docs: monthly expense baseline and emergency reserve analysis

One-off analysis, nothing built. Realistic baseline $4,140/mo -> $24,800 for
six months, against $89,770 already accessible ($81,017 loan redraw + $8,753
offset).

Records four corrections the raw data needs before any restatement:
misfiled Raiz/Vanguard/moomoo debits counted as spend, `other` credits read as
negative spend, `government` conflating ATO with rates/rego, and `fees` being
mostly annual.

CLAUDE.md gains two traps found while doing it: partial split coverage inside a
category is usually correct rather than a gap (only shared utilities and
subscriptions are split), and the loan repayment is voluntarily above contracted
($2,500 vs $1,190.54 per fortnight) with the difference recoverable via redraw.

CLAUDE.md

@@ -135,6 +135,19 @@ 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.
 
+**Partial split coverage inside a category is usually correct, not a gap.** Only
+*shared* items are split. `utilities` sits at 69% yours because Globird, OVO, GWW
+and home telecoms are split while Telstra, Vodafone, Optus and JB Hi-Fi Mobile
+are personal. `subscriptions` is 91% because Uber One, Amazon Prime and OnePass
+are shared while Claude, OpenAI, Anthropic, OpenRouter, You.com, LinkedIn, Xero,
+Billdu, Spotify and Patreon are not. `fees` and `charity` are 100% yours and
+correct. Check the merchants before concluding a rule was never applied — a
+category-level ratio that "looks wrong" usually is not.
+
+Splits exist in this app from **2026-01-09** only; earlier splits lived in
+SplitMyExpenses. So a trailing-12-month per-person series splices six months of
+gross onto six months of net. Use Feb–Jun 2026 for anything per-person.
+
 ### The shared loan
 
 The loan is a **separate ledger**, not a shared expense and not a settlement
@@ -152,6 +165,19 @@ 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.
 
+**The repayment is voluntarily above contracted, and the gap is the largest
+flexible cost in the whole picture.** Contracted is $1,190.54/fortnight
+($2,579.50/mo annualised); the actual direct debit is $2,500.00/fortnight
+($5,416.67/mo). That is $2,837.17/mo of overpayment, and it is *not* sunk — it
+shows up as `statements.redraw_available`, which grew $62,387.17 → $81,017.42
+across the two most recent loan statements. Sonu returned to $1,250/fortnight in
+July 2026 after the reduced $750 period during her leave.
+
+Treat the repayment as two figures whenever asking "what does this cost me":
+the contracted floor and the actual. `scheduled_repayment` holds the actual
+($2,500), not the contracted minimum — the contracted figure is not in the DB at
+all. See `docs/expense-baseline.md`.
+
 ### Import Date (`created_at`)
 
 `transactions.created_at` is the import timestamp (DB default `now()`). In the transactions and shared views, the "Imported" column shows:
@@ -349,6 +375,11 @@ See `README.md` → **Known Gaps / TODOs** for full details.
 - **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.
+- **Expense baseline / emergency reserve** — `docs/expense-baseline.md`. One-off
+  analysis, nothing built. Records four data corrections the raw numbers need
+  (misfiled Raiz/super/brokerage debits, `other` credits read as negative spend,
+  `government` conflating ATO with rates/rego, `fees` being mostly annual) and
+  why only Feb–Jun 2026 is trustworthy for per-person figures.
 - **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).

