fix: correct the 2026 tax tables and clear the technical debt behind #6 and #7

[devRMA]

Technical debt found while investigating #6 and #7. No new features — every change here either fixes a wrong number, an unreachable control, or a test that was not testing.

The headline: every net salary the app produced was wrong

The two tax tables disagreed on their reference year. INSS was 2025 (minimum wage 1.518,00, ceiling 8.157,41); IRRF was older still, the February 2024 table (exempt to 2.259,20). Neither had been updated for 2026, and the calculator had never implemented the two rules from Lei 15.270/2025 that dominate its target salary range.

At the app's own default salary of R$ 5.000,00 with a 220h month:

	INSS	IRRF	Net	Hourly rate
Before	509,60	347,57	4.142,83	18,83
After	501,51	0,00	4.498,49	20,45

The app understated take-home pay by R$ 355,66/month — R$ 4.267,92 a year — and "Custo da Hora", its headline output, was off by 7,9%. Above the reduction's phase-out the error narrows to a flat R$ 8,09/month of over-withheld INSS at every salary over 4.354,27.

Values are from Portaria Interministerial MPS/MF nº 13 de 09/01/2026 (Anexo II) and the 2026 IRRF table:

• INSS — 7,5% / 9% / 12% / 14% over 1.621,00 / 2.902,84 / 4.354,27 / 8.475,55, capping the contribution at 988,09.
• IRRF — exempt to 2.428,80, then 7,5% / 15% / 22,5% / 27,5% with deductions 182,16 / 394,16 / 675,49 / 908,73.
• Simplified deduction of 607,20/month, applied whenever it beats the contribution — which is every salary below roughly 5.755, i.e. most of this app's users.
• The reduction that zeroes tax to 5.000,00 and phases out linearly to 7.350,00.

The last two are interdependent, which is why implementing either alone would have been worse than neither: the published reduction formula 978,62 − 0,133145 × bruto only lands on zero at 7.350,00 when the simplified deduction is the one being applied. Implemented together the result is continuous at both ends — verified at 5.000,01 and at 7.350,00.

Rounding was also losing a cent. Math.round(value * 100) inherits binary representation dust: 1.621,00 × 7,5% evaluates to 121,57499999999999 and rounded down to 121,57 instead of 121,58.

The other wrong number: phantom overtime on night shifts

calculateSuggestedExit and calculateWorkStats disagreed about the same journey. The stats function credits the reduced night hour (CLT art. 73 — 52'30" counts as an hour); the exit suggestion did not. AUTO mode's whole contract is "leave at this time and your balance is zero", and for any shift touching 22:00–05:00 it broke that:

entry 21:00, lunch 01:00–02:00, journey 08:48
suggested 06:48 -> balance +51 min, all of it billed as overtime

Both paths now share one helper. Where the reduced hour makes an exact match unreachable — no whole number of minutes maps to 420 credited minutes, only 419 or 421 — it settles on the closest, so the residue is at most a minute.

Separately, a long lunch drove workMinutes - workedBeforeLunch negative, placing the suggested exit before lunch ended. That made the range non-chronological, which calculateWorkStats rejects by returning zeroed stats — so the UI reported a balanced day while hiding real overtime.

Overtime rates were presenting a collective agreement as law

The tiers were hardcoded at 75% and 100% under an "Extras (CLT)" heading. The statutory floor is 50% (CF art. 7º, XVI); 75% comes from collective agreements. The rates are now editable and default to the legal floor, so a 75% agreement is still representable and everyone else gets correct numbers by default.

Crashes and a control nobody could reach

• The calculator could be bricked by its own storage.JSON.parse ran unguarded on the stored extras inside a mount effect, and the project has no error boundary — one malformed entry blanked the page with no recovery short of clearing site data. Worse, Number.parseFloat on a corrupt salary produced NaN, which was then written back to localStorage, so the app stayed broken on every subsequent load.
• The delete button was unreachable on mobile.flex-1 on a bare <input> keeps min-width: auto, flooring the extras row at its intrinsic width: 380px inside a 245px container at a 375px viewport, with overflow-hidden on an ancestor. The button sat ~70px off screen — not reachable by touch, not reachable by scrolling.
• The read and write effects raced on mount. The write closed over pre-restore state and overwrote saved values with defaults before the restore landed.
• maximumScale: 1 disabled pinch zoom, failing WCAG 1.4.4 for anyone who needs to magnify the page.
• The clipboard write was unguarded and unawaited: it threw in insecure contexts and reported "Copiado!" even when the write was rejected.
• progress was clamped in one branch and not the other, so a zero-length journey produced width: "Infinity%". A zero balance also rendered as overtime in the hero card while the summary card treated it as on target.

Tests that were not testing

Coverage advertised 95.8%. Vitest's v8 provider only reports files that a test imported, so that number described 9 files out of 28. The real baseline was 49.09% statements / 37.35% branches, with fifteen source files never imported by a test at all.

	Before	After
Statements	49.09%	100%
Branches	37.35%	100%
Tests	72	286

app/layout.tsx needed no exclusion — it only failed to import because next/font/google is normally transformed by SWC during a Next build, so outside that compilation Inter arrives as an object rather than a callable. A three-line mock makes it testable, and the test now asserts viewport has no maximumScale so the pinch-zoom fix cannot regress silently.

One e2e test was passing without asserting anything: should show side ads on desktop after delay bailed out with a bare return on narrow viewports, reporting success on Mobile Chrome and Mobile Safari without running a check.

Structure

work-calculator.tsx (573 lines) and salary-calculator.tsx (387) sat loose in components/ root, which AGENT.md reserves for nothing — they are Organisms. The masked numeric input was implemented three separate times, and work-calculator carried two near-duplicate copies of the hero card (lg:hidden and hidden lg:block) that had already drifted apart. Largest component is now 291 lines.

Both files also ran their own setInterval(..., 1000), re-rendering the whole tree twice a second to advance the same clock. There is now exactly one setInterval in the codebase.

Dead code and dead config

useIsMobile had zero callers. formatMinutes was referenced only by its own test, giving five tests' worth of false confidence in a function nothing shipped. Both calculateWorkStats and calculateSuggestedExit wrapped their bodies in try/catch, but new Date(garbage) returns an Invalid Date rather than throwing — the catch blocks were unreachable. .biomeignore has done nothing since Biome 2 dropped support for it.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
workload	Ready	Preview	Aug 1, 2026 9:11pm

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!