
+
+  it("pairs foreign halves on the native figure, not the placeholder amount", async () => {
+    await ingest(req({
+      event_id: 996, kind: "transaction", direction: "debit", amount: 42,
+      amount_text: "€42.00", merchant: "Cafe A", notif_app: "Wallet",
+      occurred_at: "2026-08-20T12:00:00+10:00",
+    }));
+    await ingest(req({
+      event_id: 995, kind: "transaction", direction: "debit", amount: 17,
+      amount_text: "€17.00", merchant: "Cafe B", notif_app: "Wallet",
+      occurred_at: "2026-08-20T12:05:00+10:00",
+    }));
+    // Both have amount 0; they are different purchases and must stay two rows.
+    expect((await txns()).filter((r) => r.foreign_currency_code === "EUR")).toHaveLength(2);
+  });
+});
+
+describe("location", () => {
+  it("takes the suburb from the Wallet subtitle", async () => {
+    await ingest(req(breadworldWallet));
+    const [t] = await txns();
+    expect(t.location).toBe("Wyndham Vale, VIC");
+  });
+
+  it("falls back to the fixed-width descriptor, country included", async () => {
+    // The issuer half has no subtitle; the country code is positional.
+    await ingest(req({
+      ...breadworldBank, event_id: 75, amount: 42.44,
+      merchant: "SP PEAKDESIGN",
+      // The model's own answer for this row, included because every real
+      // interpretation carries one -- omitting it is what let a wrong
+      // precedence order pass this test the first time, and ship
+      // "SAN FRANCISOUS" into the ledger.
+      location: "SAN FRANCISOUS",
+      notif_body: "$42.44 paid from Mastercard ending...8032. SP PEAKDESIGN              SAN FRANCISOUS",
+    }));
+    const [t] = await txns();
+    expect(t.location).toBe("SAN FRANCISO US");
+    const [m] = await metas();
+    // The only foreign merchant in the whole capture window — worth surfacing.
+    expect(m.flags).toContain("foreign_country_US");
+  });
+});

src/__tests__/unit/intake-rules.test.ts

@@ -3,17 +3,21 @@ import { readFileSync } from "fs";
 import { resolve } from "path";
 import {
   classifyInput,
+  confabulationReason,
   GROCERY_PATTERN,
   merchantKey,
   merchantMatches,
   platformForSource,
+  PROMPT_PLACEHOLDER_MERCHANTS,
+  quoteShared,
   reconcileItems,
   resolveMerchantKind,
+  textCouldBeReceipt,
   toLineItems,
   validateHeader,
   type ExtractedOrder,
 } from "../../lib/intake-rules";
-import { buildMessageContent, extractPdfText, parseExtraction, EXTRACT_RESPONSE_FORMAT, ExtractionError, liteLLMExtractor } from "../../lib/intake-extract";
+import { buildMessageContent, extractPdfText, parseExtraction, EXTRACT_RESPONSE_FORMAT, ExtractionError, liteLLMExtractor, ORDER_ITEM_PROMPT } from "../../lib/intake-extract";
 import { askCard, candidatesCard, enrichedCard, type Block } from "../../lib/intake-blocks";
 
 /**
@@ -298,3 +302,118 @@ describe("liteLLMExtractor — a finish_reason that is present must be `stop`",
     await expect(extractor("tool_calls")([{ kind: "text", text: "x", name: "x" }])).rejects.toThrow(/tool_calls/);
   });
 });
+
+describe("textCouldBeReceipt — nothing that cannot be a receipt reaches the model", () => {
+  const why = (t: string | null | undefined) => {
+    const r = textCouldBeReceipt(t);
+    return r.ok ? null : r.reason;
+  };
+
+  it("the live defect: the phone sent the WORD `Image`, not the image", () => {
+    expect(textCouldBeReceipt("Image").ok).toBe(false);
+    expect(why("Image")).toMatch(/single word/);
+    expect(why("Image\n")).toMatch(/single word/);
+  });
+
+  it("a filename is not a receipt, digits in it notwithstanding", () => {
+    expect(why("IMG_0042.jpeg")).toMatch(/single word/);
+    expect(why("Receipt_19Aug2026.pdf")).toMatch(/single word/);
+  });
+
+  it("empty, blank and whitespace-only", () => {
+    expect(why("")).toMatch(/empty/);
+    expect(why("   \n\t ")).toMatch(/empty/);
+    expect(why(null)).toMatch(/empty/);
+    expect(why(undefined)).toMatch(/empty/);
+  });
+
+  it("prose with no numbers cannot be a receipt", () => {
+    expect(why("Shared from Photos")).toMatch(/no numbers/);
+    expect(why("here is the receipt for dinner last night, sorry it took so long")).toMatch(/no numbers/);
+  });
+
+  it("short but with a printed amount goes through — this door accepts a pasted line", () => {
+    expect(textCouldBeReceipt("TOTAL $17.00").ok).toBe(true);
+    expect(textCouldBeReceipt("Total 17.00").ok).toBe(true);
+    expect(textCouldBeReceipt("Gesamt 12,50 EUR").ok).toBe(true);
+  });
+
+  it("short with digits but no money figure does not — a screenshot name is not a purchase", () => {
+    expect(why("Screenshot 2026 08 22")).toMatch(/too short/);
+    expect(why("2 items")).toMatch(/too short/);
+  });
+
+  it("a real receipt body passes", () => {
+    expect(
+      textCouldBeReceipt("Burgertory (Werribee)\n19 Aug 2026\n1x Michael  14.00\n1x Joel  14.00\nSubtotal 28.00\nTotal 22.40").ok
+    ).toBe(true);
+  });
+
+  it("quoteShared flattens and truncates what goes on the card", () => {
+    expect(quoteShared("Image")).toBe("Image");
+    expect(quoteShared("  a \n b  ")).toBe("a b");
+    expect(quoteShared("x".repeat(200))).toHaveLength(80);
+    expect(quoteShared("x".repeat(200)).endsWith("…")).toBe(true);
+  });
+});
+
+describe("confabulationReason — the belt to that pre-check", () => {
+  const nothing = { textParts: ["Image"], hasImages: false };
+  const readable = { textParts: ["Total $22.40 Burgertory (Werribee) 2026-08-19"], hasImages: false };
+  const order = { merchant: "Burgertory (Braybrook)", items: [{ description: "The Pride" }] };
+
+  it("flags an order extracted from nothing readable", () => {
+    expect(confabulationReason(nothing, order)).toMatch(/nothing shared could be read/);
+  });
+
+  it("does not flag one extracted from something readable", () => {
+    expect(confabulationReason(readable, order)).toBeNull();
+  });
+
+  it("does not flag an image share — the pixels are what we asked for", () => {
+    expect(confabulationReason({ textParts: [], hasImages: true }, order)).toBeNull();
+  });
+
+  it("an empty answer to an unreadable share is honest, not a confabulation", () => {
+    expect(confabulationReason(nothing, { merchant: null, items: [] })).toBeNull();
+  });
+
+  it("the prompt's own placeholder coming back as a merchant is always a confabulation", () => {
+    for (const p of PROMPT_PLACEHOLDER_MERCHANTS) {
+      expect(confabulationReason(readable, { merchant: p, items: [] })).toMatch(/prompt's own example/);
+      expect(confabulationReason({ textParts: [], hasImages: true }, { merchant: p, items: [] })).toMatch(/prompt's own example/);
+    }
+  });
+
+  it("a real merchant that happens to be named in the prompt's platform list is NOT flagged", () => {
+    expect(confabulationReason(readable, { merchant: "Coles (Wyndham Vale)", items: [] })).toBeNull();
+    expect(confabulationReason(readable, { merchant: "Woolworths", items: [] })).toBeNull();
+  });
+});
+
+describe("the prompt carries no real merchant to echo", () => {
+  it("names no shop it could hand back as data", () => {
+    expect(ORDER_ITEM_PROMPT).not.toMatch(/Burgertory/);
+    expect(ORDER_ITEM_PROMPT).not.toMatch(/Wyndham Vale/);
+    expect(ORDER_ITEM_PROMPT).toContain(PROMPT_PLACEHOLDER_MERCHANTS[0]);
+  });
+
+  it("tells the model what to return when the input is not a receipt", () => {
+    expect(ORDER_ITEM_PROMPT).toMatch(/NOT a receipt/);
+    expect(ORDER_ITEM_PROMPT).toMatch(/Do not construct a plausible order/);
+  });
+});
+
+describe("needsHeaderCard", () => {
+  it("names the missing field and sends a dateless in-store screenshot to the app's Share button", async () => {
+    const { needsHeaderCard } = await import("@/lib/intake-blocks");
+    const dateless = needsHeaderCard(["date"], "woolworths").text;
+    expect(dateless).toContain("no date");
+    expect(dateless).toContain("Share button");
+    expect(dateless).not.toContain("total");
+    const delivery = needsHeaderCard(["total"], "uber_eats").text;
+    expect(delivery).toContain("no total");
+    expect(delivery).toContain("part of the screen");
+    expect(needsHeaderCard().text).toContain("total or date");
+  });
+});

src/__tests__/unit/wallet-descriptor.test.ts

@@ -0,0 +1,109 @@
+import { describe, it, expect } from "vitest";
+import {
+  parseWestpacDescriptor,
+  parseWalletSubtitle,
+  currencyFromAmountText,
+} from "@/lib/wallet-descriptor";
+
+/**
+ * Every input below is a real captured notification body or subtitle from
+ * capture.wallet_events (2026-08-13 .. 08-22), not an invented fixture. The
+ * model handled this one fixed-width format three different ways on this exact
+ * data, which is why the parsing moved into code.
+ */
+
+describe("parseWestpacDescriptor", () => {
+  it("reads merchant, city and country from the fixed-width block", () => {
+    expect(parseWestpacDescriptor(
+      "$13.50 paid from Mastercard ending...8032. Breadworld Bakery Cafe     WYNDHAM VAL AU"
+    )).toEqual({ merchant: "Breadworld Bakery Cafe", city: "WYNDHAM VAL", country: "AU" });
+  });
+
+  it("handles a city padded to the full field width", () => {
+    expect(parseWestpacDescriptor(
+      "$54.50 paid from Mastercard ending...8032. PLA DONATION               MELBOURNE   AU"
+    )).toEqual({ merchant: "PLA DONATION", city: "MELBOURNE", country: "AU" });
+  });
+
+  it("splits a 12-char city that runs into the country code with no space", () => {
+    // The case that defeated the model: "SAN FRANCISO" exactly fills the city
+    // field, so "US" follows immediately. Read as one token it becomes the
+    // nonsense location "SAN FRANCISOUS" and the country is lost -- on the only
+    // foreign merchant in the whole capture window.
+    expect(parseWestpacDescriptor(
+      "$42.44 paid from Mastercard ending...8032. SP PEAKDESIGN              SAN FRANCISOUS"
+    )).toEqual({ merchant: "SP PEAKDESIGN", city: "SAN FRANCISO", country: "US" });
+  });
+
+  it("reads mixed-case cities the model inconsistently stripped", () => {
+    expect(parseWestpacDescriptor(
+      "$20.99 paid from Mastercard ending...8032. Spotify P45AE035CB         Sydney      AU"
+    )).toEqual({ merchant: "Spotify P45AE035CB", city: "Sydney", country: "AU" });
+  });
+
+  it("degrades to nulls rather than guessing", () => {
+    expect(parseWestpacDescriptor(null)).toEqual({ merchant: null, city: null, country: null });
+    expect(parseWestpacDescriptor("")).toEqual({ merchant: null, city: null, country: null });
+    // No sentence terminator: not this format at all.
+    expect(parseWestpacDescriptor("You have a new notification"))
+      .toEqual({ merchant: null, city: null, country: null });
+  });
+
+  it("does not invent a country from a short merchant-only tail", () => {
+    // "...IT" here is the end of a word, not Italy. A wrong country on a
+    // foreign transaction is worse than no country.
+    const r = parseWestpacDescriptor("$9.00 paid from Mastercard ending...8032. SOME MERCHANT");
+    expect(r.country).toBeNull();
+  });
+});
+
+describe("parseWalletSubtitle", () => {
+  it("separates merchant from suburb and state", () => {
+    expect(parseWalletSubtitle("Breadworld Bakery Cafe, Wyndham Vale, VIC"))
+      .toEqual({ merchant: "Breadworld Bakery Cafe", location: "Wyndham Vale, VIC" });
+  });
+
+  it("returns no location for an online merchant", () => {
+    // 2 of the 10 Wallet subtitles captured have no place at all.
+    expect(parseWalletSubtitle("Seatsaero")).toEqual({ merchant: "Seatsaero", location: null });
+    expect(parseWalletSubtitle("Prism Accounting Bookkeeping Services"))
+      .toEqual({ merchant: "Prism Accounting Bookkeeping Services", location: null });
+  });
+
+  it("keeps a merchant name containing its own comma out of the location", () => {
+    expect(parseWalletSubtitle("Big Daddy`s Fresh Ma, Manor Lakes, VIC").location)
+      .toBe("Manor Lakes, VIC");
+  });
+
+  it("handles absent input", () => {
+    expect(parseWalletSubtitle(null)).toEqual({ merchant: null, location: null });
+    expect(parseWalletSubtitle("   ")).toEqual({ merchant: null, location: null });
+  });
+});
+
+describe("currencyFromAmountText", () => {
+  it("reads Amex's explicit marker", () => {
+    expect(currencyFromAmountText("A$5,103.49")).toBe("AUD");
+    expect(currencyFromAmountText("A$13.50")).toBe("AUD");
+  });
+
+  it("refuses to guess a bare dollar sign", () => {
+    // 14 of 21 captured transaction rows are this. Assuming AUD is harmless at
+    // home and a 23x error on a THB tap abroad, so the row goes to the
+    // reconcile queue with no currency instead.
+    expect(currencyFromAmountText("$13.50")).toBeNull();
+    expect(currencyFromAmountText("$103.30")).toBeNull();
+  });
+
+  it("reads unambiguous symbols when a card does state one", () => {
+    expect(currencyFromAmountText("€24.00")).toBe("EUR");
+    expect(currencyFromAmountText("฿500")).toBe("THB");
+    expect(currencyFromAmountText("S$18.20")).toBe("SGD");
+    expect(currencyFromAmountText("NZ$40.00")).toBe("NZD");
+  });
+
+  it("handles absent input", () => {
+    expect(currencyFromAmountText(null)).toBeNull();
+    expect(currencyFromAmountText("")).toBeNull();
+  });
+});

src/app/api/annual-charges/route.ts

@@ -0,0 +1,44 @@
+import { NextRequest, NextResponse } from "next/server";
+import { getCurrentUser } from "@/lib/auth";
+import { getUpcomingAnnualCharges } from "@/lib/annual-charges";
+
+/**
+ * Upcoming annual charges, from payments alone.
+ *
+ * Lives here rather than in home-dashboard so the detection sits beside the
+ * ledger it queries, is covered by this repo's integration suite, and can be
+ * reused by a future Slack nudge without moving. home-dashboard reads it over
+ * networks_internal the way its other widgets read their sources.
+ *
+ * Machine access uses the same shared secret as the dashboard's other reads;
+ * a browser session scopes to that user instead.
+ */
+export async function GET(req: NextRequest) {
+  // Its OWN token, with no fallback to WALLET_INGEST_TOKEN. home-dashboard's
+  // compose states the rule plainly -- "this app never holds write creds" -- and
+  // WALLET_INGEST_TOKEN unlocks POST /api/wallet/ingest and the estate-wide
+  // reconcile sweep. Handing that to a read-only surface to save minting a
+  // secret would quietly convert a dashboard into a writer. Unset means closed.
+  const token = process.env.DASH_READ_TOKEN;
+  const supplied = req.headers.get("x-ingest-token");
+  const machine = !!token && !!supplied && supplied === token;
+
+  let ownerId: number | null = null;
+  if (machine) {
+    // A machine caller names the owner explicitly rather than defaulting to 1:
+    // silently assuming an owner is how an ingested order ended up invisible in
+    // every scoped view while sitting in the table.
+    const q = req.nextUrl.searchParams.get("owner_id");
+    ownerId = q !== null && Number.isInteger(Number(q)) ? Number(q) : null;
+    if (ownerId === null) {
+      return NextResponse.json({ error: "owner_id is required for machine access" }, { status: 400 });
+    }
+  } else {
+    const user = await getCurrentUser(req);
+    if (!user) return NextResponse.json({ error: "Unauthorized" }, { status: 403 });
+    ownerId = user.id;
+  }
+
+  const charges = await getUpcomingAnnualCharges(ownerId);
+  return NextResponse.json({ charges, count: charges.length });
+}

src/app/api/slack/interactive/route.ts

@@ -387,12 +387,21 @@ async function nudgeState(transactionId: number) {
     currency: string | null;
     amount: string;
   }>(
-    `SELECT em.merchant_normalized AS merchant,
-            COALESCE(em.currency, 'AUD') AS currency,
+    // LATERAL + the shared tie-break, not a bare join: `queryRow` returns
+    // rows[0] and Postgres guarantees no ordering, so a second metadata row on
+    // one transaction would flip the merchant and currency on every card
+    // refresh at random. Every sibling reader already resolves this way.
+    `SELECT order_ctx.merchant_normalized AS merchant,
+            COALESCE(order_ctx.currency, 'AUD') AS currency,
             t.amount
        FROM transactions t
-       LEFT JOIN expense_metadata em
-         ON em.transaction_id = t.id OR em.matched_transaction_id = t.id
+       LEFT JOIN LATERAL (
+         SELECT em.merchant_normalized, em.currency
+           FROM expense_metadata em
+          WHERE em.transaction_id = t.id OR em.matched_transaction_id = t.id
+          ORDER BY (em.transaction_id = t.id) DESC NULLS LAST, em.id
+          LIMIT 1
+       ) order_ctx ON true
       WHERE t.id = $1`,
     [transactionId]
   );