docs/expense-baseline.md

@@ -0,0 +1,211 @@
+# Monthly expense baseline and emergency reserve
+
+Analysis run 2026-07-26. **One-off analysis, not a feature** — nothing in the app
+computes these numbers. Read "Reproducing this" before trusting a restated figure.
+
+The question: how much should be held in reserve to cover 6–12 months of expenses?
+
+## Answer
+
+| Scenario | $/mo | 6 months | 12 months |
+|---|---:|---:|---:|
+| Survival — contracted loan repayment, essentials only | 3,040 | 18,250 | 36,500 |
+| **Realistic — contracted loan, + dining and charity** | **4,140** | **24,800** | **49,600** |
+| Status quo — keep overpaying the loan, normal life ex-travel | 5,557 | 33,500 | 67,000 |
+
+Use the middle row. The survival row assumes dining is cut to zero and stays
+there, which is not a plan anyone executes for six months.
+
+Against that, liquidity already available (statements 133 and 131, 2026-06-30):
+
+| | |
+|---|---:|
+| Loan redraw | 81,017.42 |
+| Offset balance | 8,753.00 |
+| **Accessible** | **89,770.42** |
+
+That is 3.6× the six-month target and 1.8× the twelve-month one. Redraw grew
+$62,387.17 → $81,017.42 across the last two loan statements, matching the
+overpayment rate — the money spent killing the loan faster is still reachable.
+
+**Caveat on counting redraw as the reserve.** It is available at AMP's
+discretion, and lenders reduce or freeze it exactly when a borrower looks
+distressed — which is when it would be needed. The rate also moved 5.54% → 6.29%
+between the two statements, so redrawn funds cost more than they did. Hold some
+genuine cash; it does not need to be $50,000.
+
+## This is *your* outgoings, not household spend
+
+The app only sees accounts that get imported. Sonu's own spending on the
+household is invisible to it. Grocery spend reads as ~$300/mo gross on 4–10
+transactions, which is implausible for a household and is partly explained by her
+paying from her own account.
+
+For "how much reserve do **I** need" that blind spot does not matter — your own
+outgoings is the correct measure. Do not relabel these figures as household
+totals; they are not, and they would be wrong by an unknown amount.
+
+## The loan has two floors
+
+This is the largest single lever and the reason there are three scenarios.
+
+| | $/fortnight | $/mo annualised | Your 50% |
+|---|---:|---:|---:|
+| Contracted minimum | 1,190.54 | 2,579.50 | **1,290** |
+| Actual direct debit | 2,500.00 | 5,416.67 | **2,708** |
+| Voluntary overpayment | 1,309.46 | 2,837.17 | 1,419 |
+
+Dropping to contracted cuts your loan cost by $1,418/mo. Sonu's obligation is a
+fixed 50% of the repayment (see CLAUDE.md → "The shared loan"), so it falls with
+it. Her rate returned to $1,250/fortnight in July 2026 after the reduced $750
+period during her leave.
+
+## Baseline composition
+
+Built from **Feb–Jun 2026** — the months where split data is trustworthy — with
+annual items annualised over 12 rather than divided by the 5-month sample.
+
+| Essential | $/mo | Note |
+|---|---:|---|
+| Loan (contracted, your 50%) | 1,290 | 2,708 at the current actual rate |
+| Transport | 326 | |
+| Insurance | 280 | annualised; your 55% |
+| Utilities | 278 | shared energy/water + personal mobile |
+| Subscriptions | 204 | shared household + personal/AI |
+| Groceries | 200 | see the blind-spot note above |
+| Card + package fees | 175 | annualised — see below |
+| Health | 167 | |
+| Rates + rego | 122 | annualised, your share |
+| **Essential** | **3,042** | |
+| + dining 675, charity 422 | 4,139 | charity is a Smith Family sponsorship commitment |
+| + typical shopping | ~4,800 | median 657, **not** the 1,709 mean |
+
+**Travel is excluded throughout.** At $3,153/mo of your share even post-cutover it
+would roughly double every figure, and it is the first thing that stops.
+
+## Four corrections the raw data needed
+
+Any restatement that skips these will be wrong. None are fixed in the data yet.
+
+**Micro-investing counted as spend.** Raiz ($9,406 / 27 rows), Vanguard Super
+($500) and moomoo ($300) sit in `other` as debits — $10,206/yr, ~$850/mo of
+phantom spend. These belong in `investment`, which is already excluded.
+
+**Incoming money counted as negative spend.** 17 rows in `other` typed `credit`
+($5,622 in the window). `SPEND_SIGNED` negates credits so refunds cancel
+purchases, but these are not refunds — they are money arriving. June 2025 shows
+*minus* $7,814 of total spend because two Wise credits of ~$16.7k each landed in
+`other`.
+
+**`government` is two unrelated things.** $25,554 of ATO payments (one annual
+bill, routed through Zen B2B and RewardPay to earn points) versus $2,054 council
+rates and $875 rego. Tax is not a monthly living cost and falls with income
+anyway; rates and rego are non-negotiable. Splitting them moves this line from
+$2,411/mo to $244/mo.
+
+**`fees` is mostly annual.** Of $2,599 post-cutover, $1,750 is an annual card fee
+and $349 a loan package annual fee. Recurring is ~$175/mo annualised, not the
+$520/mo the 5-month mean implies.
+
+## Splits: what is trustworthy and what is not
+
+Splits exist in this app from **2026-01-09** only; before that they were tracked
