S67OLW

Every scenario is implemented, contract-exact, and tested through the web layer with pinned figures: budgets are the brief's table seeded per book and editable (src/money/dashboard.ts:20, migration 7 in src/db/database.ts); over-budget is a verdict plus red bar, hatched excess and 'over by' (tests/dashboard.test.ts asserts 295.00); the six-month trend stores empty months as null and draws a dashed 'no entries' frame, never zero (trendOf + test asserting 5 empty months and exact 4500.00/1276.30); net worth reuses the crypto part's cash-plus-coins figure at today's prices with last-known marked and missing rates refused, asserted at 3223.70; merchants group by receipt merchant else note with totals and counts; an empty book says 'Nothing here yet' with next steps and no svg (asserted), and an empty month says 'Nothing in September 2026' — polished after delivery (commit c44c932) to draw no bars at all, with the test tightened. Mobile is handled in CSS (stacking under 720px, scaled SVG) and the first screen element is the verdict. The player even fixed an empty-month zero-bar edge case on their own after captures were taken. Only deductions: the verdict's 'fine' counts budgeted categories only (transparent in its sub-line, a defensible reading), the done note's test count went stale (says 8 new, file now has 9), and I could not visually inspect the four delivered dashboard PNGs (no images in my evidence) — their existence and post-fix timestamps, plus the rendered-HTML code and web tests, stand in.

The new code is clean and honestly named. The pure model (src/money/dashboard.ts, 177 lines) keeps budgets, trend, merchants and the month verdict as small, well-named functions (budgetLines, trendOf, merchantsOf, verdictOf) with the PINNED_BUDGETS table straight from the brief and comments that state the null-means-empty contract. The view layer (src/web/dashboard.ts) separates statusLine/budgetSection/trendSection/worthSection/merchantsSection; every interpolated value goes through escape(). No measured duplication was supplied, but I read the whole diff: there is no copy-pasted block of any size. Nits that keep it from 9+: two unused imports (MonthPoint at src/web/dashboard.ts:8, describeMissing at src/web/dashboard.ts:10 — the missing-rates text is built inside verdictOf instead), a pure convenience re-export export { nextMonth } at src/money/dashboard.ts:177 that the module itself never uses, the !e.transferId && !e.tradeId predicate repeated four times within src/money/dashboard.ts (lines 66-67, trendOf, buildDashboard) where one named isPlainEntry helper would do, and a second month-name table (shortMonth, src/web/dashboard.ts:~160) beside MONTH_NAMES in src/web/views.ts:16.

A newcomer could change this safely. The dashboard follows the codebase's established architecture: a dependency-free pure module (money/dashboard.ts) that takes entries/budgets/rates as arguments and returns a typed Dashboard, consumed by thin Hono routes plus string views — the same split as report.ts and crypto.ts, and net worth is reused from the crypto part rather than recomputed. Functions are short (the longest, trendSection, is 25 lines), nesting is shallow, chart dimensions and pads are named constants (W/H/padL..., src/web/dashboard.ts:131), and limits are named parameters (months = 6, limit = 5). Boundaries are handled where failure is real: the /budgets POST whitelists the category, rejects negative or malformed amounts, and treats an empty amount as removal (src/web/dashboard.ts:47-63); the month query is validated with isValidMonth; unconvertible spending degrades to an honest 'Cannot tell' verdict naming the missing rate instead of a partial number. Behavior is pinned by 8 new tests (tests/dashboard.test.ts) covering the model, the page, the JSON API, the viewer-cannot-edit rule and the empty book. Minor comprehension costs: the fixed-point y-scale expression (m * BigInt(Math.round((H - padB - padT) * 1000))) / scale) / 1000 in trendSection is cleverer than it needs to be for a bar chart, the verdict serializer in /api/dashboard is one long nested ternary (src/web/dashboard.ts:77), and nextMonth/previousMonth are imported from two different modules because of the re-export.

The four delivered captures (probe ed472421) show a genuinely designed screen: the verdict is the first and largest element in all states — green 'This month is fine' with spent/left figures, a red 'Over budget' naming eating-out and 'over by 155.00', and 'Nothing in September 2026' with a link to the last month with entries. Budget bars are sorted by usage with right-aligned tabular figures, the half-year chart draws dashed 'no entries' frames instead of zero bars, net worth reads 3 128.70 EUR with a cash+coins breakdown, and merchants are ranked with counts. Small nitpicks: SVG axis text is small, and the h1 is the tiny uppercase book name rather than the verdict.

