docs: capture this session's learnings for a future pickup

CLAUDE.md gains the traps a new session would otherwise re-discover:

- Rules: a zero-condition rule matches everything (rule 43 would split all ~3,700
  transactions); preview-then-apply-by-id is the safe pattern and why it beats
  auto-applying on ingestion; how run provenance works.
- Shared expenses: transaction_splits.settled is dead data; getParticipantBalances
  is correct and must not be 'fixed'; settlement cannot be attributed per trip.
- The shared loan: separate ledger, fixed 50% with a tracked receivable, why the
  share must not be derived from actual payments, and why interest stays as spend.
- Extraction: balance assertions are the check that works, do not derive
  opening_balance or add a totals assertion (both would be tautological), Gemini
  invents summary fields it was not given, empty statements must not throw, FX is
  per-date, and CSV comparisons need millisecond ordering.

The design doc records Phase 0 as done - including that the original Phase 0 plan
was wrong, since reading the code first is what prevented breaking a working
balance page.

Known Gaps lists what is open: the unbuilt phases, 11 failing assertions, the
uncategorised Up rows, and the CSVs sitting in 030490e's history.

CLAUDE.md

@@ -109,6 +109,49 @@ it silently deletes the unlogged remainder from spend totals.
 - `participants` — people; `id=1` is "Me" (the primary user)
 - `account_owner_mappings` — persists bank+account → owner assignments
 
+### Shared expenses and settlement — read before touching
+
+The model is under active redesign. See `docs/shared-expenses-design.md` for the
+proposal and what is already decided. Three traps:
+
+**`transaction_splits.settled` is dead data.** It is `false` on every row. Its
+only writer was `/api/splits/settle`, removed in `3f04cbd` because nothing called
+it and one request could mark all of a participant's splits settled. Do not build
+on this flag until settlement contexts exist.
+
+**`getParticipantBalances` computes `splits − payments` and is correct.** Do not
+"fix" it to exclude settled splits — the payments that settled them are still
+subtracted, so you would double-count. The two settlement models (running tab vs
+per-split flag) must not be mixed.
+
+**Settlement cannot be attributed per trip.** `split_payments` records only
+from/to/amount/date. Any per-trip settled/unsettled figure is fabricated; the
+trip view used to show one and always reported 100% unsettled. Trips show share
+only, and point at `/shared` for real balances.
+
+Also: settlements already exist twice. Four of eight `split_payments` match an
+offset-account credit exactly on amount and date, with `linked_transaction_id`
+populated on only one. And Sonu's loan contributions (`…emi` in the offset
+account, 39 rows, $37,980.24) are categorised `transfers`, indistinguishable from
+ordinary internal transfers.
+
+### The shared loan
+
+The loan is a **separate ledger**, not a shared expense and not a settlement
+context — a contribution must never be able to settle a dinner. Sonu's obligation
+is a fixed 50% of the repayment; actual contributions vary, and the difference is
+a tracked receivable ($4,000.00 over 2025-07 → 2026-06).
+
+Do not derive the share from actual payments. During her leave the obligation did
+not change, only the payment did — a percentage-of-actual model would silently
+redefine her share as 30% and make the shortfall vanish.
+
+Loan interest reconciles exactly: `repayments − interest − fees = balance
+reduction`. It stays categorised `loan_interest` and counts as spend — over 12
+months $63,500 of cash left and debt fell $44,127.36, and the $16,523.64
+difference bought nothing. Excluding it would leave the balance sheet unable to
+reconcile cash out against equity gained.
+
 ### Import Date (`created_at`)
 
 `transactions.created_at` is the import timestamp (DB default `now()`). In the transactions and shared views, the "Imported" column shows:
@@ -123,6 +166,80 @@ Conditions are AND-evaluated. Fields: `merchant_normalized`, `description`, `cat
 
 `contains` and `equals` operators are case-insensitive (both sides `.toLowerCase()`).
 