src/app/api/transactions/[id]/merge/route.ts

@@ -0,0 +1,155 @@
+import { NextRequest, NextResponse } from "next/server";
+import { getCurrentUser } from "@/lib/auth";
+import { queryRaw } from "@/lib/db";
+import { applyReconcileMatch, supersedeTransaction } from "@/lib/reconcile";
+
+/**
+ * Manually resolve a duplicate pair the automation cannot see.
+ *
+ * WHY THIS EXISTS: until now there was no way to merge or remove a transaction
+ * at all. `superseded_by_id` was READ everywhere -- analytics excludes it,
+ * auto-reconcile excludes it -- but written nowhere in application code; it had
+ * only ever been set once, by migration 0023. The reconcile queue was the sole
+ * merge path, and it only offers candidates within a narrow amount band, so a
+ * pair outside it (SP PEAKDESIGN: notified $42.44, charged $43.71 -- the card's
+ * 3% foreign transaction fee) was unresolvable from the UI.
+ *
+ * There is deliberately NO delete. Every child of `transactions` is
+ * ON DELETE CASCADE, so removing a row destroys its splits, tags and metadata
+ * curation. Both modes here keep the row and its children and merely stop it
+ * counting:
+ *
+ *   reconcile -- a statementLESS row folded into its statement twin
+ *                (`reconciled_with_id`). The everyday case.
+ *   supersede -- two rows on the same footing, e.g. a statement imported twice
+ *                (`superseded_by_id`). No constraint on either side's origin.
+ *
+ * Both move curation onto the survivor FIRST, via one shared implementation in
+ * lib/reconcile.ts, because both hide the source row from every figure and
+ * splits stranded on a hidden row silently stop counting.
+ */
+export async function POST(req: NextRequest, { params }: { params: Promise<{ id: string }> }) {
+  const user = await getCurrentUser(req);
+  if (!user) return NextResponse.json({ error: "Unauthorized" }, { status: 403 });
+
+  const { id } = await params;
+  const loserId = Number(id);
+  if (!Number.isInteger(loserId)) {
+    return NextResponse.json({ error: "invalid transaction id" }, { status: 400 });
+  }
+
+  let body: { other_id?: number; mode?: string };
+  try {
+    body = await req.json();
+  } catch {
+    return NextResponse.json({ error: "invalid JSON" }, { status: 400 });
+  }
+
+  const otherId = Number(body.other_id);
+  const mode = body.mode ?? "reconcile";
+  if (!Number.isInteger(otherId)) {
+    return NextResponse.json({ error: "other_id is required" }, { status: 400 });
+  }
+  if (loserId === otherId) {
+    return NextResponse.json({ error: "a transaction cannot merge into itself" }, { status: 400 });
+  }
+  if (mode !== "reconcile" && mode !== "supersede") {
+    return NextResponse.json({ error: "mode must be 'reconcile' or 'supersede'" }, { status: 400 });
+  }
+
+  // Both rows must be the caller's. A statement row carries owner_id on the
+  // statement rather than the transaction, which is why this reads the same
+  // COALESCE every scoped query uses -- checking t.owner_id alone would reject
+  // every statement row as "not found".
+  const rows = await queryRaw<{
+    id: number;
+    statement_id: number | null;
+    reconciled_with_id: number | null;
+    superseded_by_id: number | null;
+    voided_at: Date | null;
+  }>(
+    `SELECT t.id, t.statement_id, t.reconciled_with_id, t.superseded_by_id, t.voided_at
+       FROM transactions t
+       LEFT JOIN statements s ON s.id = t.statement_id
+      WHERE t.id = ANY($1::int[]) AND COALESCE(t.owner_id, s.owner_id) = $2`,
+    [[loserId, otherId], user.id]
+  );
+  if (rows.length !== 2) {
+    return NextResponse.json({ error: "one or both transactions not found" }, { status: 404 });
+  }
+  const loser = rows.find((r) => r.id === loserId)!;
+  const other = rows.find((r) => r.id === otherId)!;
+
+  // Refuse a row that is already folded into something else, in either
+  // direction. Re-pointing it would strand the curation moved on the first
+  // pass -- and silently, since both columns hide the row from every figure.
+  for (const [label, row] of [["source", loser], ["target", other]] as const) {
+    if (row.reconciled_with_id !== null) {
+      return NextResponse.json(
+        { error: `${label} transaction ${row.id} is already reconciled with ${row.reconciled_with_id}` },
+        { status: 409 }
+      );
+    }
+    if (row.superseded_by_id !== null) {
+      return NextResponse.json(
+        { error: `${label} transaction ${row.id} is already superseded by ${row.superseded_by_id}` },
+        { status: 409 }
+      );
+    }
+    // A voided row is a charge that never posted. As a source it is already
+    // out of every figure; as a survivor it would take the other row out too.
+    if (row.voided_at !== null) {
+      return NextResponse.json(
+        { error: `${label} transaction ${row.id} is voided (a charge that never posted)` },
+        { status: 409 }
+      );
+    }
+  }
+
+  try {
+    if (mode === "reconcile") {
+      // applyReconcileMatch's contract, enforced here so a misuse is a 400 and
+      // not a row quietly linked the wrong way round.
+      if (loser.statement_id !== null) {
+        return NextResponse.json(
+          { error: "reconcile expects the source to be a statementless row; use mode 'supersede'" },
+          { status: 400 }
+        );
+      }
+      if (other.statement_id === null) {
+        return NextResponse.json(
+          { error: "reconcile expects the target to be a statement row; use mode 'supersede'" },
+          { status: 400 }
+        );
+      }
+      await applyReconcileMatch(loserId, otherId);
+    } else {
+      // Never hide a statement row behind a statementless one. The statement is
+      // the bank's own record; the statementless row is a notification, a
+      // receipt or a hand entry, and is the side that can be wrong -- a wallet
+      // row carries the amount BEFORE the card's foreign transaction fee, so
+      // burying the statement line under it would leave the ledger holding
+      // $42.44 for a purchase the bank charged $43.71 for, with the statement
+      // no longer reconciling against its own rows. The correct move is the
+      // other direction, which is exactly what `reconcile` does.
+      //
+      // Statement-under-statement stays allowed: that is the imported-twice
+      // pair migration 0023 existed for, and neither side is more true.
+      if (loser.statement_id !== null && other.statement_id === null) {
+        return NextResponse.json(
+          {
+            error:
+              "Refusing to hide a statement row behind a statementless one — the statement is the bank's record. " +
+              `Open transaction ${otherId} instead and fold it into this one.`,
+          },
+          { status: 400 }
+        );
+      }
+      await supersedeTransaction(loserId, otherId);
+    }
+  } catch (e) {
+    return NextResponse.json({ error: e instanceof Error ? e.message : String(e) }, { status: 400 });
+  }
+
+  return NextResponse.json({ ok: true, mode, source: loserId, target: otherId });
+}

src/app/api/transactions/[id]/void/route.ts

@@ -0,0 +1,64 @@
+import { NextRequest, NextResponse } from "next/server";
+import { getCurrentUser } from "@/lib/auth";
+import { queryRaw } from "@/lib/db";
+import { unvoidTransaction, voidTransaction } from "@/lib/reconcile";
+
+/**
+ * POST   /api/transactions/<id>/void  { reason }  -- a charge that never posted
+ * DELETE /api/transactions/<id>/void              -- undo it
+ *
+ * For a statementless row whose charge never reached a statement: an
+ * authorisation hold the merchant released (txn 7951, a cancelled Airbnb
+ * booking). The row stays, with its children and audit trail, and drops out of
+ * every figure and the reconcile queue. Rules and refusals: `voidTransaction`
+ * in lib/reconcile.ts.
+ */
+
+async function ownedRow(req: NextRequest, idParam: string) {
+  const user = await getCurrentUser(req);
+  if (!user) return { error: NextResponse.json({ error: "Unauthorized" }, { status: 403 }) };
+  const id = Number(idParam);
+  if (!Number.isInteger(id)) {
+    return { error: NextResponse.json({ error: "invalid transaction id" }, { status: 400 }) };
+  }
+  // Same owner scope as every other route; a statement row carries owner_id on
+  // the statement. (Statement rows are refused later, but must 400, not 404.)
+  const rows = await queryRaw<{ id: number }>(
+    `SELECT t.id FROM transactions t
+       LEFT JOIN statements s ON s.id = t.statement_id
+      WHERE t.id = $1 AND COALESCE(t.owner_id, s.owner_id) = $2`,
+    [id, user.id]
+  );
+  if (rows.length === 0) {
+    return { error: NextResponse.json({ error: "transaction not found" }, { status: 404 }) };
+  }
+  return { id };
+}
+
+export async function POST(req: NextRequest, { params }: { params: Promise<{ id: string }> }) {
+  const owned = await ownedRow(req, (await params).id);
+  if (owned.error) return owned.error;
+
+  let body: { reason?: unknown };
+  try {
+    body = await req.json();
+  } catch {
+    return NextResponse.json({ error: "invalid JSON" }, { status: 400 });
+  }
+  const reason = typeof body.reason === "string" ? body.reason.trim() : "";
+  if (!reason) return NextResponse.json({ error: "reason is required" }, { status: 400 });
+
+  try {
+    await voidTransaction(owned.id, reason);
+  } catch (e) {
+    return NextResponse.json({ error: e instanceof Error ? e.message : String(e) }, { status: 409 });
+  }
+  return NextResponse.json({ ok: true, id: owned.id, voided: true });
+}
+
+export async function DELETE(req: NextRequest, { params }: { params: Promise<{ id: string }> }) {
+  const owned = await ownedRow(req, (await params).id);
+  if (owned.error) return owned.error;
+  await unvoidTransaction(owned.id);
+  return NextResponse.json({ ok: true, id: owned.id, voided: false });
+}

src/app/api/transactions/reconcile/route.ts

@@ -1,6 +1,8 @@
 import { NextRequest, NextResponse } from "next/server";
 import { getCurrentUser } from "@/lib/auth";
-import { prisma, queryRaw } from "@/lib/db";
+import { queryRaw } from "@/lib/db";
+import { applyReconcileMatch } from "@/lib/reconcile";
+import { findAutoMatches } from "@/lib/auto-reconcile";
 
 export async function POST(req: NextRequest) {
   const user = await getCurrentUser(req);
@@ -17,7 +19,7 @@ export async function POST(req: NextRequest) {
   // Verify all manual_ids belong to this user
   const manualIds = body.matches.map((m) => m.manual_id);
   const owned = await queryRaw<{ id: number }>(
-    `SELECT id FROM transactions WHERE id = ANY($1::int[]) AND statement_id IS NULL AND owner_id = $2`,
+    `SELECT id FROM transactions WHERE id = ANY($1::int[]) AND statement_id IS NULL AND voided_at IS NULL AND owner_id = $2`,
     [manualIds, user.id]
   );
   if (owned.length !== manualIds.length) {
@@ -25,97 +27,71 @@ export async function POST(req: NextRequest) {
   }
 
   let reconciled = 0;
-
   for (const { manual_id, statement_tx_id } of body.matches) {
-    await prisma.$transaction(async (tx) => {
-      // Copy overrides: manual → statement tx
-      const override = await tx.transaction_overrides.findUnique({
-        where: { transaction_id: manual_id },
-      });
-      if (override) {
-        await tx.transaction_overrides.upsert({
-          where: { transaction_id: statement_tx_id },
-          update: {
-            category_override: override.category_override,
-            merchant_normalized: override.merchant_normalized,
-            notes: override.notes,
-            my_share_percent: override.my_share_percent,
-            updated_at: new Date(),
-          },
-          create: {
-            transaction_id: statement_tx_id,
-            category_override: override.category_override,
-            merchant_normalized: override.merchant_normalized,
-            notes: override.notes,
-            my_share_percent: override.my_share_percent,
-          },
-        });
-        await tx.transaction_overrides.deleteMany({ where: { transaction_id: manual_id } });
-      }
-
-      // Move tags: manual → statement tx
-      const tags = await tx.transaction_tags.findMany({ where: { transaction_id: manual_id } });
-      if (tags.length) {
-        await tx.transaction_tags.createMany({
-          data: tags.map((t) => ({ transaction_id: statement_tx_id, tag_id: t.tag_id })),
-          skipDuplicates: true,
-        });
-        await tx.transaction_tags.deleteMany({ where: { transaction_id: manual_id } });
-      }
-
-      // Move splits: manual → statement tx
-      const splits = await tx.transaction_splits.findMany({ where: { transaction_id: manual_id } });
-      if (splits.length) {
-        await tx.transaction_splits.createMany({
-          data: splits.map((s) => ({
-            transaction_id: statement_tx_id,
-            participant_id: s.participant_id,
-            share_percent: s.share_percent,
-          })),
-          skipDuplicates: true,
-        });
-        await tx.transaction_splits.deleteMany({ where: { transaction_id: manual_id } });
-      }
-
-      // Move provenance: manual → statement tx.
-      //
-      // Overrides, tags and splits above were always carried across; expense_metadata was
-      // the one child left behind, which did not matter while every metadata row came from
-      // an email that had created its own transaction. It matters now: a scanned grocery
-      // receipt puts its line items here, and reconciliation hides the manual row from
-      // every figure — so without this the shop's contents disappear at exactly the moment
-      // the statement line appears, and `COLES 0556 MANOR LAKES` stays as unreadable as it
-      // was before the receipt was ever scanned.
-      //
-      // transaction_id is UNIQUE, so a statement row that already has metadata (an emailed
-      // or Paperless copy got there first) keeps it. The pantry row stays attached to the
-      // reconciled manual transaction and is flagged, rather than raising a constraint
-      // violation or silently overwriting the other source.
-      const moved = await tx.$executeRawUnsafe(
-        `UPDATE expense_metadata SET transaction_id = $1
-          WHERE transaction_id = $2
-            AND NOT EXISTS (SELECT 1 FROM expense_metadata other WHERE other.transaction_id = $1)`,
-        statement_tx_id,
-        manual_id
-      );
-      if (moved === 0) {
-        await tx.$executeRawUnsafe(
-          `UPDATE expense_metadata
-              SET flags = coalesce(flags, '[]'::jsonb) || '["metadata_collision_on_reconcile"]'::jsonb
-            WHERE transaction_id = $1`,
-          manual_id
-        );
-      }
-
-      // Mark manual tx as reconciled (link to statement tx)
-      await tx.$executeRawUnsafe(
-        `UPDATE transactions SET reconciled_with_id = $1 WHERE id = $2`,
-        statement_tx_id,
-        manual_id
-      );
-    });
+    await applyReconcileMatch(manual_id, statement_tx_id);
     reconciled++;
   }
 
   return NextResponse.json({ reconciled });
 }
+
+/**
+ * Automatic pass over the reconcile queue.
+ *
+ * Applies only pairs that are unambiguous by the strict rule in
+ * `findAutoMatches` -- exact amount (or exact native pair for a foreign row),
+ * within a few days, and unique in BOTH directions. Everything else is left in
+ * the queue for a person, which is the point: this exists to keep the queue
+ * workable as the wallet lane adds ~63 rows a month, not to clear it.
+ *
+ * Validated against the 48 reconciliations a human had already made: the rule
+ * reproduces all 48 and proposes nothing they did not.
+ *
+ * `?dry=1` returns what it WOULD do without writing, which is how this should
+ * be run the first time after any change to the matching rule.
+ */
+export async function PATCH(req: NextRequest) {
+  // Two callers, two auth modes. A person in the browser reconciles their own
+  // queue; the scheduled sweep has no session and carries the machine token,
+  // and must cover EVERY owner with a queue -- 169 of the 177 rows outstanding
+  // belong to the second participant, so a sweep scoped to one user would leave
+  // almost all of it untouched.
+  const machineToken = req.headers.get("x-ingest-token");
+  const machineOk = !!machineToken && machineToken === process.env.WALLET_INGEST_TOKEN;
+
+  let ownerIds: number[];
+  if (machineOk) {
+    ownerIds = (await queryRaw<{ owner_id: number }>(
+      `SELECT DISTINCT owner_id FROM transactions
+        WHERE statement_id IS NULL AND owner_id IS NOT NULL
+          AND reconciled_with_id IS NULL AND superseded_by_id IS NULL`
+    )).map((r) => r.owner_id);
+  } else {
+    const user = await getCurrentUser(req);
+    if (!user) return NextResponse.json({ error: "Unauthorized" }, { status: 403 });
+    ownerIds = [user.id];
+  }
+
+  const dry = new URL(req.url).searchParams.get("dry") === "1";
+  const matches = (await Promise.all(ownerIds.map((id) => findAutoMatches(id)))).flat();
+
+  if (dry) {
+    return NextResponse.json({ dry_run: true, would_reconcile: matches.length, matches });
+  }
+
+  let reconciled = 0;
+  const failed: { manual_id: number; error: string }[] = [];
+  for (const m of matches) {
+    try {
+      await applyReconcileMatch(m.manual_id, m.statement_tx_id);
+      reconciled++;
+    } catch (e) {
+      // One bad pair must not abandon the rest -- each merge is its own
+      // transaction, so a failure here leaves that pair untouched and in the
+      // queue rather than half-applied.
+      failed.push({ manual_id: m.manual_id, error: e instanceof Error ? e.message : String(e) });
+    }
+  }
+
+  return NextResponse.json({ reconciled, examined: matches.length, failed });
+}

src/app/api/wallet/ingest/route.ts

@@ -0,0 +1,49 @@
+import { NextRequest, NextResponse } from "next/server";
+import { processWalletIngestion, type WalletEvent } from "@/lib/wallet-ingestion";
+
+/**
+ * Machine ingest endpoint for interpreted iOS Wallet notifications.
+ *
+ * Sibling to /api/orders/ingest and /api/receipts/ingest, and deliberately
+ * shaped like them: a shared secret rather than a browser session, because n8n
+ * calls this app-to-app.
+ *
+ * Its own token so the wallet lane's credential can be rotated without touching
+ * the order flow, which runs on a schedule nobody is watching.
+ *
+ * n8n is transport only. The interpretation already happened (Wallet Interpret,
+ * CAp1s5cuzJk2mDaI, local model on Unraid because card tails are PII); this
+ * endpoint owns every decision about what reaches the ledger.
+ */
+function authorised(req: NextRequest): boolean {
+  const expected = process.env.WALLET_INGEST_TOKEN;
+  if (!expected) return false; // fail closed when unconfigured
+  const got = req.headers.get("x-ingest-token");
+  return !!got && got === expected;
+}
+
+export async function POST(req: NextRequest) {
+  if (!authorised(req)) return NextResponse.json({ error: "Unauthorized" }, { status: 401 });
+
+  let body: WalletEvent;
+  try {
+    body = await req.json();
+  } catch {
+    return NextResponse.json({ error: "invalid JSON" }, { status: 400 });
+  }
+
+  if (!Number.isInteger(body?.event_id) || typeof body?.kind !== "string") {
+    return NextResponse.json({ error: "event_id (int) and kind are required" }, { status: 400 });
+  }
+
+  try {
+    const result = await processWalletIngestion(body);
+    return NextResponse.json(result);
+  } catch (e) {
+    // A 500 here alerts #log-alerts through the workflow error handler, which
+    // is what should happen: a write that fails is a template or schema change,
+    // not something to swallow.
+    const message = e instanceof Error ? e.message : String(e);
+    return NextResponse.json({ error: message }, { status: 500 });
+  }
+}

src/app/reconcile/page.tsx

@@ -1,7 +1,7 @@
 "use client";
 
 import { useState } from "react";
-import { usePendingReconciliations, useReconcile } from "@/lib/hooks";
+import { usePendingReconciliations, useReconcile, useVoidTransaction } from "@/lib/hooks";
 import type { ManualTxWithMatches, PotentialMatch } from "@/lib/queries";
 
 function formatDate(d: string | Date) {
@@ -156,6 +156,7 @@ export default function ReconcilePage() {
                     <span key={tag.id} className="w-2 h-2 rounded-full" style={{ backgroundColor: tag.color }} title={tag.name} />
                   ))}
                 </div>
+                <VoidButton id={tx.id} amount={tx.amount} onError={setError} />
               </div>
             ))}
           </div>
@@ -243,6 +244,7 @@ function ReconcileRow({
                 Clear
               </button>
             )}
+            <VoidButton id={tx.id} amount={tx.amount} />
             {!isSkipped && (
               <button onClick={onSkip} className="text-xs text-zinc-500 hover:text-zinc-300 px-2 py-1 rounded hover:bg-zinc-800">
                 Skip (no match)
@@ -296,3 +298,34 @@ function MatchRow({
     </button>
   );
 }
+
+/**
+ * "Never posted": a released authorisation hold, e.g. a booking cancelled
+ * before the charge settled. Drops the row out of every figure and this queue
+ * (migration 0040); the row and its history are kept.
+ */
+function VoidButton({ id, amount, onError }: { id: number; amount: number; onError?: (msg: string) => void }) {
+  const voidTx = useVoidTransaction();
+  async function handle() {
+    const reason = window.prompt(
+      `Mark this $${Number(amount).toFixed(2)} charge as never posted? It will stop counting everywhere.\n\nReason (e.g. "hold released, booking cancelled"):`
+    );
+    if (!reason?.trim()) return;
+    try {
+      await voidTx.mutateAsync({ id, reason });
+    } catch (e) {
+      const msg = e instanceof Error ? e.message : "Void failed";
+      if (onError) onError(msg); else window.alert(msg);
+    }
+  }
+  return (
+    <button
+      onClick={handle}
+      disabled={voidTx.isPending}
+      title="The charge never posted (e.g. a released hold) — stop counting it"
+      className="text-xs text-zinc-500 hover:text-red-300 px-2 py-1 rounded hover:bg-zinc-800 disabled:opacity-50 flex-shrink-0"
+    >
+      {voidTx.isPending ? "Voiding…" : "Never posted"}
+    </button>
+  );
+}

src/app/rules/page.tsx

@@ -463,7 +463,7 @@ export default function RulesPage() {
                     >
                       {expandedRun === run.id ? "▾" : "▸"}
                     </button>
-                    <span className="text-zinc-300">{new Date(run.applied_at).toLocaleString()}</span>
+                    <span className="text-zinc-300">{new Date(run.applied_at).toLocaleString("en-AU")}</span>
                     {/* Which rule ran matters most: a merchant rename and a
                         50/50 split of everything look identical as counts. */}
                     <span className="text-zinc-200">
@@ -480,7 +480,7 @@ export default function RulesPage() {
                     )}
                   </div>
                   {run.reverted_at ? (
-                    <span className="text-xs text-zinc-500">reverted {new Date(run.reverted_at).toLocaleString()}</span>
+                    <span className="text-xs text-zinc-500">reverted {new Date(run.reverted_at).toLocaleString("en-AU")}</span>
                   ) : (
                     <button
                       onClick={() => {

src/app/transactions/page.tsx

@@ -1085,7 +1085,18 @@ function TransactionsContent() {
                     ) : (
                       <>
                         {formatAmount(t.amount_aud ?? t.amount, t.transaction_type)}
-                        {t.currency && t.currency !== "AUD" && (
+                        {/*
+                          The native figure: `foreign_currency_amount` where one
+                          was recorded (a statement or wallet row, whose `amount`
+                          is AUD), else `amount` on a non-AUD statement. Printing
+                          `amount` under the foreign code put the AUD figure
+                          behind an INR label on every wallet row.
+                        */}
+                        {t.foreign_currency_code && t.foreign_currency_code !== "AUD" && t.foreign_currency_amount !== null ? (
+                          <div className="text-[10px] text-zinc-500 mt-0.5">
+                            {formatAmount(t.foreign_currency_amount, t.transaction_type, t.foreign_currency_code)}
+                          </div>
+                        ) : t.currency && t.currency !== "AUD" && (
                           <div className="text-[10px] text-zinc-500 mt-0.5">
                             {formatAmount(t.amount, t.transaction_type, t.currency)}
                           </div>

src/components/edit-transaction-modal.tsx

@@ -1,8 +1,9 @@
 "use client";
 
