fix: liability sign convention, account typing, and net-worth coverage boundary

[KenTaniguchi-R]

Fixes #84. Fixes #85. Fixes #86.

Five commits. The first two are a coupled pair and must land together; the rest are unrelated work that ended up on this branch (see the note at the bottom).

---

1. 59c1400 — Liability balance signs (#84, #85)

accounts.current_balance had no documented convention and the ingest paths disagreed:

source	credit/loan balance	evidence
Plaid	positive when owed	"For credit and loan accounts, a positive balance indicates amount owed" — Plaid docs for balances.current. plaidAmountToCents did no sign flip.
SimpleFIN	negative	real card balances arrive negative
demo seed	negative	demo.ts:225, :253
existing test	positive	accounts-queries.test.ts:91 inserted a credit card at +50000

getAccountSummary and getDashboardSummary computed netWorth = totalAssets - totalLiabilities, correct only for positive-stored liabilities. Against real data it added the debt — demo household showed $71,700 instead of $51,700, masked by the Debts tile rendering Math.abs().

Separately, SimpleFIN sends no account type and simplefin-connect-flow.tsx defaulted everything to "checking", so four credit cards and an Alpaca brokerage were filed as deposit accounts and Debts read $0.00. That is why the inverted sign went unnoticed: nothing was ever classified as a liability.

Approach — canonicalize on negative = owed, normalized at ingest (plaidBalanceToCents, kept separate from plaidAmountToCents, which also converts transaction amounts). Net worth becomes the plain sum of balances, so it no longer depends on classification at all and a mis-typed account can never invert it again. availableBalance/creditLimit are not flipped — for a credit account those are capacity, not debt. MCP tools report liabilities as a positive magnitude. Added inferAccountTypeFromName, ordered so "Discover Bank Checking" doesn't match the card-brand rule.

Verified on the real household — net worth held at $52,942.44 while Assets went $52,942.44 → $57,500.85 and Debts $0.00 → $4,558.41.

---

2. 6ed2d68 — Net-worth coverage boundary (#86)

The dashboard read ↑ $55,214.00 (2430.7%) past 6 months. Nothing earned it: five accounts holding $51,703 have their first-ever snapshot on 2026-08-28, so every earlier point counted them as $0 and plotted the credit cards alone.

2026-05-31    2 of 10 accounts
2026-06-01    3 of 10
2026-06-07    4 of 10
2026-07-29    5 of 10
2026-08-28   10 of 10   <- boundary

Carry-forward (#68) fixed gaps between snapshots; it cannot fix a gap before the first one. Per the #67 precedent this doesn't invent the missing history — it says which stretch is missing.

getNetWorthHistory now emits coveredAccounts/totalAccounts per point. coveredTrendDelta measures across the fully covered span only and returns null below two such points. The chart splits at the boundary — muted dashed line and hatched band before, solid area after — with a caption naming the shortfall in words so the treatment is never colour-alone. Zero deltas lose the arrow, which pointed somewhere the number didn't.

Three treatments were mocked against the real 84-point series before building; this is option B. Option A (trim to full coverage) collapses 6M/1Y/All to two points; option C (suppress the percentage only) leaves a line diving to −$13,456 unexplained.

Known limitation: nothing distinguishes "the account didn't exist yet" from "we have no data yet" — accounts.createdAt is when Ledgr learned of an account, not when it opened. A genuinely new account is also marked partial. Conservative, but wrong in that case.

---

3. eaad082 — README screenshots

Budgets and reports shots alongside the existing hero and Plaid Link images.

4. 44c620e — Demo seed rate and budget coverage

threshold was Math.round(30 / tmpl.frequency) — an interval in days — but it gates a counter advancing by exactly 1 per day ((dayOffset * 31) % 30), so hash < threshold yields threshold occurrences per 30-day window, not one every threshold days. A 4×/month template fired 8 times. Budget limits also covered only 5 categories, so most of the generated month landed in "Everything Else"; now 16 categories sized to the generated trend, with Groceries deliberately under it so the over-budget state is represented.

5. ad92688 — Security policy and the next dev agent-rules block

SECURITY.pdf documents private vulnerability reporting. The CLAUDE.md block is generated and re-added by next dev; committing it keeps the tree clean.

---

Test changes

accounts-queries.test.ts:91 encoded the opposite sign convention from every real data source — which is why a $20,000 error never failed a test. Fixtures across dashboard-queries, report-queries, mcp-net-worth-history, and plaid-exchange now store liabilities negative with unchanged net-worth expectations. Added a regression test asserting netWorth < totalAssets when debt is present, a plaid-exchange assertion that the normalized value reaches balance_history, and 10 new cases for the coverage boundary.

NetWorthPoint splits into NetWorthSeriesPoint (no coverage) plus the coverage-carrying dashboard shape. The reports series has the same leading-gap exposure and keeps the base type — surfacing coverage there is separate work.

Full suite green: 113/113 test files.

Why one PR

#86 was developed on a branch stacked on this one, since both touch getNetWorthHistory and NetWorthPoint. Its commit landed here rather than on that branch, and the branch ended up at main's head. Untangling meant force-pushing two branches with open PRs, so they were kept combined instead and #97 was closed.

🤖 Generated with Claude Code