+**A rule with zero conditions matches every transaction.** Both apply paths use
+`conditions.length === 0 || conditions.every(...)`. Rule 43 "Home 50/50 Sonu" has
+no conditions and a 50/50 split action — applying it blindly would split all
+~3,700 transactions with another participant. That is what `manual_only` is for:
+those rules are excluded from bulk runs and fire from the transactions page
+against a hand-picked selection.
+
+### Previewing a rule before applying it
+
+`GET /api/rules/[id]/matches` is a dry run — it writes nothing and returns only
+the transactions a rule would actually *change*, with already-correct rows
+summarised as a count. The Preview button on the rules page uses it.
+
+Apply then goes through `POST /api/transactions/bulk` with `action: "apply_rule"`
+and **explicit transaction ids**, not the conditions. That is the safety
+property: a rule whose conditions are too broad cannot reach further than what
+the preview showed and the user ticked.
+
+Prefer this over auto-applying rules on ingestion. It fails safe, works
+retroactively, and tells you which rules are consistent enough to automate later.
+
+### Rule apply history
+
+`rule_apply_runs` snapshots the before-state so a run can be reverted, and since
+migration 0017 also records `rule_id`, `rule_name` and `source`
+(`all` | `rule` | `selection`). `rule_name` is denormalised deliberately and
+there is no FK to `rules` — history must stay readable after a rule is renamed or
+deleted, and deleting a rule must not cascade away the audit trail.
+
+`GET /api/rules/runs/[id]` diffs that snapshot against current values. Rows
+changed by something else since the run are flagged, because reverting restores
+the pre-run value and discards the later edit.
+
+### Trusting extracted statement data
+
+**Balance assertions are the check that works.** `getStatements` computes
+`opening + movement − closing`; the statements page flags any statement that does
+not reconcile. Sign depends on what the balance means — on a credit card or loan
+it is what you owe, so spending increases it; on a transaction or offset account
+it is what you hold. 11 pre-existing statements currently fail, ~$4,177
+unexplained, including two adjacent ANZ statements off by exactly ±$230.38 (a
+transaction filed against the wrong one).
+
+**Do not derive `opening_balance` from `closing − movement`.** It is an accounting
+identity, so every statement would reconcile and the check would go permanently
+green. A null that reads "unverified" is worth more than a number that is right
+by construction. For the same reason, do not add a totals assertion comparing
+`total_debits` to the summed rows — those totals are now *computed* from the rows
+(see the N8N `Parse Gemini Result` node), so that check can never fail.
+
+**Gemini invents summary fields the statement does not print.** Wise PDFs show
+only a closing balance; asked for an opening balance anyway, the model produced
+11,277.08 against a truth of 0.00, and on another statement read the running
+balance of the oldest row. Every transaction was extracted perfectly in both
+cases — verified row for row against the CSV exports. When a balance assertion
+fails, suspect the summary before the transactions.
+
+**Gemini drops rows silently on long tables.** `finishReason` was `STOP`, not
+`MAX_TOKENS`, so raising `maxOutputTokens` does not help. This did *not* actually
+occur on the Wise imports (that was the summary bug above), but it is why an
+empty statement must not throw: a document that errors never gets tagged, so it
+is re-fetched every poll forever and blocks everything behind it in the queue
+(`ordering=-created`, `page_size=1`).
+
+**FX is per transaction date**, via Frankfurter (ECB daily, free, no key), with
+weekends resolving to the prior publication. A single spot rate across a 15-month
+statement is wrong by up to 20%. Wise's own rates are more accurate in principle
+but differ by only 0.05% and exist on 44 of 194 rows, so mixing bases is not
+worth it.
+
+**When comparing CSV exports to extracted data, order by full timestamp
+including milliseconds.** Two of one statement's rows are 1ms apart; dropping the
+fraction reversed them and produced a bogus opening balance.
+
 ## Development Patterns
 
 ### Adding a new API route
@@ -226,3 +343,24 @@ break or go stale. And the views are **not** owner-scoped and do **not** merge
 See `README.md` → **Known Gaps / TODOs** for full details.
 
 **Payment provider tracking**: `merchant_normalized` currently conflates payment provider (PayPal, Afterpay, Zip) with the actual merchant. Plan: add `payment_provider` column, update Gemini prompt to extract it separately, backfill from `merchant_name` patterns, surface in UI filters.
+
+### Open as of 2026-07-26
+
+- **Shared expenses redesign** — `docs/shared-expenses-design.md`. Phase 0 done;
+  Phases 1–4 unbuilt. Deliberately paused to live with the current behaviour
+  before committing to a model designed in one session.
+- **11 statements fail the balance assertion**, ~$4,177 unexplained. Predates
+  this work. One ANZ statement is off by exactly $0.50, traced to a misread digit
+  in fee rows ($5.00 vs $5.50).
+- **28 Up Bank debits are categorised `other`** ($3,760.74). Up only categorised
+  16 of 88 rows. The `Payee` field is populated throughout, so merchant rules plus
+  the rule preview should clear most of it.
+- **Up item sales are categorised `income`** ($4,238.04 across 19 credits — iPad,
+  drone, camera). Correct in that they are excluded from spend, but it mixes
+  asset disposals into the income line alongside salary.
+- **`payment_method` is not shown in the transactions list** — settable on create
+  and edit only. Worth a column or filter if cash becomes routine.
+- **Raw statement exports live in `dump/`**, gitignored since `31a8177`. They were
+  committed by accident in `030490e` and remain in that commit's history; the repo
+  has no GitHub remote, so exposure is limited to the local Gitea. Purging history
+  was offered and not actioned.