-import { useState } from "react";
+import { useEffect, useState } from "react";
 import {
   useUpdateTransaction,
+  useMergeTransaction,
   useTags,
   useAddTransactionTag,
   useRemoveTransactionTag,
@@ -93,6 +94,210 @@ function InlineTags({ transactionId, initialTags }: { transactionId: number; ini
   );
 }
 
+/**
+ * Determines which merge mode the backend will accept for (loser, candidate).
+ * Mirrors the constraint enforced server-side in the merge route: reconcile
+ * requires a statementless loser folding into a statement survivor; every
+ * other pairing is a supersede.
+ */
+function mergeModeFor(loser: TransactionRow, candidate: TransactionRow): "reconcile" | "supersede" {
+  return loser.statement_id === null && candidate.statement_id !== null ? "reconcile" : "supersede";
+}
+
+/**
+ * Collapsed by default, at the bottom, visually separate — this is the one
+ * control on the modal that changes which rows this transaction can ever be
+ * again. `transaction` is always the row that will be hidden (the loser);
+ * the row picked here is the survivor (`other_id`).
+ */
+function DuplicateSection({ transaction, onMerged }: { transaction: TransactionRow; onMerged: () => void }) {
+  const [expanded, setExpanded] = useState(false);
+  const [query, setQuery] = useState("");
+  const [results, setResults] = useState<TransactionRow[]>([]);
+  const [searching, setSearching] = useState(false);
+  const [selected, setSelected] = useState<TransactionRow | null>(null);
+  const mergeTxn = useMergeTransaction();
+
+  useEffect(() => {
+    if (!expanded || selected) return;
+    const q = query.trim();
+    // No synchronous setState here: an empty box is handled at render time, and
+    // `searching` is raised inside the timeout. Clearing state here instead
+    // cascades a render on every keystroke (react-hooks lint catches it).
+    if (q.length === 0) return;
+    const handle = setTimeout(() => {
+      setSearching(true);
+      fetch(`/api/transactions?search=${encodeURIComponent(q)}&limit=8`)
+        .then((res) => res.json())
+        .then((body: { data?: TransactionRow[] }) => {
+          setResults((body.data ?? []).filter((r) => r.id !== transaction.id));
+        })
+        .catch(() => setResults([]))
+        .finally(() => setSearching(false));
+    }, 300);
+    return () => clearTimeout(handle);
+  }, [query, expanded, selected, transaction.id]);
+
+  function reset() {
+    setExpanded(false);
+    setQuery("");
+    setResults([]);
+    setSelected(null);
+    mergeTxn.reset();
+  }
+
+  async function handleConfirm() {
+    if (!selected) return;
+    try {
+      await mergeTxn.mutateAsync({
+        id: transaction.id,
+        other_id: selected.id,
+        mode: mergeModeFor(transaction, selected),
+      });
+      onMerged();
+    } catch {
+      // Surfaced below via mergeTxn.error — section stays open so they can retry.
+    }
+  }
+
+  const mode = selected ? mergeModeFor(transaction, selected) : null;
+  // Hiding a statement row behind a statementless one is always backwards — the
+  // bank's record is the side that survives. The API refuses it too; catching it
+  // here just means the button explains itself instead of returning a 400.
+  const wrongDirection =
+    selected !== null && transaction.statement_id !== null && selected.statement_id === null;
+
+  return (
+    <div className="pt-4 border-t border-zinc-800">
+      {!expanded ? (
+        <button
+          type="button"
+          onClick={() => setExpanded(true)}
+          className="text-xs text-zinc-500 hover:text-zinc-300"
+        >
+          Mark as duplicate…
+        </button>
+      ) : (
+        <div className="space-y-2.5">
+          <div className="flex items-center justify-between">
+            <p className="font-semibold text-xs text-zinc-400">Mark as duplicate</p>
+            <button
+              type="button"
+              onClick={reset}
+              className="text-xs text-zinc-500 hover:text-zinc-300"
+            >
+              Cancel
+            </button>
+          </div>
+
+          {!selected ? (
+            <>
+              <input
+                type="text"
+                value={query}
+                onChange={(e) => setQuery(e.target.value)}
+                placeholder="Search for the surviving transaction…"
+                autoFocus
+                className="w-full bg-zinc-800 border border-zinc-700 rounded px-2 py-1.5 text-sm"
+              />
+              {searching && <p className="text-xs text-zinc-500">Searching…</p>}
+              {!searching && query.trim().length > 0 && results.length === 0 && (
+                <p className="text-xs text-zinc-600 italic">No matches</p>
+              )}
+              {/* Gated on the query too: results are no longer cleared when the
+                  box empties, so without this the last search's rows linger. */}
+              {query.trim().length > 0 && results.length > 0 && (
+                <div className="space-y-1 max-h-48 overflow-y-auto">
+                  {results.map((r) => (
+                    <button
+                      key={r.id}
+                      type="button"
+                      onClick={() => setSelected(r)}
+                      className="w-full text-left px-2 py-1.5 rounded bg-zinc-800/50 hover:bg-zinc-800 border border-zinc-700/50"
+                    >
+                      <div className="flex items-center justify-between gap-2">
+                        <span className="text-sm text-zinc-200 truncate">{r.description}</span>
+                        <span
+                          className={`text-sm font-mono shrink-0 ${SPEND_TYPES.has(r.transaction_type) ? "text-red-400" : "text-green-400"}`}
+                        >
+                          {formatAmount(r.amount, r.transaction_type)}
+                        </span>
+                      </div>
+                      <div className="flex items-center justify-between mt-0.5">
+                        <span className="text-xs text-zinc-500">
+                          {new Date(r.transaction_date).toLocaleDateString("en-AU", { day: "numeric", month: "short", year: "numeric" })}
+                        </span>
+                        <span className="text-xs text-zinc-500">
+                          {r.statement_id !== null ? "Statement row" : "Statementless"}
+                        </span>
+                      </div>
+                    </button>
+                  ))}
+                </div>
+              )}
+            </>
+          ) : (
+            <div className="space-y-2">
+              <div className="px-2 py-1.5 rounded bg-zinc-800/50 border border-zinc-700/50">
+                <div className="flex items-center justify-between gap-2">
+                  <span className="text-sm text-zinc-200 truncate">{selected.description}</span>
+                  <span
+                    className={`text-sm font-mono shrink-0 ${SPEND_TYPES.has(selected.transaction_type) ? "text-red-400" : "text-green-400"}`}
+                  >
+                    {formatAmount(selected.amount, selected.transaction_type)}
+                  </span>
+                </div>
+                <div className="flex items-center justify-between mt-0.5">
+                  <span className="text-xs text-zinc-500">
+                    {new Date(selected.transaction_date).toLocaleDateString("en-AU", { day: "numeric", month: "short", year: "numeric" })}
+                  </span>
+                  <span className="text-xs text-zinc-500">
+                    {selected.statement_id !== null ? "Statement row" : "Statementless"}
+                  </span>
+                </div>
+              </div>
+              {wrongDirection ? (
+                <p className="text-xs text-red-400">
+                  This is the wrong way round — the statement is the bank&apos;s record, so it
+                  has to be the row that survives. Close this and open{" "}
+                  <span className="text-zinc-300">{selected.description}</span> instead, then
+                  fold that one into this.
+                </p>
+              ) : (
+                <p className="text-xs text-amber-400/90">
+                  {mode === "reconcile"
+                    ? "This row will be folded into the statement row above."
+                    : "This row will be marked a duplicate of the row above."}
+                </p>
+              )}
+              {mergeTxn.error && (
+                <p className="text-red-400 text-xs">{(mergeTxn.error as Error).message}</p>
+              )}
+              <div className="flex gap-2">
+                <button
+                  type="button"
+                  onClick={() => setSelected(null)}
+                  className="px-3 py-1.5 bg-zinc-800 hover:bg-zinc-700 text-zinc-300 rounded-lg text-xs"
+                >
+                  Choose a different row
+                </button>
+                <button
+                  type="button"
+                  onClick={handleConfirm}
+                  disabled={mergeTxn.isPending || wrongDirection}
+                  className="px-3 py-1.5 bg-amber-700 hover:bg-amber-600 disabled:opacity-50 text-white rounded-lg text-xs"
+                >
+                  {mergeTxn.isPending ? "Merging…" : mode === "reconcile" ? "Fold into statement row" : "Mark as duplicate"}
+                </button>
+              </div>
+            </div>
+          )}
+        </div>
+      )}
+    </div>
+  );
+}
+
 export function EditTransactionModal({
   transaction,
   onClose,
@@ -416,6 +621,8 @@ export function EditTransactionModal({
 
             <OrderDetails transactionId={transaction.id} currency={transaction.currency ?? null} />
 
+            <DuplicateSection transaction={transaction} onMerged={onClose} />
+
           </div>
 
           {/* Footer */}

src/lib/analytics-sql.ts

@@ -71,9 +71,17 @@ export const STATEMENTS_JOIN = `LEFT JOIN statements s ON s.id = t.statement_id`
  * without being edited. Anything summing transactions without this fragment
  * still double-counts — that is the same gap that let reconciled rows into the
  * analytics routes in the first place.
+ *
+ * **3. The charge that never posted** (`voided_at`, migration 0040).
+ *
+ * A card notification is sent for an authorisation, and a released hold never
+ * reaches a statement (txn 7951: a $1,040.41 Airbnb hold on a booking cancelled
+ * the next day). There is no survivor to point at, so it gets its own column —
+ * and it lives in this fragment for the same reason case 2 does.
  */
 export const EXCLUDE_RECONCILED_SOURCE = `NOT (t.statement_id IS NULL AND t.reconciled_with_id IS NOT NULL)
-  AND t.superseded_by_id IS NULL`;
+  AND t.superseded_by_id IS NULL
+  AND t.voided_at IS NULL`;
 
 /**
  * The currency `t.amount` is actually denominated in.
@@ -89,10 +97,17 @@ export const EXCLUDE_RECONCILED_SOURCE = `NOT (t.statement_id IS NULL AND t.reco
  *    figure and `foreign_currency_code` names it (order-ingestion.ts leaves
  *    `amount_aud` NULL rather than asserting an FX rate it does not have).
  *
+ *  - A wallet row also has no statement, but follows the STATEMENT convention:
+ *    `amount` = `amount_aud` = a provisional AUD figure, and the native figure
+ *    lives in `foreign_currency_amount` (wallet-ingestion.ts). `amount_aud`
+ *    being present is what says `amount` is already AUD. Without that arm, an
+ *    INR 524 tap stored as A$9.42 rendered as "INR 9.42" -- the AUD number
+ *    under the foreign label (txn 8162, 2026-10-03).
+ *
  * Reading `s.currency` alone labels every foreign order row as AUD.
  * `s` must be the statements alias in scope.
  */
-export const NATIVE_CURRENCY = `COALESCE(s.currency, t.foreign_currency_code, 'AUD')`;
+export const NATIVE_CURRENCY = `COALESCE(s.currency, CASE WHEN t.amount_aud IS NULL THEN t.foreign_currency_code END, 'AUD')`;
 
 /**
  * True when a row's AUD value is unknown: it is denominated in something other

