Fix top review findings: coverage labeling, error handling, UX dead ends

[daphnehanse11]

Fixes the highest-priority findings from a full code review of the app (backend, frontend, and a live walkthrough).

What users see wrong today

1. "No coverage" when the real answer is "full-price marketplace." Anyone with a $0 premium tax credit (income over the 400% FPL cliff, which is back in 2026) or anyone 65+ showed no coverage pill at all — reading as "you lose insurance."
2. Good news framed as bad. An income drop that moves the whole family onto Medicaid ($300/mo → $0) was headlined "Your new income ends your ACA tax credit."
3. Silent zeros. Any PolicyEngine calculation error was swallowed into $0, so bad input produced a confident all-zeros result instead of an error.
4. Raw platform errors. Simulations regularly take 30–60s on a Modal cold start; without maxDuration/timeouts, users saw Unexpected token 'A'... when the proxy tried to JSON-parse an HTML error page.
5. Seven questions, then a dead end. Households with employer coverage answered the whole wizard before being told the tool can't help them.

Changes

Backend

• Coverage typing: Marketplace (full price) for $0-PTC people, Medicare for 65+ (dual-eligibles surface as Medicare); counterfactual remaps carry the new field through
• _run_simulation raises on failed core variables (medicaid, chip, premium_tax_credit, slcsp); lcbp/chip_premium stay soft-default 0 since their absence is a data gap
• Move events: after-side bronze estimate now uses the destination state's bronze/silver ratio (was ~$1,000/yr off on e.g. TX→CA)
• lcbp/chip_premium added to metadata with a non-benefit category, so totalBenefits no longer counts the gross bronze premium as a benefit; also fixes the raw chip_premium label leaking to the API
• 500 responses no longer include internal exception text (logged server-side instead)

Frontend

• ESI heads-up moved to wizard step 5, with an explicit "Continue without employer coverage" path
• Result headlines prioritize per-person coverage transitions (Medicaid gained/lost, Medicare) over tax-credit mechanics; lost-PTC copy now says "you'd pay full price"
• Stale-response guard (requestSeq): a slow simulation can no longer save a scenario for a household you've already left or rewrite the URL; error card and change wizard are now mutually exclusive
• API routes: maxDuration = 120, 110s backend fetch timeout with a friendly 504, content-type-checked error parsing, deduplicated into src/lib/backendProxy.ts
• Share links: decodeScenario restores pregnantMember (pregnancy scenarios previously recomputed against a non-pregnant baseline)
• ResultsView hooks moved above the early return (hooks-order crash); added app/error.tsx error boundary
• Move-specific silver explainer sentence only shows for move scenarios
• README live URL fixed (old link 404s)

Verification

• pytest tests/ — 34 passed
• npm run build + tsc + eslint clean (3 pre-existing dead-code warnings untouched)
• Manual walkthrough on localhost against the real backend: above-cliff couple shows Marketplace (full price) pills; 66-year-old shows Medicare; family income drop headlines "qualifies you and your family for Medicaid"; ESI "Yes" blocks at step 5; invalid state returns a clean JSON error
• Known upstream issue (pre-existing, not addressed here): for a childless CA couple, PolicyEngine reports both PTC=$0 and Medicaid=$0 in a ~$30k–32k band (~142–151% FPL) — the two eligibility checks disagree. Flagged for separate investigation.

🤖 Generated with Claude Code

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
cc-clean	Ready	Preview, Comment	Jul 27, 2026 6:35pm
coverage-compass	Ready	Preview, Comment	Jul 27, 2026 6:35pm

[daphnehanse11]

Second review round (adversarial pass over this diff) found and fixed in 3d2bef1:

• Error sanitization was incomplete: the new raise-on-failure path embedded raw PolicyEngine exception text and returned it as a 400, and modal_app.py (the production backend) had no sanitization at all. Now a dedicated SimulationComputeError carries a user-safe message (details logged server-side), the Flask API returns it as 500, and Modal matches.
• Marriage/divorce counterfactual side-simulations now run non-strict — their aggregated results are discarded, so a failed variable there no longer aborts a whole comparison that previously succeeded.
• restoreScenario now invalidates in-flight simulations (clicking a saved scenario during a run no longer gets silently overridden).
• decodeScenario rejects unknown event types from crafted share links (previously rendered "undefined …" headlines).
• Backing out of the ESI heads-up fully resets the answer in the "So far" summary.

Full backend suite: 34 passed. One deploy-time check for the reviewer: maxDuration = 120 in the API routes requires Vercel Pro or Hobby with Fluid compute (limit 300s); legacy Hobby without Fluid caps at 60s.

🤖 Generated with Claude Code