Markup is strong: lang="en", viewport meta, skip link, header/nav/main landmarks, sections with aria-labelledby, budget bars as role="img" with spoken aria-labels (spent of budget), a screen-reader table behind the trend SVG carrying 'no entries' per month, labelled budget form, focus-visible outlines, and a dark scheme. Over-budget is not color alone — red text says 'over by 155.00' and the category sorts first. Contrast of ink/muted/green/red on the warm background reads well in the captures. Minor: the key verdict is a

not the h1, warn-color text is borderline contrast, and 12px SVG axis labels are small (mitigated by the sr-only table).

The ~375px capture shows the page holding up: sections stack, budget rows reflow to tag/bar/amount lines, no horizontal scrolling or truncation, and the verdict ('Over budget') is the first thing visible, as the brief demands. The trend chart scales down proportionally without overflow. Deductions: chart text becomes very small at phone width, the month-nav wraps so the next-month button sits alone on its own line, and the nav/prev-next touch targets are on the small side.

The dashboard commit adds a purposeful 8-test suite (tests/dashboard.test.ts) with exact-value assertions for every scenario that is testable in code: budgets seeded from the pinned table, spending per category against budget (62.40 groceries, rent exactly on budget at ratio 1), the over-budget verdict naming eating-out and over-by 295.00, six-month trend with empty months as null not zero (model-level and rendered/API-level: 'no entries', empty:5), merchants with receipt-naming and totals (24.08), net worth 3223.70 through both page and JSON API, empty-book guidance with no SVG drawn, and the viewer permission denial. Unit level (pure buildDashboard/trendOf/verdictOf) is complemented by real integration through the Hono app. A deterministic probe confirmed 63 tests pass, 0 fail, on the participant's machine. Gaps keep it from 9+: the new 'unknown'/'Cannot tell' missing-rate verdict branch and negative-budget validation are implemented but never exercised by a test at dashboard level, and no automated check exists that the verdict renders first (the 'in the hand' ordering rests on screenshots alone).

The pinned budget table is declared once and matches the brief exactly; it is seeded into a budgets table (migration 7 + new books) and the dashboard always reads it from the DB, never from markup. Every displayed figure is derived on request from the book: category spending via totalInHome, six-month trend with no-entry months as null (drawn as a labelled dashed frame, never a zero bar), merchants from confirmed receipts' merchant names else the note, and net worth reusing the crypto part's cash-plus-coins valuation at today's/last-known prices with stale marking. Missing rates refuse with a 'Cannot tell' verdict instead of partial numbers; the empty book renders no chart. Data layer separation is clean: pure shaping in src/money/dashboard.ts, SQL confined to src/db/database.ts, presentation only formats. Deductions: the headline verdict's spent figure counts only budgeted categories (unbudgeted spending shows in its section with 'no budget' but is excluded from the headline), and a non-reportable month in the trend is drawn as a zero-height (class-marked) bar rather than a gap.

The dashboard is added as a textbook slice of the existing layering: pure domain model (src/money/dashboard.ts: budgetLines, trendOf, merchantsOf, verdictOf, buildDashboard — no I/O, empty months as null, MonthVerdict as a discriminated union), storage isolated in database.ts (MIGRATION_7 budgets table + three prepared-statement methods), presentation in separate view functions. Dependencies flow one way (web→money/db; db→money for PINNED_BUDGETS seeding); merchantOf is injected as a callback so the domain never knows about receipts, and today is a parameter keeping the model testable. Features map to files and tests mirror them (tests/dashboard.test.ts covers model and HTTP separately). Right-sized: no chart library, server-rendered SVG, no needless layering. Minor deductions: src/web/dashboard.ts (212 lines) fuses routing, model assembly, JSON API serialization, and HTML views in one file (though sectioned); the netWorth/today helpers it reuses live in src/web/crypto.ts rather than the money/ domain layer, a pre-existing reach-across the dashboard inherits; database.ts remains a 33KB single-file data-access blob.