src/lib/annual-charges.ts

@@ -0,0 +1,275 @@
+import { queryRaw } from "@/lib/db";
+
+/**
+ * Upcoming annual charges, derived from payments alone.
+ *
+ * Independent of the email lane on purpose: a recurring CHARGE is visible in the
+ * ledger whether or not anything was ever emailed about it, which is the only
+ * way a card annual fee or a loan package fee shows up at all -- banks announce
+ * those on the statement, not in a renewal notice.
+ *
+ * WHAT DOES NOT WORK, measured before this was written: "two charges about 365
+ * days apart" is useless on its own. Ranked that way the top hit on this ledger
+ * is a bakery with 202 such pairs, then a donation, Patreon, Coles -- merchants
+ * used often enough that some pair is inevitably a year apart. Frequency is the
+ * confound. Even "every gap is annual" still surfaces a Target visited twice
+ * ($30, then $198) and a Specsavers ($24, then $224). Two visits a year apart
+ * are not a fee.
+ *
+ * So there are two signals, and they are kept apart because they are not equally
+ * trustworthy -- a guess must never render beside a certainty as if it were one:
+ *
+ *   confirmed  the bank called it a fee AND it has repeated on an annual cadence
+ *   likely     the bank called it a fee, seen once -- the kind is certain, the
+ *              cadence is inferred (an annual fee charged once is still annual)
+ *   possible   no fee wording; a merchant whose every gap is annual AND whose
+ *              amount is stable. Catches Amazon Prime, Uber One, VicRoads rego,
+ *              none of which carry fee wording. Also where false positives live.
+ */
+
+export type Confidence = "confirmed" | "likely" | "possible";
+
+/**
+ * `expected_not_seen` is not a lesser version of `upcoming` -- it is the more
+ * interesting row. The estimate has passed and no charge landed, which means one
+ * of three things, and `account_days_stale` is what tells them apart:
+ *   the account stopped being imported  (HSBC: ledger ends 2026-01-07, while the
+ *                                        mail shows statements to 2026-02-07)
+ *   the account genuinely closed
+ *   the charge landed under a different label (VicRoads paid 2026-03-16 as
+ *                                        "VICROADS ONLINE PAYM", not "VICROADS")
+ * Reporting it as upcoming would be a lie; hiding it would lose the signal.
+ */
+export type ChargeStatus = "upcoming" | "expected_not_seen";
+
+export type AnnualCharge = {
+  label: string;
+  status: ChargeStatus;
+  bank_name: string | null;
+  account_number: string | null;
+  confidence: Confidence;
+  observations: number;
+  last_charged: string;
+  last_amount: string;
+  amount_low: string;
+  amount_high: string;
+  median_gap_days: number | null;
+  next_due_estimate: string;
+  /** Ledger liveness for the account: a fee that "stopped" may just be an
+   *  account that stopped being imported. Null for merchant-signal rows, which
+   *  are not tied to one account. */
+  account_last_statement: string | null;
+  account_days_stale: number | null;
+};
+
+/**
+ * `transaction_type = 'fee'` is the universe, not a description match.
+ *
+ * The statement extractor already classifies fees structurally -- 131 rows
+ * across the ledger -- and every card and loan annual fee on this ledger is
+ * typed that way. An earlier version of this filtered `transaction_type =
+ * 'debit'` and returned NONE of them, which is the whole reason this comment
+ * exists: the fees are not debits.
+ *
+ * That universe is broader than "annual", though. It also holds ANZ's monthly
+ * ACCOUNT SERVICING FEE, Zip's MONTHLY ACCOUNT SERVICE FEE, CommBank's MONTHLY
+ * FEE, Citibank LATE PAYMENT FEE, HSBC OVERSEAS TRANSACTION FEE and every Wise
+ * per-transfer charge. Cadence separates them: an annual fee repeats about a
+ * year apart, the rest repeat monthly, on an event, or never (Wise labels each
+ * transfer uniquely, so they can never group to two observations at all).
+ */
+const FEE_TYPE = "fee";
+
+/**
+ * Wording that establishes the KIND when there is only one observation.
+ *
+ * A single charge cannot demonstrate a cadence, but a bank calling something an
+ * annual fee has already told us what it is -- Westpac's CARD FEE $295 and AMP's
+ * Professional Pack Annual Fee each appear once and are certainly annual.
+ * Deliberately narrow: bare "fee" would readmit every per-transaction charge.
+ */
+const ANNUAL_WORDING =
+  "(annual|yearly)[[:space:]]*(card[[:space:]]*)?(fee|charge)" +
+  "|card[[:space:]]+fee" +
+  "|(package|membership|facility)[[:space:]]+fee";
+
+/**
+ * Resolve every statement to a canonical account identity, following
+ * account_links forward. One card can appear under two bank names and two
+ * numbers after an issuer change (Citi Prestige -> NAB MyCard); without this the
+ * $700 fee charged 2024-12-13 and 2025-11-13 reads as two dead one-offs.
+ *
+ * Recursive because the table permits a chain (A->B->C); `uq_account_links_from`
+ * forbids a fork, and `account_links_no_self` a self-link, so the walk
+ * terminates. The depth guard is belt-and-braces against a cycle someone inserts
+ * by hand (A->B, B->A), which no constraint can catch.
+ */
+const RESOLVED_ACCOUNTS = `
+  WITH RECURSIVE chain AS (
+    SELECT s.id AS statement_id, s.bank_name, s.account_number,
+           s.bank_name AS cur_bank, s.account_number_key AS cur_key, 0 AS depth
+    FROM statements s
+    UNION ALL
+    SELECT c.statement_id, c.bank_name, c.account_number,
+           l.to_bank_name, l.to_account_key, c.depth + 1
+    FROM chain c
+    JOIN account_links l
+      ON l.from_bank_name = c.cur_bank AND l.from_account_key = c.cur_key
+    WHERE c.depth < 10
+  ),
+  resolved AS (
+    SELECT DISTINCT ON (statement_id)
+      statement_id, bank_name, account_number, cur_bank AS acct_bank, cur_key AS acct_key
+    FROM chain ORDER BY statement_id, depth DESC
+  )`;
+
+export async function getUpcomingAnnualCharges(ownerId: number): Promise<AnnualCharge[]> {
+  // Signal 1 -- the bank's own fee wording, grouped by resolved account identity
+  // and by the fee's wording so two different fees on one card stay apart
+  // (ANZ charges "ANNUAL FEE" $120 and "ADDITIONAL CARD FEE" $10 on the same
+  // statement, and they are not the same obligation).
+  const feeRows = await queryRaw<AnnualCharge>(
+    `${RESOLVED_ACCOUNTS},
+     fees AS (
+       SELECT r.acct_bank, r.acct_key, r.bank_name, r.account_number,
+              upper(trim(regexp_replace(t.description, '[[:space:]]+', ' ', 'g'))) AS label,
+              t.transaction_date, t.amount
+       FROM transactions t
+       JOIN resolved r ON r.statement_id = t.statement_id
+       JOIN statements s ON s.id = t.statement_id
+       WHERE t.superseded_by_id IS NULL AND t.reconciled_with_id IS NULL
+         AND t.amount > 0
+         AND COALESCE(t.owner_id, s.owner_id) = $1
+         -- Structural first, wording only as a widening: a bank that types a
+         -- fee as a plain debit still gets picked up if it names it annual.
+         AND (t.transaction_type = $3 OR t.description ~* $2)
+     ),
+     gapped AS (
+       SELECT f.*,
+              transaction_date - lag(transaction_date)
+                OVER (PARTITION BY acct_key, label ORDER BY transaction_date) AS gap
+       FROM fees f
+     ),
+     agg AS (
+       SELECT acct_key, acct_bank, label,
+              count(*)::int AS observations,
+              max(transaction_date) AS last_charged,
+              min(amount) AS amount_low, max(amount) AS amount_high,
+              percentile_cont(0.5) WITHIN GROUP (ORDER BY gap)
+                FILTER (WHERE gap IS NOT NULL) AS median_gap,
+              count(*) FILTER (WHERE gap IS NOT NULL) AS gaps,
+              count(*) FILTER (WHERE gap BETWEEN 300 AND 430) AS annual_gaps,
+              (array_agg(bank_name ORDER BY transaction_date DESC))[1] AS bank_name,
+              (array_agg(account_number ORDER BY transaction_date DESC))[1] AS account_number,
+              (array_agg(amount ORDER BY transaction_date DESC))[1] AS last_amount
+       FROM gapped GROUP BY acct_key, acct_bank, label
+     ),
+     liveness AS (
+       SELECT r.acct_key, max(s.billing_end_date) AS last_statement
+       FROM statements s JOIN resolved r ON r.statement_id = s.id
+       GROUP BY r.acct_key
+     )
+     SELECT a.label,
+            -- The account's CURRENT identity, not the bank that happened to
+            -- print the last fee. That $700 fee last appeared under "Citibank"
+            -- in Nov 2025, but the account is NAB/MyCard now and the next one
+            -- will come from there; naming the dead brand would send you to the
+            -- wrong bank.
+            a.acct_bank AS bank_name, a.account_number,
+            CASE WHEN (a.last_charged + COALESCE(round(a.median_gap)::int, 365)) >= CURRENT_DATE
+                 THEN 'upcoming' ELSE 'expected_not_seen' END AS status,
+            -- One observation cannot prove a cadence, but the bank naming it an
+            -- annual fee already tells us the kind. Repeated on an annual gap it
+            -- is confirmed; seen once it is likely and says so.
+            CASE WHEN a.gaps > 0 AND a.annual_gaps = a.gaps THEN 'confirmed'
+                 ELSE 'likely' END AS confidence,
+            a.observations,
+            a.last_charged::text                              AS last_charged,
+            a.last_amount::text                               AS last_amount,
+            a.amount_low::text                                AS amount_low,
+            a.amount_high::text                               AS amount_high,
+            round(a.median_gap)::int                          AS median_gap_days,
+            (a.last_charged + COALESCE(round(a.median_gap)::int, 365))::text AS next_due_estimate,
+            l.last_statement::text                            AS account_last_statement,
+            (CURRENT_DATE - l.last_statement)::int            AS account_days_stale
+     FROM agg a LEFT JOIN liveness l ON l.acct_key = a.acct_key
+     -- Two ways to earn a place, and nothing else gets one. Filtered here, not
+     -- in the UI, so the API can never be read wrongly by a future caller.
+     WHERE
+       -- Repeated on an annual cadence. Excludes ANZ's monthly ACCOUNT
+       -- SERVICING FEE and Zip's MONTHLY ACCOUNT SERVICE FEE, whose gaps are
+       -- ~30 days, and Citibank's LATE PAYMENT FEE, which is event-driven.
+       (a.gaps > 0 AND a.annual_gaps = a.gaps AND a.median_gap >= 300)
+       -- Or seen once, but named as an annual/card/package fee. Westpac's CARD
+       -- FEE and AMP's Professional Pack Annual Fee are each a single charge and
+       -- unmistakably annual. A Wise per-transfer fee is also a single charge
+       -- and says nothing of the sort, which is what keeps it out.
+       OR (a.gaps = 0 AND a.label ~* $2)`,
+    [ownerId, ANNUAL_WORDING, FEE_TYPE]
+  );
+
+  // Signal 2 -- merchant-charged annual items with no fee wording. Requires
+  // BOTH that every gap is annual AND that the amount is stable: VicRoads rego
+  // rises about 3% a year ($825.23 -> $852.88 -> $875.48), so equality is too
+  // strict, while no amount test at all readmits the twice-visited restaurant.
+  // Keyed on merchant, never on card -- that rego was paid on three different
+  // banks, and a card-keyed view would see three unrelated one-offs.
+  const merchantRows = await queryRaw<AnnualCharge>(
+    `WITH t AS (
+       SELECT upper(trim(COALESCE(t.merchant_normalized, t.merchant_name, t.description))) AS label,
+              t.transaction_date, t.amount
+       FROM transactions t
+       LEFT JOIN statements s ON s.id = t.statement_id
+       WHERE t.superseded_by_id IS NULL AND t.reconciled_with_id IS NULL
+         AND t.voided_at IS NULL
+         AND t.transaction_type = 'debit' AND t.amount > 0
+         AND COALESCE(t.owner_id, s.owner_id) = $1
+         AND t.transaction_type = 'debit' AND t.description !~* $2
+     ),
+     gapped AS (
+       SELECT t.*, transaction_date - lag(transaction_date)
+                     OVER (PARTITION BY label ORDER BY transaction_date) AS gap
+       FROM t
+     ),
+     agg AS (
+       SELECT label, count(*)::int AS observations,
+              max(transaction_date) AS last_charged,
+              min(amount) AS amount_low, max(amount) AS amount_high,
+              percentile_cont(0.5) WITHIN GROUP (ORDER BY gap)
+                FILTER (WHERE gap IS NOT NULL) AS median_gap,
+              count(*) FILTER (WHERE gap IS NOT NULL) AS gaps,
+              count(*) FILTER (WHERE gap BETWEEN 300 AND 430) AS annual_gaps,
+              (array_agg(amount ORDER BY transaction_date DESC))[1] AS last_amount
+       FROM gapped GROUP BY label
+     )
+     SELECT label, NULL::text AS bank_name, NULL::text AS account_number,
+            CASE WHEN (last_charged + round(median_gap)::int) >= CURRENT_DATE
+                 THEN 'upcoming' ELSE 'expected_not_seen' END AS status,
+            'possible' AS confidence, observations,
+            last_charged::text AS last_charged,
+            last_amount::text  AS last_amount,
+            amount_low::text   AS amount_low,
+            amount_high::text  AS amount_high,
+            round(median_gap)::int AS median_gap_days,
+            (last_charged + round(median_gap)::int)::text AS next_due_estimate,
+            NULL::text AS account_last_statement,
+            NULL::int  AS account_days_stale
+     FROM agg
+     WHERE gaps >= 1 AND annual_gaps = gaps
+       -- Amount stability is what separates a subscription from a restaurant.
+       -- 20% absorbs CPI drift and an FX-priced renewal without admitting a
+       -- $30-then-$198 pair.
+       AND amount_high <= amount_low * 1.20
+       -- Two observations is the floor for a cadence claim, and this lane has no
+       -- wording to fall back on the way the fee lane does.
+       AND observations >= 2`,
+    [ownerId, ANNUAL_WORDING]
+  );
+
+  // Upcoming first and soonest-first within that, so the thing you can still
+  // act on is at the top; the expected-but-missing rows follow as a review list.
+  const rank = (c: AnnualCharge) => (c.status === "upcoming" ? 0 : 1);
+  return [...feeRows, ...merchantRows].sort(
+    (a, b) => rank(a) - rank(b) || a.next_due_estimate.localeCompare(b.next_due_estimate)
+  );
+}

src/lib/auto-reconcile.ts

@@ -0,0 +1,235 @@
+import { queryRaw, type Queryable } from "@/lib/db";
+import { merchantMatches } from "@/lib/intake-rules";
+
+/**
+ * Unambiguous statement matches for statementless ("manual") transactions.
+ *
+ * WHY THIS EXISTS
+ * ---------------
+ * The reconcile queue is worked by hand and is already ~177 deep. The wallet
+ * lane writes a transaction for every tap (DECISIONS.md ING-11, 2026-08-22),
+ * which at the observed rate adds ~63 rows a month, ~767 a year. Without an
+ * automatic pass for the easy cases that queue stops being workable, and the
+ * tempting alternative -- filtering the new rows back out of it -- is
+ * explicitly ruled out by the `awaitsStatementLine()` removal note in
+ * queries.ts: "an imported row that has no statement belongs in the queue;
+ * volume is solved upstream, never downstream by hiding what is there."
+ *
+ * WHAT COUNTS AS UNAMBIGUOUS
+ * --------------------------
+ * Deliberately much stricter than the queue's own candidate ranking, which
+ * offers the top 5 within +/-1% and +/-3 days for a human to choose between.
+ * A machine gets one shot and no judgement, so:
+ *
+ *   - EXACT amount equality, not the queue's +/-1% band. A near-miss is exactly
+ *     the case a person should look at.
+ *   - Same owner, and the statement row must not already be reconciled against.
+ *   - UNIQUE BOTH WAYS: the manual row must have exactly one candidate, AND
+ *     that candidate must be claimed by exactly one manual row. This is the
+ *     order_transaction_bridge rule, and it is what stops two identical $13.50
+ *     bakery taps in one window from being paired arbitrarily.
+ *
+ * Foreign rows match on the NATIVE pair (foreign_currency_amount +
+ * foreign_currency_code), never on `amount`: `amount` is the AUD the bank
+ * charged after applying its own rate, so a wallet row cannot predict it, while
+ * the native figure is exact on both sides. Same reasoning as
+ * order_transaction_bridge's FX path.
+ *
+ * This function only PROPOSES. Applying a pair goes through the same merge the
+ * manual route uses, so overrides/tags/splits/metadata move identically.
+ */
+
+export type AutoMatch = {
+  manual_id: number;
+  statement_tx_id: number;
+  amount: string;
+  manual_date: string;
+  statement_date: string;
+  description: string;
+  basis: "aud" | "native";
+};
+
+/** Days either side of the manual row's date to consider. Authorisation and
+ *  settlement are rarely the same day; 3 matches the queue's own window and
+ *  the wallet reconciliation check's default. */
+export const AUTO_MATCH_WINDOW_DAYS = 3;
+
+export async function findAutoMatches(
+  ownerId: number,
+  windowDays: number = AUTO_MATCH_WINDOW_DAYS
+): Promise<AutoMatch[]> {
+  return queryRaw<AutoMatch>(
+    `WITH pairs AS (
+       SELECT m.id                                   AS manual_id,
+              t.id                                   AS statement_tx_id,
+              t.amount::text                         AS amount,
+              m.transaction_date::text               AS manual_date,
+              t.transaction_date::text               AS statement_date,
+              t.description                          AS description,
+              CASE WHEN m.foreign_currency_code IS NOT NULL
+                    AND m.foreign_currency_amount IS NOT NULL
+                   THEN 'native' ELSE 'aud' END      AS basis
+         FROM transactions m
+         JOIN transactions t
+           ON t.statement_id IS NOT NULL
+          AND t.transaction_type = 'debit'
+          AND t.transaction_date BETWEEN m.transaction_date - $2::int
+                                     AND m.transaction_date + $2::int
+          AND (
+                -- Foreign: compare what both sides actually recorded natively.
+                (m.foreign_currency_code IS NOT NULL
+                 AND m.foreign_currency_amount IS NOT NULL
+                 AND t.foreign_currency_code   = m.foreign_currency_code
+                 AND t.foreign_currency_amount = m.foreign_currency_amount)
+             OR -- AUD side: exact equality, never a tolerance band. The
+                -- statement row MAY be foreign here, and that is not a mistake:
+                -- all 48 reconciliations a human has made on this ledger are
+                -- exactly this shape -- a manual row carrying no currency,
+                -- matched on exact AUD against a statement row that does carry
+                -- one (EUR, CHF). Overseas receipts and orders record what was
+                -- charged in AUD; only the wallet lane records natively, which
+                -- is why the branch above exists. Requiring both sides to be
+                -- non-foreign was tried and rejected: it matched 0 of those 48.
+                -- The coincidence risk (a domestic tap whose amount happens to
+                -- equal a foreign charge's AUD figure) is already covered by
+                -- unique-both-ways, which refuses as soon as a second candidate
+                -- exists.
+                (m.foreign_currency_code IS NULL
+                 AND t.amount = m.amount)
+              )
+         JOIN statements s ON s.id = t.statement_id
+        WHERE m.statement_id IS NULL
+          AND m.owner_id = $1
+          AND m.reconciled_with_id IS NULL
+          AND m.superseded_by_id IS NULL
+          AND m.voided_at IS NULL
+          AND (m.payment_method IS NULL OR m.payment_method NOT IN ('cash', 'credits'))
+          AND t.superseded_by_id IS NULL
+          AND COALESCE(t.owner_id, s.owner_id) = $1
+          -- The statement row must be unclaimed.
+          AND NOT EXISTS (SELECT 1 FROM transactions mt WHERE mt.reconciled_with_id = t.id)
+     )
+     SELECT manual_id, statement_tx_id, amount, manual_date, statement_date, description, basis
+       FROM pairs p
+      WHERE (SELECT count(*) FROM pairs q WHERE q.manual_id       = p.manual_id)       = 1
+        AND (SELECT count(*) FROM pairs q WHERE q.statement_tx_id = p.statement_tx_id) = 1
+      ORDER BY manual_id`,
+    [ownerId, windowDays]
+  );
+}
+
+/**
+ * An existing statementless transaction that a newly-arriving receipt or order
+ * describes -- i.e. the row the wallet lane already created for this purchase.
+ *
+ * WHY: signals for one purchase arrive in sequence, not in competition. The
+ * wallet notification lands within seconds of the tap; the receipt is shared
+ * hours or days later. Measured on txn 6858: the wallet event arrived 35h25m
+ * before anything created a transaction, and pantry then created one from
+ * scratch -- so the same Woolworths shop would have had two rows. The receipt
+ * must adopt the row that already exists and enrich it.
+ *
+ * Deliberately narrow, for the same reason findAutoMatches() is: exact amount,
+ * a few days either side, unclaimed, and UNIQUE. Two candidates means two
+ * plausible purchases, and inventing a link between the wrong pair moves one
+ * shop's line items onto another shop's row. Returning null is always safe --
+ * the caller falls back to creating its own row, which is today's behaviour.
+ */
+/**
+ * The row a newly-arriving WALLET notification describes -- the mirror of
+ * findAdoptableTransaction, for the case where the signals arrived in the other
+ * order.
+ *
+ * WHY: the sequencing this estate is built on -- wallet first, receipt second,
+ * statement last -- is a tendency, not a guarantee. The wallet lane can be
+ * parked (its first 21 rows were backfilled days late), the interpreter runs
+ * every 15 minutes, the phone can be offline. Whenever a receipt reaches the
+ * ledger before the notification for the same tap, the wallet writer used to
+ * insert unconditionally: measured on the 2026-08-18 Woolworths shop, txn 6858
+ * (pantry, 23 items) and txn 6927 (wallet) were the same $103.30 purchase, and
+ * the second sat in the reconcile queue waiting for a statement line that the
+ * first had already claimed.
+ *
+ * Deliberately NOT the same predicate as findAdoptableTransaction, in the one
+ * way that matters: that one refuses a row which already carries line items,
+ * because for a receipt such a row is a DIFFERENT purchase that happens to
+ * share an amount. Here it is the opposite -- an itemised row is the receipt we
+ * are trying to fold into, and refusing it is the bug.
+ *
+ * Losing that guard means the merchant has to supply it. A wallet row adopts
+ * only when the merchant agrees (the same whole-word containment the intake
+ * matcher uses: "Woolworths" vs "WOOLWORTHS 3345"), because amount-and-date
+ * alone would let a notification silently disappear into an unrelated purchase
+ * of the same value -- and a wallet row that is never written is a tap with no
+ * trace, which is worse than a visible duplicate. When in doubt, return null:
+ * the caller then writes its own row, which is the pre-existing behaviour.
+ */
+export async function findWalletAdoptableTransaction(
+  ownerId: number,
+  amount: number,
+  date: string,
+  merchant: string | null,
+  windowDays: number = AUTO_MATCH_WINDOW_DAYS,
+  // wallet-ingestion calls this while holding its advisory lock, and must pass
+  // that transaction's connection: borrowing a second pooled one per request
+  // is how a burst of concurrent posts exhausts the pool and deadlocks.
+  db: Pick<Queryable, "queryRaw"> = { queryRaw }
+): Promise<{ id: number; metadata_id: number | null } | null> {
+  // No usable merchant, no adoption. "Card purchase" is wallet-ingestion's own
+  // fallback for a notification that named no merchant at all.
+  if (!merchant || merchant === "Card purchase") return null;
+  const rows = await db.queryRaw<{ id: number; metadata_id: number | null; merchant: string | null }>(
+    `SELECT t.id, em.id AS metadata_id,
+            COALESCE(em.merchant_normalized, t.merchant_normalized, t.merchant_name, t.description) AS merchant
+       FROM transactions t
+       LEFT JOIN expense_metadata em ON em.transaction_id = t.id
+      WHERE t.statement_id IS NULL
+        AND t.owner_id = $1
+        AND t.amount = $2::numeric
+        AND t.transaction_date BETWEEN $3::date - $4::int AND $3::date + $4::int
+        AND t.reconciled_with_id IS NULL
+        AND t.superseded_by_id IS NULL
+        -- A voided row is a charge that never posted; a real charge for the same
+        -- amount (a rebooking after a released hold) must not fold into it.
+        AND t.voided_at IS NULL
+        -- A wallet row is the OTHER HALF of this tap, not a purchase to fold
+        -- into; findPairedRow() merges those, with different rules.
+        AND COALESCE(t.source, '') <> 'wallet'
+      LIMIT 5`,
+    [ownerId, amount, date, windowDays]
+  );
+  const hits = rows.filter((r) => merchantMatches(r.merchant, merchant));
+  return hits.length === 1 ? { id: hits[0].id, metadata_id: hits[0].metadata_id } : null;
+}
+
+export async function findAdoptableTransaction(
+  ownerId: number,
+  amount: number,
+  date: string,
+  windowDays: number = AUTO_MATCH_WINDOW_DAYS
+): Promise<number | null> {
+  const rows = await queryRaw<{ id: number }>(
+    `SELECT t.id
+       FROM transactions t
+      WHERE t.statement_id IS NULL
+        AND t.owner_id = $1
+        AND t.amount = $2::numeric
+        AND t.transaction_date BETWEEN $3::date - $4::int AND $3::date + $4::int
+        AND t.reconciled_with_id IS NULL
+        AND t.superseded_by_id IS NULL
+        -- A voided row is a charge that never posted; a real charge for the same
+        -- amount (a rebooking after a released hold) must not fold into it.
+        AND t.voided_at IS NULL
+        -- Never adopt a row that another source already fully described; that
+        -- is a different purchase that happens to share an amount.
+        AND NOT EXISTS (
+          SELECT 1 FROM expense_metadata em
+           WHERE (em.transaction_id = t.id OR em.matched_transaction_id = t.id)
+             AND jsonb_typeof(em.line_items) = 'array'
+             AND jsonb_array_length(em.line_items) > 0
+        )
+      LIMIT 2`,
+    [ownerId, amount, date, windowDays]
+  );
+  return rows.length === 1 ? rows[0].id : null;
+}

src/lib/db.ts

@@ -38,7 +38,12 @@ export type Queryable = {
  * next statement runs. Anything that needs a lock to still be held on its next
  * query has to be in here.
  */
-export async function withTransaction<T>(fn: (q: Queryable) => Promise<T>): Promise<T> {
+export async function withTransaction<T>(
+  fn: (q: Queryable) => Promise<T>,
+  // Prisma's defaults (maxWait 2s, timeout 5s) count time spent WAITING on an
+  // advisory lock, so a caller that serialises a burst on one needs more.
+  opts?: { maxWait?: number; timeout?: number }
+): Promise<T> {
   return prisma.$transaction(async (tx) => {
     const q: Queryable = {
       queryRaw: <R>(sql: string, params: unknown[] = []) => tx.$queryRawUnsafe<R[]>(sql, ...params),
@@ -48,5 +53,5 @@ export async function withTransaction<T>(fn: (q: Queryable) => Promise<T>): Prom
       },
     };
     return fn(q);
-  });
+  }, opts);
 }

src/lib/hooks.ts

@@ -177,6 +177,40 @@ export function useUpdateTransaction() {
   });
 }
 
+export function useMergeTransaction() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: async ({
+      id,
+      other_id,
+      mode,
+    }: {
+      id: number;
+      other_id: number;
+      mode: "reconcile" | "supersede";
+    }) => {
+      const res = await fetch(`/api/transactions/${id}/merge`, {
+        method: "POST",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({ other_id, mode }),
+      });
+      if (!res.ok) {
+        const body = await res.json().catch(() => ({}));
+        throw new Error(body.error || "Failed to merge transaction");
+      }
+      return res.json();
+    },
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: ["transactions"] });
+      qc.invalidateQueries({ queryKey: ["transaction"] });
+      qc.invalidateQueries({ queryKey: ["splits"] });
+      qc.invalidateQueries({ queryKey: ["analytics"] });
+      qc.invalidateQueries({ queryKey: ["participant-balances"] });
+      qc.invalidateQueries({ queryKey: ["shared-transactions"] });
+    },
+  });
+}
+
 export function useBulkAction() {
   const qc = useQueryClient();
   return useMutation({
@@ -1128,6 +1162,27 @@ export function useReconcile() {
   });
 }
 
