fix(messages): make New Message modal full-screen below md

[catomean]

Summary

• Visitor feedback on https://orangecat.ch/messages at 320×800: the New Message modal's search-result rows shared screen space with the fixed mobile bottom nav and the "Ask Cat" FAB in the bottom-right corner — at risk of a tap landing on whichever fixed control happened to be under a row instead of the row itself.
• Investigation: I could not reproduce the reported hit-testing failure as a literal z-index-losing bug against current main — NewConversationModal's backdrop (z-modal = 55) legitimately outranks both MobileBottomNav (inline zIndex: 45) and the Ask Cat FAB (z-40) in the same stacking context (no ancestor of the modal sets a transform/filter/etc. that would trap it in a lower stacking context). I built a full integration reproduction — the real MessagePanel/NewConversationModal/MobileBottomNav/GlobalCatLauncher components, real classNames, a faked Supabase session via the Zustand auth store (this sandbox has no live credentials) — and at every coordinate matching the visitor's exact reported rects, elementFromPoint resolved to the modal's own content, not the nav or FAB. Detail in the session handoff.
• Regardless of the precise original mechanism, the report describes a real structural risk this PR removes outright: on a short/narrow viewport, a centered max-w-md card with backdrop padding leaves the fixed nav/FAB corner sharing the same screen coordinates as the bottom of the results list — and the FleetCrown feedback-widget FAB (external, not in this repo, always z-index: 2147483000) genuinely can't be beaten by any z-index we set, so that corner is never fully safe while any content renders there.
• Fix: below md (768px, the same breakpoint MessagePanel already uses for its own mobile/desktop layout split), the modal is now a full-bleed sheet — h-full w-full, no rounding/border, no backdrop padding — covering the entire viewport. Nothing renders where fixed chrome could ever intercept a tap. The results list fills the remaining height (flex-1) and reserves env(safe-area-inset-bottom) so the last row also clears a device's home-indicator/gesture-bar inset. md+ (tablet/desktop) is untouched.

Test plan

• npx tsc --noEmit — clean
• npx eslint src/components/messaging/NewConversationModal.tsx — clean
• Full local npm run test:unit (2638 tests) and full pre-push suite — green
• Playwright verification against the real running app at 320×800 (real MessagePanel/NewConversationModal/MobileBottomNav/GlobalCatLauncher, /api/profiles mocked with 10 fake rows, real auth faked via the Zustand store since no live Supabase credentials exist in this sandbox): before the fix, the modal card ended around y676 leaving the fixed nav (y719-800) and FAB (y664-720) exposed underneath; after the fix, the card is exactly {x:0,y:0,w:320,h:800} and every previously-exposed coordinate — the old FAB center, the old nav center, the viewport's bottom-right corner — now resolves to real modal row content, not the nav or FAB.
• Verified md+ (1024×768) is pixel-identical to before: card {x:288,y:113,w:448,h:542}, border-radius:6px, border-width:1px — matches the original max-w-md rounded-md border treatment exactly.
• Screenshots and full investigation notes captured in the session handoff (~/.fleetcrown/sessions/orangecat.md); the debug scaffolding used to reproduce this (a temp preview route, a temp window.__debugAuthStore hook, a temp routes.ts entry) was fully reverted before committing — this diff is exactly the one file above.

🤖 Generated with Claude Code