pirania: fix dead captive portal redirect server (uhttpd-mod-lua contract)

[luandro]

Problem

Pirania's HTTP capture redirect server (pirania-uhttpd, port 59080) never listens: procd gives up after five respawn attempts and every HTTP capture redirect lands on a closed port, so the captive portal is dead on any build that includes ba4f3824 ("fix: luacheck: pirania", merged 2026-08-05).

Root cause

That commit localized handle_request in packages/pirania/files/www/paradise... packages/pirania/files/www/pirania-redirect/redirect and added return handle_request.
That is valid Lua and it keeps the busted tests green, because they load the file as a module via require(). But the real consumer, uhttpd-mod-lua, uses a different contract: it loadfile()s the handler, executes it, and then looks up the global handle_request. With a local function the lookup fails, uhttpd exits with Error: Lua handler ... provides no handle_request() callback, procd respawns five times and gives up.

Fix

• Keep handle_request global (with a -- luacheck: globals handle_request directive so lint stays clean) and keep the module-style return, so both consumers work.
• Add a test that loads the file the way uhttpd does (loadfile() + assert global), so this regression cannot pass CI again.

Verification

Found live on LibreMesh master builds flashed onto UniFi 6 Lite nodes: /etc/init.d/pirania-uhttpd status = not running; manual invocation of the init command line reproduced Error: Lua handler /www/pirania-redirect/redirect provides no handle_request() callback instantly; after the fix the service stays up, port 59080 answers 302 to the portal URL, and /portal/auth.html answers 200. CI will also run the new contract test.

Also in this PR (second commit)

The packaged default catch_bridged_interfaces only listed wlan0-ap, so on two-radio nodes (UniFi 6 Lite) 5 GHz clients bypassed the portal. Added wlan1-ap; consumer is an nft set of type ifname, so the extra default entry is harmless on single-radio devices.

[a-gave]

Thanks! Assuming it has been tested/fixed also on a real device: approving
And personally sorry for having introduced the regression, the luacheck globals is more appropriate!