+/** Mark a statementless row as a charge that never posted (migration 0040). */
+export function useVoidTransaction() {
+  const qc = useQueryClient();
+  return useMutation({
+    mutationFn: async ({ id, reason }: { id: number; reason: string }) => {
+      const res = await fetch(`/api/transactions/${id}/void`, {
+        method: "POST",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({ reason }),
+      });
+      if (!res.ok) throw new Error((await res.json()).error || "Void failed");
+      return res.json();
+    },
+    onSuccess: () => {
+      qc.invalidateQueries({ queryKey: ["transactions"] });
+      qc.invalidateQueries({ queryKey: ["reconcile-pending"] });
+      qc.invalidateQueries({ queryKey: ["analytics"] });
+    },
+  });
+}
+
 export function useMerchantTransactions(merchant: string | null) {
   return useQuery<{ transactions: MerchantTxnRow[] }>({
     queryKey: ["analytics", "merchant-txns", merchant],

src/lib/intake-blocks.ts

@@ -120,13 +120,51 @@ export function candidatesCard(c: CardInput, candidates: Candidate[]): { blocks:
   return { blocks, text };
 }
 
-export function forwardedCard(c: CardInput & { captureId: number; enriching: boolean }): { blocks: Block[]; text: string } {
-  const text = `${esc(c.merchant) || "Grocery receipt"} ${c.total !== null ? money(c.total, c.currency) : ""} → pantry board (${c.itemCount} item${c.itemCount === 1 ? "" : "s"}) — finance ${c.enriching ? "updates" : "books it"} on confirm`.replace(/\s+/g, " ");
+export function forwardedCard(
+  c: CardInput & {
+    captureId: number;
+    enriching: boolean;
+    /** Pantry already holds this exact receipt (same barcode tuple). */
+    alreadyCaptured?: { captureId: number; confirmed: boolean } | null;
+    /** Finance found more than one row this could be, so it picked none. */
+    ambiguous?: boolean;
+  }
+): { blocks: Block[]; text: string } {
+  const what = `${esc(c.merchant) || "Grocery receipt"} ${c.total !== null ? money(c.total, c.currency) : ""}`.trim();
+  const items = `${c.itemCount} item${c.itemCount === 1 ? "" : "s"}`;
+  // The warning comes FIRST and the emoji changes, because this card is read at
+  // a glance on a phone: a duplicate warning appended to a cheerful "→ pantry
+  // board" line is a duplicate warning nobody reads. Confirming both copies
+  // counts the same groceries twice — pantry's addItem accumulates — so the
+  // card has to say which capture to dismiss, not merely that there are two.
+  if (c.alreadyCaptured) {
+    const state = c.alreadyCaptured.confirmed ? "already confirmed" : "still waiting to be confirmed";
+    const text = `Already on the pantry board — ${what} is capture ${c.alreadyCaptured.captureId} (${state}). This upload is capture ${c.captureId}; dismiss it unless it replaces the first, because confirming both counts the same groceries twice.`.replace(/\s+/g, " ");
+    return { text, blocks: [section(`:repeat: ${text}`)] };
+  }
+  const tail = c.ambiguous
+    ? "— finance found more than one row it could be, so it picked none; reconcile it there"
+    : `— finance ${c.enriching ? "updates" : "books it"} on confirm`;
+  const text = `${what} → pantry board (${items}) ${tail}`.replace(/\s+/g, " ");
   return { text, blocks: [section(`:shopping_trolley: ${text}`)] };
 }
 
-export function needsHeaderCard(): { blocks: Block[]; text: string } {
-  const text = "I can see items but no order total or date — share the top of the screen too so I can match it.";
+/**
+ * Names what is missing, because the advice differs. A delivery-app screenshot
+ * that lost its header needs the top of the screen; a Woolworths eReceipt
+ * SCREENSHOT has no date anywhere on it (seen 2026-08-22 — store, total, items,
+ * payment, and not one date), while the app's own Share button sends the raw
+ * receipt WITH the date. So for a dateless in-store receipt the right advice is
+ * the Share button, or pantry's own upload — not "share more of the screen".
+ */
+export function needsHeaderCard(missing: string[] = ["total", "date"], source?: string | null): { blocks: Block[]; text: string } {
+  const what = missing.length ? missing.join(" or ") : "total or date";
+  const inStore = source === "coles" || source === "woolworths";
+  const hint =
+    missing.length === 1 && missing[0] === "date" && inStore
+      ? "an app screenshot carries no date — use the app's Share button instead, or upload the picture to pantry directly"
+      : "share the part of the screen that shows it";
+  const text = `I can see items but no ${esc(what)} — ${hint}, so I can match it.`;
   return { text, blocks: [section(`:warning: ${text}`)] };
 }
 
@@ -155,7 +193,12 @@ export function askCard(c: CardInput): { blocks: Block[]; text: string } {
 
 export function duplicateCard(c: { merchant: string | null; total: number | null; transactionId: number | null; captureId: number | null; batchId: string | null; currency?: string | null }): { blocks: Block[]; text: string } {
   const where = c.transactionId ? `transaction ${c.transactionId}` : c.captureId ? `pantry capture ${c.captureId}` : c.batchId ? `batch ${esc(c.batchId)}` : "an earlier upload";
-  const text = `Already handled — ${esc(c.merchant) || "this order"} ${c.total !== null ? money(c.total, c.currency) : ""} is ${where}. Nothing was changed.`.replace(/\s+/g, " ");
+  // The header (merchant, date, total) IS the identity of a purchase here, so a
+  // genuine SECOND identical purchase — two of the same coffee, one day, one shop
+  // — reads as a re-share and is refused. That is the right default (a re-share is
+  // far commoner than a repeat) but it is a dead end without this sentence: the
+  // card must say what to do instead, or the ledger quietly loses the second one.
+  const text = `Already handled — ${esc(c.merchant) || "this order"} ${c.total !== null ? money(c.total, c.currency) : ""} is ${where}. Nothing was changed — if this really is a second, separate purchase, add it in the app.`.replace(/\s+/g, " ");
   return { text, blocks: [section(`:repeat: ${text}`)] };
 }
 

src/lib/intake-extract.ts

@@ -1,5 +1,5 @@
 import { extractText, getDocumentProxy } from "unpdf";
-import type { ExtractedOrder, ExtractedItem } from "@/lib/intake-rules";
+import { PROMPT_PLACEHOLDER_MERCHANTS, type ExtractedOrder, type ExtractedItem } from "@/lib/intake-rules";
 
 /**
  * Extraction for the intake router: one joint model call per batch.
@@ -31,13 +31,22 @@ export async function extractPdfText(bytes: Buffer): Promise<string> {
   return (Array.isArray(text) ? text.join("\n") : text).trim();
 }
 
+/**
+ * The merchant examples are FORMAT PLACEHOLDERS, never real shops
+ * (`PROMPT_PLACEHOLDER_MERCHANTS`). Given an input with nothing in it, a model
+ * still answers, and it answers out of its prompt: on 2026-08-22 the word
+ * `Image` produced a complete Burgertory receipt because the prompt named
+ * Burgertory. The "not a receipt" rule below and the placeholder names are the
+ * two halves of closing that.
+ */
 export const ORDER_ITEM_PROMPT = `You are reading a purchase receipt: an online food or grocery order (Uber Eats, DoorDash, Coles, Woolworths, ...) or an in-store till receipt. It arrives as one or more screenshots in scroll order, as the text of a PDF, or as pasted text.
 
 Return ONE JSON object with exactly these keys:
 { "merchant", "order_date", "total", "subtotal", "currency", "items": [{ "quantity", "description", "unit_price", "line_total", "options" }], "adjustments": [{ "label", "amount" }] }
 
 Rules:
-- merchant: the store or restaurant name exactly as printed, including any branch in parentheses (e.g. "Coles (Wyndham Vale)", "Burgertory (Werribee)"). Never the delivery platform. null if not shown.
+- If what you are given is NOT a receipt or an order — it is blank, a filename, a single word such as "Image", an app screen with no purchase on it, or anything else you cannot read a purchase off — return { "merchant": null, "order_date": null, "total": null, "subtotal": null, "currency": null, "items": [], "adjustments": [] }. Do not construct a plausible order. Nothing in these instructions is data: never return an example from this prompt as if it were printed on the input.
+- merchant: the store or restaurant name exactly as printed, including any branch in parentheses (format: "${PROMPT_PLACEHOLDER_MERCHANTS[0]}"). Never the delivery platform. null if not shown.
 - order_date: the order or purchase date printed, as YYYY-MM-DD. null if not shown.
 - total: the amount actually charged — the "Total" / amount paid — as a number. null if not shown.
 - subtotal: the printed item subtotal before fees and discounts, as a number, or null.

src/lib/intake-pantry.ts

@@ -20,7 +20,21 @@ export type PantryForwardArgs = {
   financeAction: "enrich" | "create";
 };
 
-export type PantryForwardResult = { captureId: number; itemCount: number | null };
+/**
+ * `alreadyCaptured` is pantry telling us this exact receipt is already on its
+ * board -- same store, register, receipt number and date, which is the tuple
+ * both Coles and Woolworths encode in their barcode. Pantry deliberately still
+ * ACCEPTS the second capture (the first may have been a bad crop being
+ * replaced) and says so instead of refusing; dropping the field on the floor
+ * turned "said so" into silence, and a duplicate that is not shown is a
+ * duplicate that gets confirmed -- which double-counts the groceries, because
+ * pantry's addItem accumulates.
+ */
+export type PantryForwardResult = {
+  captureId: number;
+  itemCount: number | null;
+  alreadyCaptured: { captureId: number; confirmed: boolean } | null;
+};
 
 export class PantryDuplicateError extends Error {
   constructor(
@@ -65,12 +79,23 @@ export function pantryForwarder(opts: { baseUrl?: string; token?: string; fetchI
     } catch (e) {
       throw new PantryUnavailableError(`pantry unreachable: ${e instanceof Error ? e.message : String(e)}`);
     }
-    const body = (await response.json().catch(() => ({}))) as { captureId?: number; error?: string; items?: unknown[]; reason?: string };
+    const body = (await response.json().catch(() => ({}))) as {
+      captureId?: number; error?: string; items?: unknown[]; reason?: string;
+      alreadyCaptured?: { captureId?: number; confirmed?: boolean } | null;
+    };
     if (response.status === 409) {
       throw new PantryDuplicateError(typeof body.captureId === "number" ? body.captureId : null, body.reason ?? body.error ?? "already captured");
     }
     if (!response.ok) throw new PantryUnavailableError(`pantry returned ${response.status}: ${body.error ?? ""}`.trim());
     if (typeof body.captureId !== "number") throw new PantryUnavailableError("pantry returned no captureId");
-    return { captureId: body.captureId, itemCount: Array.isArray(body.items) ? body.items.length : null };
+    const prior = body.alreadyCaptured;
+    return {
+      captureId: body.captureId,
+      itemCount: Array.isArray(body.items) ? body.items.length : null,
+      alreadyCaptured:
+        prior && typeof prior.captureId === "number" && prior.captureId !== body.captureId
+          ? { captureId: prior.captureId, confirmed: prior.confirmed === true }
+          : null,
+    };
   };
 }

src/lib/intake-router.ts

@@ -2,9 +2,13 @@ import { randomUUID } from "node:crypto";
 import { queryRaw, queryRow, withTransaction } from "@/lib/db";
 import { EXCLUDE_RECONCILED_SOURCE, NATIVE_CURRENCY } from "@/lib/analytics-sql";
 import { DEFAULT_OWNER_ID } from "@/lib/order-ingestion";
+import { findAdoptableTransaction } from "@/lib/auto-reconcile";
 import {
   classifyInput,
+  confabulationReason,
   merchantKey,
+  quoteShared,
+  textCouldBeReceipt,
   merchantMatches,
   platformForSource,
   reconcileItems,
@@ -373,7 +377,12 @@ export async function routeFood(ctx: RouteContext): Promise<IntakeResult> {
 
 /** Grocery: the original files go to pantry's board; finance rides along on its confirm (C3/C4). */
 export async function routeGrocery(ctx: RouteContext, files: StoredFile[], pantry: PantryForward): Promise<IntakeResult> {
-  const { match } = await matchTransaction(ctx.header, ctx.platform);
+  // `ambiguous` is not a miss. Two ledger rows for one shop is precisely the
+  // state that produced the 2026-08-18 Woolworths double-book, and forwarding
+  // with financeAction 'create' as though nothing matched is what turns it into
+  // a third row on confirm. The groceries still go to the board — they have to,
+  // it is where they are worked — but the card must say finance could not pick.
+  const { match, ambiguous } = await matchTransaction(ctx.header, ctx.platform);
   try {
     const r = await pantry({
       files,
@@ -384,7 +393,14 @@ export async function routeGrocery(ctx: RouteContext, files: StoredFile[], pantr
       financeAction: match ? "enrich" : "create",
     });
     await updateBatch(ctx.batchId, { kind: "grocery", status: "forwarded_to_pantry", transaction_id: match?.transaction_id ?? null, capture_id: r.captureId, items: ctx.items });
-    const card = forwardedCard({ ...cardInput(ctx), itemCount: r.itemCount ?? ctx.items.length, captureId: r.captureId, enriching: !!match });
+    const card = forwardedCard({
+      ...cardInput(ctx),
+      itemCount: r.itemCount ?? ctx.items.length,
+      captureId: r.captureId,
+      enriching: !!match,
+      alreadyCaptured: r.alreadyCaptured,
+      ambiguous,
+    });
     return finish(ctx, "grocery", "forwarded_to_pantry", card, { capture_id: r.captureId, transaction_id: match?.transaction_id });
   } catch (e) {
     if (e instanceof PantryDuplicateError) {
@@ -444,6 +460,29 @@ export async function createFromBatch(batch: BatchRow, kind: "grocery" | "food")
   }
 
   const ownerId = await ownerForEmail(batch.shared_by);
+
+  // Before minting a row: is one already here that nothing has described yet?
+  //
+  // matchTransaction above cannot see it. Its key includes the platform, and a
+  // wallet notification has none — so an Uber Eats screenshot never matches the
+  // wallet row for the very tap that paid for it, and the Create button books
+  // the purchase a second time. That is the normal arrival order, not an edge
+  // case: the notification lands within seconds of the tap and the screenshot
+  // is shared later.
+  //
+  // findAdoptableTransaction is the same helper the receipt lane uses and
+  // carries the same guards — exact amount, ±3 days, unclaimed, UNIQUE, and
+  // never a row that already carries line items. Its uniqueness check runs
+  // outside the advisory lock below, so two simultaneous taps could both adopt;
+  // the second would then find the items in place and fall through to the lock,
+  // which is where duplicates are actually resolved.
+  const adopted = await findAdoptableTransaction(ownerId, header.total, header.order_date);
+  if (adopted) {
+    await replaceLineItems(adopted, items, { source: "intake", platform: header.platform, batchId: batch.id, currency: header.currency });
+    await updateBatch(batch.id, { kind, status: "enriched", transaction_id: adopted, items });
+    return finish(ctx, kind, "enriched", enrichedCard({ ...cardInput(ctx), transactionId: adopted }), { transaction_id: adopted });
+  }
+
   const merchant = header.merchant ?? "Unknown merchant";
   // Not `orderDescription()`: it always prefixes "Order - ", which is right for
   // a delivery receipt and wrong for the in-store till receipt this path also
@@ -548,6 +587,8 @@ export async function runIntake(input: IntakeInput, deps: IntakeDeps): Promise<I
   // 1. classify each file and build the parts, in the order shared.
   const parts: InputPart[] = [];
   const usable: StoredFile[] = [];
+  /** Text parts dropped before the model saw them — see `textCouldBeReceipt`. */
+  const rejectedText: { name: string; text: string; reason: string }[] = [];
   for (const f of input.files) {
     const kind = classifyInput(f);
     if (kind === "pdf") {
@@ -565,7 +606,20 @@ export async function runIntake(input: IntakeInput, deps: IntakeDeps): Promise<I
     } else if (kind === "image") {
       parts.push({ kind: "image", mime: f.type || "image/jpeg", bytes: f.bytes, name: f.name });
     } else if (kind === "text") {
-      parts.push({ kind: "text", text: f.bytes.toString("utf-8"), name: f.name });
+      // Nothing that cannot be a receipt reaches the model. A share that
+      // arrives as the word "Image" (the phone coerced the photo to text) is
+      // not a thin receipt, and asking anyway produces a whole invented one —
+      // 2026-08-22, `textCouldBeReceipt`. Dropped like an unreadable file type
+      // rather than failing the batch, so one junk text part beside a good
+      // screenshot still goes through; if nothing usable is left the card below
+      // quotes what the phone actually sent.
+      const text = f.bytes.toString("utf-8");
+      const check = textCouldBeReceipt(text);
+      if (!check.ok) {
+        rejectedText.push({ name: f.name, text, reason: check.reason });
+        continue;
+      }
+      parts.push({ kind: "text", text, name: f.name });
     } else {
       continue;
     }
@@ -573,7 +627,16 @@ export async function runIntake(input: IntakeInput, deps: IntakeDeps): Promise<I
   }
   if (!parts.length) {
     await updateBatch(batchId, { kind: "unsupported", status: "unsupported" });
-    return fail("unsupported", "unsupported", unsupportedCard());
+    const dropped = rejectedText[0];
+    return fail(
+      "unsupported",
+      "unsupported",
+      unsupportedCard(
+        dropped
+          ? `the share arrived as text and ${dropped.reason}: "${quoteShared(dropped.text)}" — send the picture or the PDF itself`
+          : undefined
+      )
+    );
   }
 
   // A re-share of the very same files, before spending a model call on it.
@@ -604,11 +667,32 @@ export async function runIntake(input: IntakeInput, deps: IntakeDeps): Promise<I
   }
   const items = toLineItems(extracted.items);
 
