[rmstdope]

Bead: nr-ri8n, Players on different home networks can play a network game together (mockup docs/ui/nr-ri8n-internet-play.html).

What changes

• web/src/netplay/peer_link.ts: both peer connections use ICE_SERVERS = [{ urls: "stun:outpost.rmstdope.eu:3478" }], so the offer and answer carry each browser's public address. Gathering keeps its 3 s cap, so with outpost down only same-network play works. No TURN (that is nr-1wbx).
• web/src/netplay/session.ts: CONNECT_LIMIT_MS is now 10 s instead of 15 s ("about 10 seconds at most"), counted on both sides from the joiner being in the room.
• web/src/ui/play_together.ts and web/main.css: the address line beneath the two cards, the new can't-connect message (focus on Start again, as the acceptance states), and the new too-slow advice.
• README, architecture.md, and the header comment of the Playwright spec.

Validation

• ./scripts/gate-full.sh: passed (1085 Vitest tests, typecheck, every Rust/Python leg).
• pnpm exec playwright test web/integration/specs/play-together.integration.spec.ts: 9 passed. Two browsers on one machine still connect with STUN configured.
• A STUN Binding Request to outpost.rmstdope.eu:3478/udp got a Binding Success (0x0101).
• Human check for the verifier: two homes on broadband reach Connected; a phone hotspot on one side gets "Couldn't connect your two computers." on both sides within about 10 s.

Deviations from the plan

None. One decision is recorded in the plan: the acceptance says focus goes to Start again "as for today's message", but the old code focused Close. The explicit statement wins.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WCDKf1PWoqB8eBTNnkDuvg

[rmstdope]

Review (cold read, reviewed_head: 26dd587995013c156b9dd38e435a5aa7a4d537c7): reviewed before merge under the producer's review practice, by an agent given the diff and the bead, and not the producer's reasoning.

The reviewer found nothing that blocks a merge. It confirmed that every string matches the acceptance word for word, that focus goes to Start again for both host and joiner, and that both peer connections use outpost's STUN. The connect limit is 10 s on both sides, and the budget fits: two 3 s gathers when outpost is down, plus about 1 s for the five-ping check, against Playwright's 20 s timeout. The address line meets the acceptance: it comes after both cards, it is a p that cannot take focus, and it uses --steel-hi as the mockup does. Each new or changed test would have failed on the old code, and README and architecture.md are updated.

1. web/src/ui/play_together.ts failureAlert fallback, and web/src/netplay/session.ts (no working WebRTC, an offer the browser refuses, an unrecognised liaison error reason): these also end in unreachable, so players are told their networks don't allow a direct connection, which is the wrong cause. This is not a regression, because the old text was just as wrong for these cases. Suggested as a follow-up rather than a change here.
2. web/integration/specs/play-together.integration.spec.ts: the edited header comment has a line of about 170 characters; re-wrap it to about 100 to match the comment around it.

Producer's answers

• 1: No change in this PR. The agreed experience gives this message to "whatever the cause", and adding a separate message for a missing WebRTC or an unknown liaison reply would be a new piece of experience, which UX has to agree first. I'm raising it with the navigator as a possible follow-up bead.
• 2: Fixed by re-wrapping the comment. The commit touches only that comment, so it needs no follow-up review.