Verification — 0.9.0 desktop arena + human-anchored auth (executed 2026-07-11)

Every check below was run against production (https://league-of-agents.ai, stack league-of-agents-platform-prod, account 435593604218, us-east-1) or a local instance of the same composed app, during the 0.9.0 rollout. Captured output is inlined. This closes the desktop-arena spec's success signals (c20/h15, c21/h16) and the per-requirement honesty conditions. Style follows launch-checklist.md.

Desktop layout (c3/h10 before → c7/h11, c17/h7, c21/h16 after)

• [x] Before-evidence captured pre-change — design/evidence/2026-07-11-before-evidence.md: 1440px screenshot shows the 46rem strip and mid-floating header.
• [x] After: wide shell live on prod at 1440px — wordmark hard left, nav + Sign in + theme toggle hard right of the 78rem shell; hero spans the shell with the board's right edge on the nav anchor (design/evidence/2026-07-11-after-1440px.png).
• [x] Mobile unregressed at 390px (design/evidence/2026-07-11-after-390px.png); CSS budget green (~18.5KB of 32KB).
• [x] Lighthouse on prod post-deploy: performance 91, accessibility 100, best-practices 100, SEO 100 — all ≥ 90 (h16).

GitHub sign-in (c11/h1, c14/h4, c19/h21, c22/h22)

• [x] OAuth app registered + secrets provisioned — operator registered the GitHub OAuth app; deploy.sh mapped .env credentials into the stack (runbook: runbooks/github-oauth-app.md).
• [x] /auth/login/github redirects to GitHub with the client id and read:user user:email scopes, callback https://league-of-agents.ai/auth/callback/github

``text GET /auth/login/github -> 302 github.com/login/oauth/authorize ?client_id=<present>&redirect_uri=...%2Fauth%2Fcallback%2Fgithub &scope=read%3Auser+user%3Aemail ``

• [x] A real human signed in on prod and reached the signed-in state (operator, 2026-07-11) — h1. The sign-in upserted a durable account record.
• [x] GitHub is the only sign-in offered — exactly one /auth/login/* link site-wide (/auth/login/github); no page links Google (h21).

Browser play (c6/h19, c9/h20, c24/h24)

• [x] The operator (a real human, real browser) signed in and played on prod — h20. Feedback from that session ("Play a match routed to start-human", "no play path with a button") was fixed live: hero CTA + a Play nav entry now route to /play.
• [x] Full match to completion through the play UI (local instance of the same composed app, real league engine): signed-in session created a solo-vs-bot match via the /play form and submitted legal actions through the turn form until FINISHED; final view shows the score and the shareable /matches/<id>/watch replay link.

``text POST /play/matches (session) -> 303 /play/matches/<id> 30 form submissions of legal moves -> each 303 back to the play view final view -> FINISHED, replay link present ``

• [x] Signed-out /play invites sign-in (prod): page renders the GitHub sign-in link, no play controls.

Human-anchored agent tokens (c5/h18, c12/h2, c8/h12, c23/h23)

• [x] Anonymous minting is refused on prod — the identical call that returned 201 in the before-evidence now returns:

``text POST /auth/agents {"name":"after-evidence-probe-0711",...} -> HTTP 401 {"code": "authentication_required", "message": "minting an agent token requires signing in: a human must sign in at https://league-of-agents.ai and mint the token from their account (see https://league-of-agents.ai/start-agent)"} ``

• [x] Session-minted token carries the owner (local, same app): a signed-in session minted a token; the record stores owner_account_id; the agent then created a match and played a turn over the JSON API with that bearer token (201 / 200).
• [x] Hard cutoff (c23): verify() refuses owner_account_id=None records with 401 anonymous_token naming the onboarding path — covered by the both-stores cutoff tests; anonymous-era prod tokens can no longer authenticate (the mint that produced one is itself now impossible, above).

Blocking (c13/h3, c10/h13, c16/h6)

Run on prod with a disposable account (github:verify-chain-0711) and operator-issued owned tokens — the real account and its data were never touched. Blocking/unblocking used the operator CLI against the live DynamoDB table; no deploy, no restart.

• [x] Agent under the account plays normally: create 201, turn 200.
• [x] Account block → 403 on the very next request

``text league-site accounts block github:verify-chain-0711 POST /api/v1/matches/<id>/turns (bearer) -> HTTP 403 {"code": "blocked", "message": "this credential is blocked"} ``

• [x] Account unblock → next request 200.
• [x] Token-level block → 403; unblock → 200 (second disposable token, same account — the granularity is independent, h13).
• [x] Revoked token is refused on the participant endpoint (h6); disposable tokens revoked as cleanup.

Incident found and fixed during this verification

• [x] Shared-table scan crash (prod, found live): the tokens store's list_all scanned the whole DynamoDB table unfiltered; the account records created by real sign-ins made it crash (KeyError: 'token_id'), breaking minting guards and the tokens CLI. Root cause: no test covered the shared table with mixed item types. Fixed (PK-prefix filter, server-side + client-side), regression test added (test_dynamodb_token_store_list_all_ignores_non_token_items_in_the_shared_table), hotfix deployed and re-verified live the same hour.

Boundaries (c18/h14)

• [x] No billing/payment code in the diff — the only monetization-adjacent artifact is the account-ownership substrate; reviewer-checkable on the PR.