+  // 2b. did it answer out of the upload, or out of its prompt?
+  const confab = confabulationReason(
+    {
+      textParts: parts.filter((p): p is Extract<InputPart, { kind: "text" }> => p.kind === "text").map((p) => p.text),
+      hasImages: parts.some((p) => p.kind === "image"),
+    },
+    extracted
+  );
+  if (confab) {
+    await updateBatch(batchId, { kind: "unsupported", status: "unsupported" });
+    return fail("unsupported", "unsupported", unsupportedCard(confab));
+  }
+
   // 3. header.
+  //
+  // This path stores the partial header and the items on the batch row, which
+  // is right for a genuine screenshot missing its top — the next tap has
+  // something to work from. It is wrong for a confabulation, and on 2026-08-22
+  // it stored one: an invented Burgertory header with two invented lines, held
+  // back from the matcher only by an absent date. Nothing here could tell the
+  // two apart, which is why the pre-check above and 2b now keep confabulations
+  // from reaching this point at all rather than trying to filter them here.
   const headerCheck = validateHeader(extracted);
   if (!headerCheck.ok) {
     await updateBatch(batchId, { header: { merchant: extracted.merchant, order_date: extracted.order_date, total: extracted.total, platform }, items, status: "needs_header" });
-    return fail("held", "needs_header", needsHeaderCard(), 200, { items });
+    return fail("held", "needs_header", needsHeaderCard(headerCheck.missing, input.source), 200, { items });
   }
   const header = headerCheck.header;
   const storedHeader = { ...header, platform, subtotal: extracted.subtotal ?? null, currency: extracted.currency ?? "AUD" };

src/lib/intake-rules.ts

@@ -250,3 +250,94 @@ export function merchantMatches(stored: string | null | undefined, extracted: st
   if (!a.length || !b.length) return false;
   return containsWordRun(a, b) || containsWordRun(b, a);
 }
+
+// ─── is this text even a receipt? (2026-08-22) ───────────────────────────────
+
+/**
+ * The merchant names the extraction prompt uses to show the FORMAT. They are
+ * placeholders rather than real shops on purpose.
+ *
+ * On 2026-08-22 an iOS Shortcut coerced a shared photo to text, so the intake
+ * router received one text part whose entire content was the five characters
+ * `Image`. The model was handed that, and returned a whole receipt:
+ * `Burgertory (Braybrook)`, two "The Pride" lines at $22.80, a $41.15 total.
+ * Nothing on the receipt existed — "Burgertory" was in the prompt, as the
+ * example of how to print a branch in parentheses. Only the absent date kept it
+ * out of the matcher; the invented header and items were stored on the batch.
+ *
+ * Same shape as the ingestion engine's 2026-08-05 empty-body defect: a stage
+ * that can be handed nothing has to say so, because the model downstream will
+ * always answer, and it answers out of its prompt.
+ *
+ * Exported from here rather than written into the prompt string so the prompt
+ * and the guard that rejects an echo of it cannot drift apart.
+ */
+export const PROMPT_PLACEHOLDER_MERCHANTS = ["<Store> (<Suburb>)"] as const;
+
+/**
+ * Below this a text part needs a money figure to be believable. NOT a floor on
+ * its own: `TOTAL $17.00` is twelve characters and is exactly the kind of
+ * pasted line this door exists to accept.
+ */
+export const MIN_RECEIPT_TEXT_CHARS = 40;
+
+/** `$17.00`, `17.00`, `12,50` — a printed amount, not merely a digit. */
+const MONEY_SHAPE = /(?:[$€£¥]\s*\d|\d[.,]\d{2}(?!\d))/;
+
+export type TextCheck = { ok: true } | { ok: false; reason: string };
+
+/**
+ * Can this text possibly be a receipt? Deterministic, before any model call.
+ *
+ * A receipt has more than one word and at least one number. Everything the
+ * phone substitutes when a share goes wrong — the literal word `Image`, a bare
+ * filename like `IMG_0042.jpeg`, an empty body — fails one of those, and none
+ * of them fails in a way a model will admit to.
+ *
+ * Applies to TEXT parts only. A PDF is checked for readable text separately and
+ * an image is the thing we actually asked for.
+ */
+export function textCouldBeReceipt(raw: string | null | undefined): TextCheck {
+  const text = (raw ?? "").trim();
+  if (!text) return { ok: false, reason: "it was empty" };
+  if (!/\s/.test(text)) return { ok: false, reason: "it is a single word" };
+  if (!/\d/.test(text)) return { ok: false, reason: "it has no numbers in it" };
+  if (text.length < MIN_RECEIPT_TEXT_CHARS && !MONEY_SHAPE.test(text)) {
+    return { ok: false, reason: "it is too short to be a receipt" };
+  }
+  return { ok: true };
+}
+
+/** What the phone actually sent, short enough to put on a card. */
+export function quoteShared(raw: string, max = 80): string {
+  const flat = (raw ?? "").replace(/\s+/g, " ").trim();
+  return flat.length > max ? `${flat.slice(0, max - 1)}…` : flat;
+}
+
+/**
+ * Belt to `textCouldBeReceipt`'s braces: did the extractor answer out of its
+ * prompt rather than out of the upload?
+ *
+ * Two tells, both cheap and both deterministic:
+ *  - the merchant IS one of the prompt's format placeholders; or
+ *  - nothing readable was sent at all (no images, and no text part that could
+ *    be a receipt) and yet a merchant or items came back.
+ *
+ * The second is reachable even with the pre-check in place, because the
+ * pre-check only inspects text parts: a PDF that extracts to a few characters
+ * of junk still reaches the model. This is `unsupported`, never `needs_header`
+ * — a confabulated header is not a partial one.
+ */
+export function confabulationReason(
+  input: { textParts: string[]; hasImages: boolean },
+  extracted: { merchant: string | null; items: unknown[] }
+): string | null {
+  const key = merchantKey(extracted.merchant);
+  if (key && PROMPT_PLACEHOLDER_MERCHANTS.some((p) => merchantKey(p) === key)) {
+    return "the merchant came back as the prompt's own example rather than a name off the receipt";
+  }
+  if (input.hasImages) return null;
+  if (input.textParts.some((t) => textCouldBeReceipt(t).ok)) return null;
+  if (!extracted.merchant && !extracted.items.length) return null;
+  return "nothing shared could be read as a receipt, yet an order came back";
+}

src/lib/order-parse.ts

@@ -526,7 +526,11 @@ export function parseOrderHTML(html: string, meta: MessageMeta): ParsedOrder {
   } else {
     // DoorDash carries no order id anywhere in the receipt, so the provider
     // message id is the only stable identity available. That is correct for
-    // ingestion idempotency: one receipt is one order.
+    // ingestion idempotency: one receipt is one order — PROVIDED every caller
+    // sends the same id for the same email. Two lanes post here (n8n Order
+    // Processor and ingestion-engine order_handoff); both must send the bare
+    // RFC Message-ID. n8n sent Gmail's own id until 2026-10-03 and one order
+    // became txns 7268 + 7269.
     if (!meta.messageId) {
       throw new OrderParseError("no order id in body and no messageId supplied");
     }

src/lib/order-reviews.ts

@@ -167,9 +167,24 @@ export async function merchantVerdict(
        FROM order_reviews r
        JOIN transactions t ON t.id = r.transaction_id
        JOIN participants p ON p.id = r.participant_id
-       JOIN expense_metadata em
-         ON em.transaction_id = r.transaction_id
-         OR em.matched_transaction_id = r.transaction_id
+       -- (No backticks in here: this SQL lives in a TS template literal and a
+       -- backtick ends the string -- TS1005 on a line that looks like a comment.)
+       -- LATERAL, not a plain JOIN. A plain join on
+       -- transaction_id OR matched_transaction_id fans out: one review
+       -- joined to N metadata rows appears N times in the result, and the loops
+       -- below count every appearance -- inflating the loved/never tallies
+       -- that drive the pre-order warning. LIMIT 50 applies to the fanned-out
+       -- set too, so duplicates would also push older reviews out of the
+       -- window. Same tie-break as every other reader, so the merchant matched
+       -- here is the merchant the UI shows.
+       JOIN LATERAL (
+         SELECT em2.merchant_normalized
+           FROM expense_metadata em2
+          WHERE em2.transaction_id = r.transaction_id
+             OR em2.matched_transaction_id = r.transaction_id
+          ORDER BY (em2.transaction_id = r.transaction_id) DESC NULLS LAST, em2.id
+          LIMIT 1
+       ) em ON true
       -- Case-folded: the platforms capitalise the same restaurant differently
       -- ("TEG Kebabs & Biryani" on Uber Eats, "TEG KEBABS & BIRYANI" on
       -- DoorDash). An exact match split one restaurant's history in two, so a

src/lib/queries.ts

@@ -851,6 +851,15 @@ export async function getPendingReconciliations(ownerId: number): Promise<Manual
        WHERE ts.transaction_id = t.id
      ) txn_splits ON true
      WHERE t.statement_id IS NULL AND t.owner_id = $1 AND t.reconciled_with_id IS NULL
+       -- A superseded row is already resolved, just not by reconciling. Without
+       -- this the queue kept offering it: a duplicate wallet half hidden under
+       -- its twin (txn 7037, 2026-10-03) sat beside that twin with the same
+       -- statement line as its match, and accepting it on the hidden row would
+       -- have used the line up and stranded the real one.
+       AND t.superseded_by_id IS NULL
+       -- A voided row is resolved too: the charge never posted, so no statement
+       -- line will ever arrive for it (migration 0040).
+       AND t.voided_at IS NULL
        AND ${needsCardMatch("t")}
      ORDER BY t.transaction_date DESC, t.row_index ASC`,
     [ownerId]
@@ -887,12 +896,39 @@ export async function getPendingReconciliations(ownerId: number): Promise<Manual
        FROM transactions m
        JOIN transactions t ON t.statement_id IS NOT NULL
          AND t.transaction_date BETWEEN m.transaction_date - 3 AND m.transaction_date + 3
-         AND t.amount BETWEEN m.amount * 0.99 AND m.amount * 1.01
+         -- +-5%, NOT +-1%. A foreign card purchase is charged at the notified
+         -- amount PLUS the card's foreign transaction fee (3% on the Westpac
+         -- Altitude), so the wallet row and its statement line can never match
+         -- exactly and the old +-1% band excluded the statement row outright:
+         -- SP PEAKDESIGN was notified at $42.44 and charged $43.71 (x1.03),
+         -- so the one row that WAS its twin was the one the picker could not
+         -- offer. Widening is safe here because this list is capped at 5 by
+         -- the rn window and ordered by amount closeness; a looser band changes
+         -- which five appear, never how many, and a person confirms each one.
+         -- The AUTO-matcher stays exact on purpose; see findAutoMatches().
+         AND (t.amount BETWEEN m.amount * 0.95 AND m.amount * 1.05
+              -- Or the exact native pair, whatever the AUD figures say. A
+              -- wallet row's AUD is provisional: in Oct 2026 it priced INR at
+              -- 0.0180 while CommBank charged ~0.0149, 20% under, so the band
+              -- excluded every rupee twin. findAutoMatches() matches on this
+              -- pair already, but refuses when two lines share it inside its
+              -- window (INR 112 on 10/4 and again on 10/7) -- and those are
+              -- exactly the rows that land here for a person to choose.
+              OR (m.foreign_currency_code IS NOT NULL
+                  AND t.foreign_currency_code = m.foreign_currency_code
+                  AND t.foreign_currency_amount = m.foreign_currency_amount))
        JOIN statements s ON s.id = t.statement_id
        LEFT JOIN transaction_overrides o ON o.transaction_id = t.id
        WHERE m.statement_id IS NULL
          AND m.owner_id = $1
          AND m.reconciled_with_id IS NULL
+         AND m.superseded_by_id IS NULL
+         AND m.voided_at IS NULL
+         -- Nor offer a superseded STATEMENT row as the match: the ANZ
+         -- statements 107/142/143 left 31 of those, and a manual row
+         -- reconciled to one inherits its exclusion from every figure.
+         -- findAutoMatches() has had both filters all along.
+         AND t.superseded_by_id IS NULL
          AND ${needsCardMatch("m")}
          AND COALESCE(t.owner_id, s.owner_id) = $1
          AND NOT EXISTS (
@@ -990,6 +1026,11 @@ export async function getSharedTransactions(ownerId: number, tagIds?: number[],
       SELECT em.platform
         FROM expense_metadata em
        WHERE em.transaction_id = t.id OR em.matched_transaction_id = t.id
+       -- Same tie-break as getTransactions / replaceLineItems / matchTransaction,
+       -- so every reader resolves the row a write lands on. Without the ORDER BY
+       -- this LIMIT 1 is bounded but arbitrary: it would disagree with the main
+       -- transaction list for the same transaction the day multiplicity appears.
+       ORDER BY (em.transaction_id = t.id) DESC NULLS LAST, em.id
        LIMIT 1
     ) order_ctx ON true
     LEFT JOIN transactions src ON src.reconciled_with_id = t.id AND src.statement_id IS NULL

src/lib/receipt-ingestion.ts

@@ -1,5 +1,6 @@
 import { queryRaw, queryRow } from "@/lib/db";
 import { DEFAULT_OWNER_ID } from "@/lib/order-ingestion";
+import { findAdoptableTransaction } from "@/lib/auto-reconcile";
 
 /**
  * Grocery receipts scanned in pantry-app, arriving as candidate spend.
@@ -154,16 +155,75 @@ export async function processReceiptIngestion(receipt: ParsedReceipt): Promise<R
     const paymentMethod = paymentMethodFor(leg.class);
     const amount = round2(Number(leg.amount));
     const description = receipt.store_detail ? `${receipt.merchant_name} ${receipt.store_detail}` : receipt.merchant_name;
-    const txn = await queryRow<{ id: number }>(
-      `INSERT INTO transactions (
-         transaction_date, description, amount, amount_aud, category, payment_method,
-         merchant_name, merchant_normalized, transaction_type, owner_id
-       ) VALUES ($1,$2,$3,$3,$4,$5,$6,$6,'debit',$7) RETURNING id`,
-      [receipt.transaction_date, description, amount, RECEIPT_CATEGORY, paymentMethod, receipt.merchant_name, DEFAULT_OWNER_ID]
-    );
+
+    // Adopt the row the wallet lane already created for this tap, if there is
+    // one. Signals for a purchase arrive in sequence -- the notification within
+    // seconds, the receipt hours later -- so creating a second transaction here
+    // is how one shop ends up as two rows. Measured on txn 6858: the wallet
+    // event arrived 35h25m before this path created its own row for the same
+    // $103.30 Woolworths shop. Null means no unambiguous candidate, and the
+    // fallback is exactly the old behaviour.
+    const adoptedId = await findAdoptableTransaction(DEFAULT_OWNER_ID, amount, receipt.transaction_date);
+    const txn = adoptedId
+      ? await queryRow<{ id: number }>(
+          // The receipt read the register tape; the notification read a push
+          // banner. So the receipt wins on what was bought and where, and the
+          // column list here IS the ownership boundary -- `location` is absent
+          // because the wallet notification's suburb is the better value and
+          // must survive.
+          `UPDATE transactions
+              SET description = $2, merchant_name = $3, merchant_normalized = $3,
+                  category = $4, payment_method = $5
+            WHERE id = $1 RETURNING id`,
+          [adoptedId, description, receipt.merchant_name, RECEIPT_CATEGORY, paymentMethod]
+        )
+      : await queryRow<{ id: number }>(
+          `INSERT INTO transactions (
+             transaction_date, description, amount, amount_aud, category, payment_method,
+             merchant_name, merchant_normalized, transaction_type, owner_id
+           ) VALUES ($1,$2,$3,$3,$4,$5,$6,$6,'debit',$7) RETURNING id`,
+          [receipt.transaction_date, description, amount, RECEIPT_CATEGORY, paymentMethod, receipt.merchant_name, DEFAULT_OWNER_ID]
+        );
 
     const carriesLines = leg.leg_index === cardLeg?.leg_index;
-    const meta = await queryRow<{ id: number }>(
+
+    // An adopted row may already carry a metadata row (the wallet lane writes
+    // one). expense_metadata.transaction_id is UNIQUE, so a second INSERT would
+    // raise -- and must not, because the two describe the same purchase.
+    // Enrich in place instead, with the same ownership split as the transaction
+    // update above: everything the receipt read wins, and anything only the
+    // notification knew (issuer, and the transaction's location) is absent from
+    // the column list rather than conditionally skipped, so it cannot be
+    // clobbered by a bug here. Mirrors order_transaction_bridge's UPGRADE pass.
+    const priorMeta = adoptedId
+      ? await queryRow<{ id: number }>(
+          `SELECT id FROM expense_metadata WHERE transaction_id = $1`, [adoptedId]
+        )
+      : null;
+
+    const meta = priorMeta
+      ? await queryRow<{ id: number }>(
+          `UPDATE expense_metadata
+              SET source = 'pantry', order_reference = $2, line_items = $3::jsonb,
+                  subtotal = $4, amount = $5, merchant_normalized = $6,
+                  transaction_date = $7, card_last4 = COALESCE($8, card_last4),
+                  currency = 'AUD', flags = flags || $9::jsonb,
+                  payment_method = $10, payment_method_detail = $11,
+                  tender_raw = COALESCE($12, tender_raw),
+                  receipt_sha256 = $13, receipt_group = $14,
+                  extraction_model = 'cloud-budget'
+            WHERE id = $1 RETURNING id`,
+          [
+            priorMeta.id, reference,
+            JSON.stringify(carriesLines ? receipt.line_items ?? [] : []),
+            receipt.tax_amount ?? null, amount, receipt.merchant_name,
+            receipt.transaction_date, leg.card_last4 ?? null,
+            JSON.stringify(carriesLines ? [...flags, "enriched_prior_row"] : [...flags, "line_items_on_card_leg", "enriched_prior_row"]),
+            paymentMethod, leg.card_product ?? null, receipt.tender_raw ?? null,
+            receipt.image_sha256 ?? null, group,
+          ]
+        )
+      : await queryRow<{ id: number }>(
       `INSERT INTO expense_metadata (
          transaction_id, source, order_reference, line_items, subtotal, amount,
          merchant_normalized, transaction_date, card_last4, currency, flags,