+in SplitMyExpenses. So a trailing-12-month per-person series splices six months
+of *gross* onto six months of *net* and is not a series at all. Use Feb–Jun 2026.
+
+**Partial split coverage within a category is usually correct, not a gap.** This
+was misdiagnosed once during the analysis. Verified composition:
+
+| Category | Your share | Split | Unsplit |
+|---|---:|---|---|
+| utilities | 69% | Globird, OVO, GWW, home telecoms | Telstra, Vodafone, Optus, JB Hi-Fi Mobile |
+| subscriptions | 91% | Uber One, Amazon Prime, OnePass | Claude, OpenAI, Anthropic, OpenRouter, You.com, LinkedIn, Xero, Billdu, Spotify, Patreon |
+| fees | 100% | — | credit card fees are personal |
+| charity | 100% | — | personal commitment |
+
+Only *shared* utilities and subscriptions are split. Mobile bills, AI
+subscriptions and card fees are personal and correctly sit at 100%. Do not
+"fix" these ratios.
+
+## Known weak spots
+
+- **`other` is $424–773/mo of your share and unclassified.** The single biggest
+  lever on accuracy. Clearing it with the rule preview
+  (`GET /api/rules/[id]/matches`) improves every other view at the same time.
+- **Five months is a short sample**, and it contains the Europe trip and the ATO
+  bill. Both are excluded, but they crowd out the ordinary months. The
+  SplitMyExpenses CSVs would stretch it to 18 months — worth having, not
+  blocking, since only an *aggregate ratio per category* is needed, not row
+  matching, so the combined-transaction problem does not bite.
+- **`/api/analytics/monthly` does not filter `reconciled_with_id IS NOT NULL`.**
+  48 rows are double-counted app-wide. Small, but real, and unrelated to this
+  analysis. The queries below do filter it.
+
+## Reproducing this
+
+Working queries are not checked in; they were run ad hoc against
+`postgres-personal`. The shape that matters:
+
+```sql
+-- Per-category monthly distribution, your share, Feb-Jun 2026.
+-- Mirrors src/lib/analytics-sql.ts, plus the corrections above.
+WITH s AS (
+  SELECT to_char(t.transaction_date,'YYYY-MM') m,
+    COALESCE(o.category_override, t.category, 'other') cat,
+    SUM((CASE WHEN t.transaction_type IN ('refund','credit') THEN -1 ELSE 1 END)
+        * (CASE WHEN t.interest_amount IS NOT NULL
+                THEN t.interest_amount ELSE COALESCE(t.amount_aud, t.amount) END)
+        * COALESCE(ts.share_percent, o.my_share_percent,
+            100 - COALESCE((SELECT SUM(x.share_percent) FROM transaction_splits x
+                            WHERE x.transaction_id = t.id
+                              AND x.participant_id <> 1), 0)) / 100)::numeric(12,2) amt
+  FROM transactions t
+  LEFT JOIN transaction_overrides o ON o.transaction_id = t.id
+  LEFT JOIN transaction_splits ts ON ts.transaction_id = t.id AND ts.participant_id = 1
+  LEFT JOIN statements st ON st.id = t.statement_id
+  WHERE COALESCE(t.owner_id, st.owner_id) = 1
+    AND t.reconciled_with_id IS NULL                      -- see weak spots
+    AND ((t.transaction_type IN ('debit','fee','interest')
+          OR t.interest_amount IS NOT NULL)
+         OR t.transaction_type IN ('refund','credit'))
+    AND COALESCE(o.category_override, t.category, 'other')
+        NOT IN ('transfers','investment','income')
+    -- correction 1: misfiled investments
+    AND COALESCE(t.merchant_normalized, t.merchant_name, '')
+        !~* '^(Raiz|Vanguard Super|moomoo)'
+    -- correction 2: incoming money miscategorised as 'other' credits
+    AND NOT (COALESCE(o.category_override, t.category, 'other') = 'other'
+             AND t.transaction_type = 'credit')
+    AND t.transaction_date >= '2026-02-01' AND t.transaction_date < '2026-07-01'
+  GROUP BY 1, 2
+)
+SELECT cat, count(*) mths, round(avg(amt),0) mean,
+       round(percentile_cont(0.5) WITHIN GROUP (ORDER BY amt)::numeric, 0) median
+FROM s GROUP BY cat ORDER BY mean DESC;
+```
+
+Use the **median** for anything recurring monthly and the **annualised total** for
+lumpy-but-certain items (insurance, rates, rego, annual card fees). Means are
+badly skewed here — shopping is 1,709 mean against 657 median.
+
+Corrections 3 and 4 (`government`, `fees`) are not expressible as filters; they
+need the category split by merchant, done by hand above.
+
+## If this becomes a feature
+
+The blocker is `other`, not the split model — the split data held up under
+scrutiny. Sequence:
+
+1. Recategorise Raiz / Vanguard Super / moomoo to `investment` (27 rows,
+   unambiguous) via the rule preview.
+2. Split `government` so rates and rego separate from ATO. Probably a new
+   category or a tag; `government` conflates a recurring obligation with an
+   annual tax bill.
+3. Work `other` down with the rule preview.
+4. Then a baseline panel on Insights, which already carries the
+   Regular/Occasional split this analysis is a stricter version of
+   (`REGULAR_CATEGORIES` in `src/lib/categories.ts`).
+
+A **liquidity vs baseline** view would be genuinely new rather than a restatement:
+`statements.redraw_available`, offset closing balance, and the
+contracted-vs-actual repayment gap are all in the schema and nothing reads them.