Release 1.23.1

Two release-watch fixes found by the release-lifecycle acceptance scenarios (D46, roadmap WS-P task 16; OpenVibe.Host scripts/release-acceptance.js). release-watch.js stays within its 5 KB brotli budget (4.99 KB; comments shortened).

• An account switch during an update (scenario 9). An in-place update fetches its regions with the session (credentials: 'same-origin'), so a region can be rendered for whoever is signed in. On sites that switch accounts in place (Live: the openvibe-auth-changed window event), a region fetched for the previous account and still waiting for focus to leave it, or still in flight when the account changed, was committed as fetched: the new account (or a guest, after signing out) saw what the server had rendered for the previous one. Now openvibe-auth-changed drops such regions and fetches them again with the new session, and an answer that arrives after a switch is never committed (counted as deferred: account, a new reason /release-metrics accepts). The switch itself still reloads nothing, prompts nothing and opens no second stream.
• Resuming an offline tab (scenario 14). Back online, the release stream reconnected at once only when it had given up (6 failures); a tab that came back while still backing off waited out the backoff (up to 8 minutes) before hearing release notifications again. online now reconnects at once from the cursor in both cases, like notification-live.js does.

New tests: test/release-account-switch.test.js and test/release-offline-resume.test.js (offline: no prompt, no reload, nothing counted as failed; back online: one /release.json read straight to the newest release, in place; a connection lost mid-update leaves the page whole and prompts once; past the window, never reloaded under a focused field; a hidden tab resumes its stream when shown). And test/release-reading.test.js (scenario 4, a feed update while reading: a region above the viewport that grows or shrinks moves the page by exactly that, the region keeps its own scroll, a region in view is not compensated; nothing to fix). All three print [metric] name=value lines the acceptance runner reads.