src/lib/reconcile.ts

@@ -0,0 +1,248 @@
+import { prisma } from "@/lib/db";
+
+/**
+ * Merge a statementless ("manual") transaction into its statement twin.
+ *
+ * Extracted from POST /api/transactions/reconcile so the manual queue and the
+ * automatic pass cannot drift apart. There must be exactly one implementation
+ * of this: it moves curation (overrides, tags, splits) and provenance
+ * (expense_metadata) off a row that reconciliation is about to hide from every
+ * figure, and a second copy of that logic would eventually move a different set.
+ *
+ * The manual row is NOT deleted -- it is linked via `reconciled_with_id`.
+ * Deleting it would cascade its children away; superseding/linking keeps the
+ * history and lets the row stay visible as provenance.
+ */
+export async function applyReconcileMatch(
+  manual_id: number,
+  statement_tx_id: number
+): Promise<void> {
+  await prisma.$transaction(async (tx) => {
+    await moveCurationWithin(tx, manual_id, statement_tx_id);
+
+    // Mark manual tx as reconciled (link to statement tx)
+    await tx.$executeRawUnsafe(
+      `UPDATE transactions SET reconciled_with_id = $1 WHERE id = $2`,
+      statement_tx_id,
+      manual_id
+    );
+  });
+}
+
+/**
+ * Move overrides, tags, splits and expense_metadata from one transaction to
+ * another, inside a caller-supplied transaction.
+ *
+ * Callers: `applyReconcileMatch` (then sets `reconciled_with_id`) and
+ * `supersedeTransaction` (then sets `superseded_by_id`). Both hide the source
+ * row from every figure, so both must carry its curation across first; keeping
+ * one implementation is what the file header insists on.
+ */
+async function moveCurationWithin(
+  tx: Parameters<Parameters<typeof prisma.$transaction>[0]>[0],
+  manual_id: number,
+  statement_tx_id: number
+): Promise<void> {
+  {
+    // Copy overrides: manual → statement tx
+    const override = await tx.transaction_overrides.findUnique({
+      where: { transaction_id: manual_id },
+    });
+    if (override) {
+      await tx.transaction_overrides.upsert({
+        where: { transaction_id: statement_tx_id },
+        update: {
+          category_override: override.category_override,
+          merchant_normalized: override.merchant_normalized,
+          notes: override.notes,
+          my_share_percent: override.my_share_percent,
+          updated_at: new Date(),
+        },
+        create: {
+          transaction_id: statement_tx_id,
+          category_override: override.category_override,
+          merchant_normalized: override.merchant_normalized,
+          notes: override.notes,
+          my_share_percent: override.my_share_percent,
+        },
+      });
+      await tx.transaction_overrides.deleteMany({ where: { transaction_id: manual_id } });
+    }
+
+    // Move tags: manual → statement tx
+    const tags = await tx.transaction_tags.findMany({ where: { transaction_id: manual_id } });
+    if (tags.length) {
+      await tx.transaction_tags.createMany({
+        data: tags.map((t) => ({ transaction_id: statement_tx_id, tag_id: t.tag_id })),
+        skipDuplicates: true,
+      });
+      await tx.transaction_tags.deleteMany({ where: { transaction_id: manual_id } });
+    }
+
+    // Move splits: manual → statement tx.
+    //
+    // `settled` and `settled_at` MUST be carried. This is a delete-and-recreate,
+    // and the column defaults to false — so omitting it silently marks every
+    // settled split unsettled again and resurrects a debt that was already
+    // paid. Nothing errors and nothing in the UI says the flag was reset; it
+    // just reappears as owed. No queue row carries a settled split today, but
+    // the auto-matcher runs this path unattended, so a latent version of this
+    // would surface as money, not as an exception.
+    const splits = await tx.transaction_splits.findMany({ where: { transaction_id: manual_id } });
+    if (splits.length) {
+      await tx.transaction_splits.createMany({
+        data: splits.map((s) => ({
+          transaction_id: statement_tx_id,
+          participant_id: s.participant_id,
+          share_percent: s.share_percent,
+          settled: s.settled ?? false,
+          settled_at: s.settled_at,
+        })),
+        skipDuplicates: true,
+      });
+      await tx.transaction_splits.deleteMany({ where: { transaction_id: manual_id } });
+    }
+
+    // Move provenance: manual → statement tx.
+    //
+    // Overrides, tags and splits above were always carried across; expense_metadata was
+    // the one child left behind, which did not matter while every metadata row came from
+    // an email that had created its own transaction. It matters now: a scanned grocery
+    // receipt puts its line items here, and reconciliation hides the manual row from
+    // every figure — so without this the shop's contents disappear at exactly the moment
+    // the statement line appears, and COLES 0556 MANOR LAKES stays as unreadable as it
+    // was before the receipt was ever scanned.
+    //
+    // transaction_id is UNIQUE, so a statement row that already has metadata (an emailed
+    // or Paperless copy got there first) keeps it. The manual row's metadata stays
+    // attached to the reconciled manual transaction and is flagged, rather than raising a
+    // constraint violation or silently overwriting the other source.
+    const moved = await tx.$executeRawUnsafe(
+      `UPDATE expense_metadata SET transaction_id = $1
+        WHERE transaction_id = $2
+          AND NOT EXISTS (SELECT 1 FROM expense_metadata other WHERE other.transaction_id = $1)`,
+      statement_tx_id,
+      manual_id
+    );
+    if (moved === 0) {
+      await tx.$executeRawUnsafe(
+        `UPDATE expense_metadata
+            SET flags = coalesce(flags, '[]'::jsonb) || '["metadata_collision_on_reconcile"]'::jsonb
+          WHERE transaction_id = $1`,
+        manual_id
+      );
+    }
+  }
+}
+
+/**
+ * Manually mark `loser_id` as superseded by `survivor_id`.
+ *
+ * The escape hatch for a duplicate that `applyReconcileMatch` cannot express:
+ * that one merges a statementLESS row into its statement twin, so it has
+ * nothing to say about two rows on the SAME footing -- the pair left by a
+ * statement imported twice (migration 0023 superseded 31 such pairs, $42,040.68,
+ * by hand), or two feed rows for one purchase.
+ *
+ * Curation moves FIRST, using the same block `applyReconcileMatch` runs, because
+ * `superseded_by_id` hides the row from every figure (`analytics-sql.ts`) and
+ * splits or tags left on it would silently stop counting. That is why this lives
+ * beside its sibling and calls into the same movement rather than re-deriving
+ * it: the header of this file warns that a second copy would eventually move a
+ * different set, and a supersede that forgot splits would resurrect a debt in
+ * exactly the way `applyReconcileMatch`'s own comment describes.
+ *
+ * NOT a delete. Every child of `transactions` is ON DELETE CASCADE, so removing
+ * the row would destroy the splits, tags and metadata that make it legible --
+ * and a duplicate is evidence of how the ledger was assembled, which is worth
+ * keeping. `transactions_no_self_supersede` blocks a row pointing at itself; the
+ * survivor is re-checked here anyway so the error is a 400 and not a 500.
+ */
+export async function supersedeTransaction(
+  loser_id: number,
+  survivor_id: number
+): Promise<void> {
+  if (loser_id === survivor_id) throw new Error("A transaction cannot supersede itself");
+  await prisma.$transaction(async (tx) => {
+    // Refuse to bury a row underneath one that is itself already buried, which
+    // would make the curation we just moved unreachable in a second hop.
+    const survivor = await tx.transactions.findUnique({
+      where: { id: survivor_id },
+      select: { superseded_by_id: true },
+    });
+    if (!survivor) throw new Error(`Survivor transaction ${survivor_id} not found`);
+    if (survivor.superseded_by_id !== null) {
+      throw new Error(`Transaction ${survivor_id} is itself superseded; pick the surviving row`);
+    }
+    await moveCurationWithin(tx, loser_id, survivor_id);
+    await tx.$executeRawUnsafe(
+      `UPDATE transactions SET superseded_by_id = $1 WHERE id = $2`,
+      survivor_id,
+      loser_id
+    );
+  });
+}
+
+/**
+ * Mark a statementless row as a charge that never posted (migration 0040).
+ *
+ * The case neither sibling above can express: a card notification is sent for
+ * an AUTHORISATION, and when the merchant releases the hold no statement line
+ * ever arrives. Txn 7951 was a $1,040.41 Airbnb hold on a booking cancelled the
+ * next day -- it sat in the reconcile queue with nothing to match and counted as
+ * spend in every figure. There is no survivor to move curation onto, so nothing
+ * moves: the row and its children stay, and `EXCLUDE_RECONCILED_SOURCE` stops
+ * them counting.
+ *
+ * Refusals, all deliberate:
+ *  - a statement row: that is the bank's record of a posted charge, and the
+ *    statement must keep summing to its header (`transactions_void_statementless`
+ *    enforces this in the database too). A wrong statement row is a supersede.
+ *  - a row already reconciled or superseded: it is already out of every figure,
+ *    and a reconciled one evidently DID post.
+ *  - a row with a SETTLED split: someone has already paid their share of it.
+ *    Voiding would silently turn that payment into a credit on their balance;
+ *    that needs a person to look at it, not a side effect.
+ */
+export async function voidTransaction(id: number, reason: string): Promise<void> {
+  const why = reason.trim();
+  if (!why) throw new Error("A void needs a reason");
+  await prisma.$transaction(async (tx) => {
+    const rows = await tx.$queryRawUnsafe<
+      { statement_id: number | null; reconciled_with_id: number | null; superseded_by_id: number | null; settled: number }[]
+    >(
+      `SELECT t.statement_id, t.reconciled_with_id, t.superseded_by_id,
+              (SELECT count(*)::int FROM transaction_splits ts
+                WHERE ts.transaction_id = t.id AND ts.settled) AS settled
+         FROM transactions t WHERE t.id = $1 FOR UPDATE`,
+      id
+    );
+    const row = rows[0];
+    if (!row) throw new Error(`Transaction ${id} not found`);
+    if (row.statement_id !== null) {
+      throw new Error(`Transaction ${id} is a statement line; a posted charge cannot be voided (supersede it instead)`);
+    }
+    if (row.reconciled_with_id !== null) {
+      throw new Error(`Transaction ${id} is reconciled with ${row.reconciled_with_id}, so it posted`);
+    }
+    if (row.superseded_by_id !== null) {
+      throw new Error(`Transaction ${id} is already superseded by ${row.superseded_by_id}`);
+    }
+    if (row.settled > 0) {
+      throw new Error(`Transaction ${id} has ${row.settled} settled split(s); resolve those before voiding`);
+    }
+    await tx.$executeRawUnsafe(
+      `UPDATE transactions SET voided_at = now(), void_reason = $1 WHERE id = $2 AND voided_at IS NULL`,
+      why,
+      id
+    );
+  });
+}
+
+/** Undo `voidTransaction` -- e.g. the "released" hold posted after all. */
+export async function unvoidTransaction(id: number): Promise<void> {
+  await prisma.$executeRawUnsafe(
+    `UPDATE transactions SET voided_at = NULL, void_reason = NULL WHERE id = $1`,
+    id
+  );
+}

src/lib/wallet-descriptor.ts

@@ -0,0 +1,120 @@
+/**
+ * Deterministic parsing of the bits of a wallet notification that have a fixed
+ * shape.
+ *
+ * Everything here was previously left to the interpreting model, which handled
+ * ONE fixed-width format three different ways on the same day's data: it kept
+ * the country code on "WYNDHAM VAL AU", stripped it on "Sydney" and "Perth",
+ * and swallowed "SAN FRANCISOUS" whole into the location field. A fixed-width
+ * field is not a judgement call, and the same argument that put the order
+ * parsers in code rather than in a Gemini call applies: a regex fails loudly on
+ * a template change, where a model returns a plausible wrong answer.
+ *
+ * The model still decides what the amount and merchant ARE -- it is better at
+ * that than a pattern, because five cards emit three shapes. This only handles
+ * the parts that are positional.
+ */
+
+/**
+ * Westpac writes the merchant/city/country block as a fixed-width statement
+ * descriptor: 27-char merchant, 12-char city, 2-char ISO country. Verified
+ * across every Westpac event captured:
+ *
+ *   Breadworld Bakery Cafe     WYNDHAM VAL AU
+ *   PLA DONATION               MELBOURNE   AU
+ *   SP PEAKDESIGN              SAN FRANCISOUS   <- "SAN FRANCISO" + "US"
+ *
+ * The last example is why this is positional rather than a split on whitespace:
+ * a 12-character city fills its field exactly and runs straight into the
+ * country code with no separator.
+ */
+export type WestpacDescriptor = {
+  merchant: string | null;
+  city: string | null;
+  /** ISO-3166 alpha-2, upper-case. Null when the tail is not a country code. */
+  country: string | null;
+};
+
+const MERCHANT_WIDTH = 27;
+const CITY_WIDTH = 12;
+
+/**
+ * `body` is the whole notification body, e.g.
+ *   "$13.50 paid from Mastercard ending...8032. Breadworld Bakery Cafe     WYNDHAM VAL AU"
+ * Returns nulls rather than throwing: a template change must degrade to "we
+ * don't know", never to a wrong country on a foreign transaction.
+ */
+export function parseWestpacDescriptor(body: string | null | undefined): WestpacDescriptor {
+  const none: WestpacDescriptor = { merchant: null, city: null, country: null };
+  if (!body) return none;
+
+  // The descriptor is whatever follows the first sentence ("... ending...8032. ").
+  // [\s\S] rather than the `s` flag: the tsconfig target predates es2018.
+  const m = body.match(/^[\s\S]*?\.\s([\s\S]*)$/);
+  if (!m) return none;
+  const tail = m[1].replace(/\s+$/, "");
+  if (!tail) return none;
+
+  const merchant = tail.slice(0, MERCHANT_WIDTH).trim() || null;
+  const rest = tail.slice(MERCHANT_WIDTH);
+  if (!rest) return { merchant, city: null, country: null };
+
+  // Country is the final two characters of the fixed block, but only when the
+  // block is actually long enough to contain one -- a short tail is a merchant
+  // name with no location, not a merchant whose last two letters are a country.
+  const block = rest.slice(0, CITY_WIDTH + 2);
+  if (block.length < CITY_WIDTH + 2) {
+    return { merchant, city: block.trim() || null, country: null };
+  }
+  const city = block.slice(0, CITY_WIDTH).trim() || null;
+  const tailTwo = block.slice(CITY_WIDTH).trim().toUpperCase();
+  const country = /^[A-Z]{2}$/.test(tailTwo) ? tailTwo : null;
+  return { merchant, city, country: country };
+}
+
+/**
+ * The Apple Wallet notification's subtitle, which is comma-delimited and clean:
+ *   "Breadworld Bakery Cafe, Wyndham Vale, VIC"  -> location "Wyndham Vale, VIC"
+ *   "Seatsaero"                                  -> no location (online merchant)
+ *
+ * Only the first comma separates merchant from place; the rest is the place.
+ */
+export function parseWalletSubtitle(subtitle: string | null | undefined): {
+  merchant: string | null;
+  location: string | null;
+} {
+  if (!subtitle || !subtitle.trim()) return { merchant: null, location: null };
+  const parts = subtitle.split(",").map((p) => p.trim()).filter(Boolean);
+  if (parts.length <= 1) return { merchant: parts[0] ?? null, location: null };
+  return { merchant: parts[0], location: parts.slice(1).join(", ") };
+}
+
+/**
+ * Currency marker, if the notification states one.
+ *
+ * Only Amex does: it writes "A$5,103.49". Westpac, CommBank and Apple Wallet
+ * write a bare "$", which is honestly recorded as unknown rather than assumed
+ * to be AUD -- 14 of 21 captured transaction rows. Domestically that assumption
+ * would be harmless; abroad it is a 23x error on a THB tap, so it is not made
+ * here. The statement settles it within weeks, and the reconcile queue holds
+ * the row until then.
+ */
+export function currencyFromAmountText(amountText: string | null | undefined): string | null {
+  if (!amountText) return null;
+  const t = amountText.trim();
+  if (/^A\$/i.test(t)) return "AUD";
+  if (/^(NZ\$)/i.test(t)) return "NZD";
+  if (/^US\$/i.test(t)) return "USD";
+  if (/^€/.test(t)) return "EUR";
+  if (/^£/.test(t)) return "GBP";
+  if (/^¥/.test(t)) return "JPY";
+  if (/^฿/.test(t)) return "THB";
+  if (/^S\$/i.test(t)) return "SGD";
+  if (/^₹/.test(t)) return "INR";
+  // CommBank's approval prompts write the ISO code after the figure
+  // ("672006 VND", "1340.00 INR"); some issuers put it first ("USD 10.00").
+  const iso = t.match(/^[\d.,\s]+([A-Z]{3})$/) ?? t.match(/^([A-Z]{3})\s*[\d.,]+$/);
+  if (iso) return iso[1];
+  // A bare "$" is genuinely ambiguous -- do not guess.
+  return null;
+}

src/lib/wallet-ingestion.ts

@@ -0,0 +1,481 @@
+import { queryRow, withTransaction, type Queryable } from "@/lib/db";
+import { DEFAULT_OWNER_ID } from "@/lib/order-ingestion";
+import { findWalletAdoptableTransaction } from "@/lib/auto-reconcile";
+import {
+  parseWestpacDescriptor,
+  parseWalletSubtitle,
+  currencyFromAmountText,
+} from "@/lib/wallet-descriptor";
+
+/**
+ * Wallet notifications -> transactions.
+ *
+ * A transaction is created for EVERY tap (DECISIONS.md ING-11, 2026-08-22),
+ * because overseas trips are where live spend matters most: the statement lag
+ * is the same 6-35 days abroad but knowing what you have spent matters far more
+ * in an unfamiliar currency.
+ *
+ * SEQUENCING IS THE WHOLE DESIGN. Wallet is the FIRST signal, not a late one --
+ * measured on txn 6858, the notification arrived 35h25m before anything created
+ * a transaction. So this creates the row, the receipt enriches it
+ * (receipt-ingestion.ts adopts it), and the statement confirms it via the
+ * reconcile queue, which a wallet row enters unaided: it is statementless with
+ * payment_method 'card'. Nothing here needs to know about statements.
+ *
+ * But that sequence is a TENDENCY, not a guarantee -- these first 21 rows were
+ * themselves backfilled days after the taps -- so this also adopts a row a
+ * receipt got to first (findWalletAdoptableTransaction), and a stored-value
+ * instrument is written as 'credits' so it never enters a queue no statement
+ * line will ever resolve.
+ */
+
+export type WalletEvent = {
+  /** capture.wallet_events.id -- the idempotency key. */
+  event_id: number;
+  /** 'transaction' | 'not_transaction' | 'unsure' from the interpreter. */
+  kind: string;
+  direction?: string | null;
+  amount: number | null;
+  /** The exact substring the model read the amount from, e.g. "A$5,103.49". */
+  amount_text?: string | null;
+  currency?: string | null;
+  merchant?: string | null;
+  card_tail?: string | null;
+  issuer?: string | null;
+  location?: string | null;
+  notif_app?: string | null;
+  /** The notification's title line, e.g. "Visa Gift Card" or "Payment". */
+  notif_title?: string | null;
+  notif_body?: string | null;
+  notif_subtitle?: string | null;
+  occurred_at?: string | null;
+};
+
+export type WalletIngestResult = {
+  transactionId: number | null;
+  metadataId: number | null;
+  skipped?:
+    | "not_a_transaction" | "already_ingested" | "no_amount" | "merged_pair" | "adopted"
+    | "account_notice" | "stored_value_drawdown";
+  flags: string[];
+};
+
+/**
+ * How far apart the two halves of ONE tap can arrive. Measured 2-8 seconds on
+ * every captured pair; ten minutes leaves room for a phone that was offline,
+ * and is still nowhere near the +/-1 DAY the matcher used to allow -- which
+ * folded the UNICEF $50 donation and an unrelated $50 bar charge 17 hours later
+ * into one row (event 275 into txn 8091, 2026-10-01).
+ */
+const PAIR_WINDOW_MINUTES = 10;
+
+/** The ledger is dated the way the bank statements are: Melbourne local. */
+const LEDGER_TZ = "Australia/Melbourne";
+
+/**
+ * The calendar date of an instant, in LEDGER_TZ.
+ *
+ * Slicing the ISO string took whatever offset n8n serialised -- UTC in
+ * practice -- so a Melbourne-evening tap landed on the previous day (28 of the
+ * first 134 rows). Statements, which this row is reconciled against, date in
+ * Australian local time.
+ */
+export function ledgerDate(occurredAt: string | null | undefined, now: Date = new Date()): string {
+  const d = occurredAt ? new Date(occurredAt) : now;
+  const at = Number.isNaN(d.getTime()) ? now : d;
+  // en-CA formats as YYYY-MM-DD.
+  return new Intl.DateTimeFormat("en-CA", {
+    timeZone: LEDGER_TZ, year: "numeric", month: "2-digit", day: "2-digit",
+  }).format(at);
+}
+
+/**
+ * A ledger-derived AUD rate for a currency: the median of what the bank
+ * actually charged divided by the native figure, over statement rows.
+ *
+ * Same source order_transaction_bridge uses for its plausibility guard. This is
+ * only ever a PROVISIONAL value -- transactions.amount is NOT NULL and always
+ * AUD, and a notification cannot know the AUD figure because the bank sets the
+ * rate at settlement. The statement overwrites it at reconcile time.
+ * Returns null when there is not enough history to be worth anything.
+ */
+async function medianRateToAud(code: string): Promise<number | null> {
+  const row = await queryRow<{ rate: string | null }>(
+    `SELECT percentile_cont(0.5) WITHIN GROUP (
+              ORDER BY amount / NULLIF(foreign_currency_amount, 0)
+            )::text AS rate
+       FROM transactions
+      WHERE foreign_currency_code = $1
+        AND foreign_currency_amount > 0
+        AND superseded_by_id IS NULL
+        -- Statement rows only: a wallet row's amount is itself a provisional
+        -- guess, and one stored before this filter (txn 7316, 672006 VND at a
+        -- 1:1 "rate") would otherwise feed the next estimate.
+        AND statement_id IS NOT NULL
+      HAVING count(*) >= 5`,
+    [code]
+  );
+  const rate = row?.rate ? Number(row.rate) : null;
+  return rate && Number.isFinite(rate) && rate > 0 ? rate : null;
+}
+
+/**
+ * The ECB reference rate for the day, via Frankfurter (free, no key; the same
+ * source the statement processor converts with).
+ *
+ * The fallback when the ledger has no history for a currency. Hong Kong was the
+ * first trip to need it: no HKD statement row existed, so every HKD tap stored
+ * A$0.00 and the trip read as free until the statement landed weeks later. The
+ * ledger median is still preferred -- it carries the card's ~3% foreign fee,
+ * which the ECB mid-rate does not -- so this one is flagged as such.
+ *
+ * Any failure is null, never a throw: no rate is a state this lane already
+ * handles (a flagged 0), and a slow FX API must not lose the purchase.
+ */
+async function ecbRateToAud(code: string, date: string): Promise<number | null> {
+  try {
+    const res = await fetch(
+      `https://api.frankfurter.dev/v1/${date}?from=${encodeURIComponent(code)}&to=AUD`,
+      { signal: AbortSignal.timeout(5000) }
+    );
+    if (!res.ok) return null;
+    const body = (await res.json()) as { rates?: { AUD?: number } };
+    const rate = body?.rates?.AUD;
+    return typeof rate === "number" && Number.isFinite(rate) && rate > 0 ? rate : null;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * The other half of a paired notification, if it has already been ingested.
+ *
+ * One tap emits TWO notifications ~2-8s apart: the issuer app (which has the
+ * card tail) and Apple Wallet (which has the better merchant and the suburb).
+ * Neither alone is complete, so they must be MERGED, never deduped -- dropping
+ * either loses a field nothing else supplies. The notifications share no
+ * identifier, so a pair is recognised by what a pair IS: the same amount,
+ * minutes apart, from two different apps. Two notifications from one app are
+ * two charges, however close together.
+ */
+async function findPairedRow(
+  db: Queryable,
+  amount: number,
+  notifiedAt: string,
+  notifApp: string | null,
+  fx: { amount: number; code: string } | null = null
+) {
+  // A foreign row pairs on its native pair: its `amount` is a provisional
+  // conversion (or a 0 placeholder), so two unrelated foreign taps could share it.
+  const rows = await db.queryRaw<{ id: number; meta_id: number }>(
+    `SELECT t.id, em.id AS meta_id
+       FROM transactions t
+       JOIN expense_metadata em ON em.transaction_id = t.id
+      WHERE t.source = 'wallet'
+        AND t.statement_id IS NULL
+        AND t.superseded_by_id IS NULL
+        AND CASE WHEN $4::text IS NULL
+                 THEN t.amount = $1::numeric AND t.foreign_currency_code IS NULL
+                 ELSE t.foreign_currency_code = $4::text
+                      AND t.foreign_currency_amount = $5::numeric END
+        AND em.notified_at BETWEEN $2::timestamptz - make_interval(mins => $3::int)
+                               AND $2::timestamptz + make_interval(mins => $3::int)
+        AND em.notif_app IS DISTINCT FROM $6::text
+      LIMIT 2`,
+    [amount, notifiedAt, PAIR_WINDOW_MINUTES, fx?.code ?? null, fx?.amount ?? null, notifApp]
+  );
+  return rows.length === 1 ? rows[0] : null;
+}
+
+/**
+ * A stored-value instrument: a gift card, a prepaid travel card, a store card.
+ *
+ * It matters because payment_method decides queue membership. 'card' means "a
+ * statement line is coming"; findAutoMatches and the reconcile queue both
+ * exclude 'credits' and 'cash' for exactly that reason. A $103.30 Woolworths
+ * tap on a Visa gift card written as 'card' parks in the queue forever, being
+ * offered as an unmatched row against statements that can never contain it --
+ * three of the first 21 wallet rows (events 80, 92, 101) were gift-card taps.
+ *
+ * The evidence is the notification TITLE, which is where Apple Wallet prints
+ * the instrument's own name ("Visa Gift Card"); the interpreter reduces that to
+ * issuer "Visa", which a plain Visa debit card also produces, so the title is
+ * the only place the distinction survives. Read as a prior, not a contract: a
+ * miss leaves today's behaviour, which is the queue.
+ */
+const STORED_VALUE = /gift\s*card|prepaid|travel\s*(?:money|card)|store\s*card/i;
+
+/**
+ * A transit card reporting a drawdown of its own balance ("HKD 7.40 for Train"
+ * / "Your new balance is HKD 169.50.").
+ *
+ * Unlike a gift card, a transit card is loaded FROM a bank card -- that load is
+ * on the statement and is the spend. Each fare is that money being used, so
+ * booking it again counts it twice. The interpreter already called 11 of 13
+ * Octopus fares not_transaction; the two it did not (events 191, 200) became
+ * rows, so the outcome depended on the model's mood. This makes it a rule. The
+ * body wording is the signal because it is what every such card says; the
+ * issuer name is not (Octopus, Opal, myki, Suica each name themselves).
+ */
+const BALANCE_DRAWDOWN = /\byour (?:new|current) balance is\b/i;
+
+/**
+ * Account notices that state an amount without money moving: "Your Smart Access
+ * 1910 is overdrawn by $21.21. To avoid fees and interest, ..." (event 235)
+ * became a $21.21 purchase named "To avoid fees and interest,". The model is
+ * right that the text names an amount; it is not a payment.
+ */
+const ACCOUNT_NOTICE = /\boverdrawn\b|\bto avoid fees\b|\bavailable balance\b|\blow balance\b/i;
+
+/** How an adopted or merged row records which wallet event folded into it. */
+const adoptedFlag = (eventId: number) => `wallet_evt:${eventId}`;
+
+export function isStoredValueInstrument(ev: Pick<WalletEvent, "notif_title" | "issuer" | "notif_app">): boolean {
+  return [ev.notif_title, ev.issuer, ev.notif_app].some((v) => !!v && STORED_VALUE.test(v));
+}
+
+/** Already landed: a row this lane wrote, or one it adopted or merged into. */
+async function alreadyIngested(db: Pick<Queryable, "queryRow">, eventId: number) {
+  return db.queryRow<{ id: number }>(
+    `SELECT id FROM transactions WHERE source = 'wallet' AND source_ref = $1
+      UNION ALL
+     SELECT em.transaction_id AS id FROM expense_metadata em
+      JOIN transactions t ON t.id = em.transaction_id
+      WHERE em.flags @> $2::jsonb AND t.superseded_by_id IS NULL
+      LIMIT 1`,
+    [`evt:${eventId}`, JSON.stringify([adoptedFlag(eventId)])]
+  );
+}
+
+export async function processWalletIngestion(ev: WalletEvent): Promise<WalletIngestResult> {
+  const flags: string[] = [];
+
+  // An open-world classifier feeding a closed-world writer drops the unknown
+  // and calls it a refusal (ING-14). `unsure` is a real answer and must not
+  // silently become a transaction.
+  if (ev.kind !== "transaction") {
+    return { transactionId: null, metadataId: null, skipped: "not_a_transaction", flags };
+  }
+  if (ev.amount === null || !Number.isFinite(ev.amount) || ev.amount <= 0) {
+    return { transactionId: null, metadataId: null, skipped: "no_amount", flags };
+  }
+  const text = [ev.notif_subtitle, ev.notif_body].filter(Boolean).join("\n");
+  if (ACCOUNT_NOTICE.test(text)) {
+    return { transactionId: null, metadataId: null, skipped: "account_notice", flags };
+  }
+  if (BALANCE_DRAWDOWN.test(ev.notif_body ?? "")) {
+    return { transactionId: null, metadataId: null, skipped: "stored_value_drawdown", flags };
+  }
+
+  // Two places this event may already have landed: a row this lane wrote (which
+  // carries the ref natively), or a row it ADOPTED or MERGED into, whose
+  // source_ref belongs to whoever created it -- marked in the metadata flags
+  // rather than by overwriting a field another lane owns. Both must be checked,
+  // because bumping the interpreter's prompt_version deliberately re-interprets
+  // and re-posts every event ever captured -- that is the replay mechanism, and
+  // it must not mint a second row for a purchase that has one. Checked here
+  // without the lock to skip the FX lookup on a replay, and again under it.
+  const seen = await alreadyIngested({ queryRow }, ev.event_id);
+  if (seen) return { transactionId: seen.id, metadataId: null, skipped: "already_ingested", flags };
+
+  const sourceRef = `evt:${ev.event_id}`;
+  const notifiedAt = ev.occurred_at ?? new Date().toISOString();
+  const date = ledgerDate(ev.occurred_at);
+
+  // Deterministic parsing of the positional fields. The model decides what the
+  // amount and merchant ARE; it does not get to parse a fixed-width block.
+  // Only Westpac writes that block: run over a CommBank overdraft notice it
+  // produced the merchant "To avoid fees and interest," (event 235).
+  const desc = /westpac/i.test(ev.notif_app ?? "")
+    ? parseWestpacDescriptor(ev.notif_body)
+    : { merchant: null, city: null, country: null };
+  const sub = parseWalletSubtitle(ev.notif_subtitle);
+  const merchant = sub.merchant ?? ev.merchant ?? desc.merchant ?? "Card purchase";
+  // Two sources for a place, and they are not equal. The Apple Wallet subtitle
+  // gives "Wyndham Vale, VIC"; the issuer's fixed-width descriptor gives
+  // "WYNDHAM VAL AU" -- a 12-character truncation in caps. The subtitle wins
+  // outright, which matters for the pair merge below: whichever half arrives
+  // second must be able to REPLACE a descriptor-derived place, or the result
+  // depends on arrival order.
+  const subtitleLocation = sub.location ?? null;
+  // Precedence: the Wallet subtitle, then the DETERMINISTIC descriptor parse,
+  // and only then the model's own answer. Getting this order wrong defeats the
+  // point of parsing positionally at all -- the model returned the mangled
+  // "SAN FRANCISOUS" and the padded "MELBOURNE   AU" for exactly these rows,
+  // which is what wallet-descriptor.ts exists to correct.
+  const descriptorLocation = desc.city && desc.country
+    ? `${desc.city} ${desc.country}`
+    : desc.city;
+  const location = subtitleLocation ?? descriptorLocation ?? ev.location ?? null;
+
+  // Currency: only ever what the notification actually states. A bare "$" is
+  // recorded as unknown, never assumed AUD -- harmless domestically, a 23x
+  // error on a THB tap abroad.
+  const stated = currencyFromAmountText(ev.amount_text)
+    ?? (ev.currency && /^[A-Z]{3}$/.test(ev.currency) ? ev.currency : null);
+  if (!stated) flags.push("currency_unverified");
+  if (desc.country && desc.country !== "AU") flags.push(`foreign_country_${desc.country}`);
+
+  // Foreign: the native figure is what both sides can agree on exactly, so it
+  // is what gets stored. `amount` still needs a value (NOT NULL) and takes a
+  // provisional conversion, flagged, which the statement overwrites. Resolved
+  // BEFORE the lock: the ECB fallback is a network call, and the lock
+  // serialises every wallet post behind whatever runs inside it.
+  let amountAud = ev.amount;
+  let fxAmount: number | null = null;
+  let fxCode: string | null = null;
+  if (stated && stated !== "AUD") {
+    fxAmount = ev.amount;
+    fxCode = stated;
+    const ledgerRate = await medianRateToAud(stated);
+    const rate = ledgerRate ?? (await ecbRateToAud(stated, date));
+    if (rate) {
+      amountAud = Number((ev.amount * rate).toFixed(2));
+      flags.push("aud_provisional");
+      if (!ledgerRate) flags.push("rate_ecb");
+    } else {
+      // No rate from either source. The native figure must NOT fall through as
+      // AUD -- 672006 VND once landed as A$672,006. Zero is a flagged
+      // placeholder; the statement supplies the real AUD at reconcile time,
+      // which matches on the native pair, not on this column.
+      amountAud = 0;
+      flags.push("aud_provisional", "no_rate_reference");
+    }
+  }
+
+  // A gift card has no statement line coming, so it must not enter the
+  // reconcile queue (which filters on exactly this column).
+  const storedValue = isStoredValueInstrument(ev);
+  const paymentMethod = storedValue ? "credits" : "card";
+  if (storedValue) flags.push("stored_value_instrument");
+
+  // Everything from the idempotency re-check to the INSERT holds one lock.
+  // n8n's HTTP node posts a poll's items CONCURRENTLY, and both halves of a tap
+  // usually land in the same 15-minute poll -- so each ran findPairedRow before
+  // the other had inserted, neither found its partner, and one tap became two
+  // rows: 11 pairs, A$3,841.40, every pair with an identical created_at. A
+  // single lock for the whole lane, not one per amount: wallet volume is a few
+  // dozen a day, and a coarse lock has no key to get wrong.
+  return withTransaction(async (db) => {
+    await db.queryRaw(`SELECT pg_advisory_xact_lock(hashtextextended('wallet-ingest', 0))::text`);
+
+    const existing = await alreadyIngested(db, ev.event_id);
+    if (existing) {
+      return { transactionId: existing.id, metadataId: null, skipped: "already_ingested" as const, flags };
+    }
+
+    // Merge the paired notification rather than creating a second row for one
+    // tap. The issuer half brings the card tail; the Wallet half brings the
+    // merchant and suburb. COALESCE so whichever arrives second only ever fills
+    // gaps -- it can never blank a field the first half supplied.
+    const pair = await findPairedRow(
+      db, amountAud, notifiedAt, ev.notif_app ?? null,
+      fxCode && fxAmount !== null ? { amount: fxAmount, code: fxCode } : null
+    );
+    if (pair) {
+      await db.queryRaw(
+        `UPDATE transactions
+            SET merchant_name       = COALESCE(NULLIF($2,'Card purchase'), merchant_name),
+                merchant_normalized = COALESCE(NULLIF($2,'Card purchase'), merchant_normalized),
+                -- A subtitle-derived place overwrites; anything else only fills a
+                -- gap. Without the $4 branch the merged row depends on which
+                -- notification happened to arrive first, and the issuer half
+                -- usually wins by a few seconds -- keeping the worse string.
+                location            = CASE WHEN $4::boolean THEN COALESCE($3, location)
+                                           ELSE COALESCE(location, $3) END
+          WHERE id = $1`,
+        [pair.id, merchant, location, subtitleLocation !== null]
+      );
+      // The merged half records itself, or a replay finds nothing and merges
+      // it again (or, once the partner is superseded, mints a new row).
+      await db.queryRaw(
+        `UPDATE expense_metadata
+            SET card_last4            = COALESCE(card_last4, $2),
+                payment_method_detail = COALESCE(payment_method_detail, $3),
+                merchant_normalized   = COALESCE(NULLIF($4,'Card purchase'), merchant_normalized),
+                flags                 = COALESCE(flags, '[]'::jsonb) || $5::jsonb
+          WHERE id = $1`,
+        [pair.meta_id, ev.card_tail ?? null, ev.issuer ?? null, merchant, JSON.stringify([adoptedFlag(ev.event_id)])]
+      );
+      return { transactionId: pair.id, metadataId: pair.meta_id, skipped: "merged_pair" as const, flags };
+    }
+
+    // The receipt got here first. Fold into its row instead of writing a second
+    // one for the same purchase (see findWalletAdoptableTransaction). The column
+    // list IS the ownership boundary, the mirror of receipt-ingestion.ts's: the
+    // receipt read the register tape, so merchant, description, category and the
+    // items are its business and are absent here; `location` is wallet's own --
+    // nothing else writes a suburb -- and payment_method is filled only when the
+    // other lane left it unknown.
+    const adopted = await findWalletAdoptableTransaction(DEFAULT_OWNER_ID, amountAud, date, merchant, undefined, db);
+    if (adopted) {
+      await db.queryRaw(
+        `UPDATE transactions
+            SET location       = COALESCE($2, location),
+                payment_method = COALESCE(payment_method, $3)
+          WHERE id = $1`,
+        [adopted.id, location, paymentMethod]
+      );
+      let metadataId = adopted.metadata_id;
+      if (metadataId) {
+        await db.queryRaw(
+          `UPDATE expense_metadata
+              SET card_last4            = COALESCE(card_last4, $2),
+                  payment_method_detail = COALESCE(payment_method_detail, $3),
+                  currency              = COALESCE(currency, $4),
+                  flags                 = COALESCE(flags, '[]'::jsonb) || $5::jsonb
+            WHERE id = $1`,
+          [metadataId, ev.card_tail ?? null, ev.issuer ?? null, stated, JSON.stringify([...flags, adoptedFlag(ev.event_id)])]
+        );
+      } else {
+        // expense_metadata.transaction_id is UNIQUE, so this branch only runs
+        // when the adopted row genuinely has none -- and it must produce one, or
+        // the card tail and the flag that makes this idempotent have nowhere to
+        // live.
+        const meta = await db.queryRow<{ id: number }>(
+          `INSERT INTO expense_metadata (
+             transaction_id, source, source_message_id, merchant_normalized, amount,
+             transaction_date, card_last4, currency, payment_method,
+             payment_method_detail, line_items, flags, extraction_model,
+             notified_at, notif_app
+           ) VALUES ($1,'wallet',$2,$3,$4,$5,$6,$7,$8,$9,'[]'::jsonb,$10::jsonb,'wallet-interpret',$11,$12)
+           RETURNING id`,
+          [
+            adopted.id, sourceRef, merchant, ev.amount, date, ev.card_tail ?? null, stated,
+            paymentMethod, ev.issuer ?? null, JSON.stringify([...flags, adoptedFlag(ev.event_id)]),
+            notifiedAt, ev.notif_app ?? null,
+          ]
+        );
+        metadataId = meta!.id;
+      }
+      return { transactionId: adopted.id, metadataId, skipped: "adopted" as const, flags };
+    }
+
+    const txn = await db.queryRow<{ id: number }>(
+      `INSERT INTO transactions (
+         transaction_date, description, amount, amount_aud, category, payment_method,
+         merchant_name, merchant_normalized, transaction_type, owner_id, location,
+         foreign_currency_amount, foreign_currency_code, source, source_ref, row_index
+       ) VALUES ($1,$2,$3,$3,'other',$10,$4,$4,'debit',$5,$6,$7,$8,'wallet',$9,0)
+       RETURNING id`,
+      [date, merchant, amountAud, merchant, DEFAULT_OWNER_ID, location, fxAmount, fxCode, sourceRef, paymentMethod]
+    );
+
+    const meta = await db.queryRow<{ id: number }>(
+      `INSERT INTO expense_metadata (
+         transaction_id, source, source_message_id, merchant_normalized, amount,
+         transaction_date, card_last4, currency, payment_method,
+         payment_method_detail, line_items, flags, extraction_model,
+         notified_at, notif_app
+       ) VALUES ($1,'wallet',$2,$3,$4,$5,$6,$7,$8,$9,'[]'::jsonb,$10::jsonb,'wallet-interpret',$11,$12)
+       RETURNING id`,
+      [
+        txn!.id, sourceRef, merchant, ev.amount, date,
+        ev.card_tail ?? null, stated, paymentMethod, ev.issuer ?? null, JSON.stringify(flags),
+        notifiedAt, ev.notif_app ?? null,
+      ]
+    );
+
+    return { transactionId: txn!.id, metadataId: meta!.id, flags };
+  }, { maxWait: 10_000, timeout: 30_000 });
+}