Passwords could be guessed on two addresses that need no sign-in. The sign-in brake named three doors where a password is checked. There was a fourth: resending the confirmation mail and correcting an unconfirmed address. Measured: 100 wrong passwords, 100 times 401, not a single 429, and the answer told the caller when a guess was right. Anyone able to reach Nexview from the network could try passwords for every account, the operator's included, without holding one of their own.
Access keys survived "sign out everywhere". They take a second path through the sign-in that never asked whether the session was still valid. Somebody whose laptop was stolen locked the thief out of sessions but not out of keys. "Sign out everywhere" now revokes them and says so beforehand; a routine password change leaves them alone and points at where to revoke.
A slow Jellyfin or Emby no longer loses the whole library sync. Every request to a media server had fifteen seconds, whatever it asked for. A library big enough, or a server slow enough, missed that, and because one missed page takes the entire run with it, the card read "Not synced yet." rather than "half done". Reported for Jellyfin while Emby on the same installation worked; both run the same code, the Jellyfin server was simply the slower of the two. Reported by @ldoctoru in #7. The two screenshots in that report are what made it findable: the card saying "Not synced yet." and a log line ending in 502 in 15011ms: three attempts, three times the same figure, which is our own clock rather than the server's. The time limit is no longer a fixed number. Nexview now asks in chunks whose size the server decides: a chunk that misses its time is halved and the same place asked again, and what gets through is kept for the rest of the run. A library twice the size needs twice as many chunks, but each one fits. Nothing to configure. Timeouts also reach the log now, naming the request that died. Media servers were the only outward connection that stayed silent about it, so the report that started this could not be traced from the logs.
A silent Radarr deleted the whole storage accounting. No answer meant no items, and everything the instance would have reported counted as gone. Measured: usage fell from 28 GB to 20, the house from 2 GB to 0, and because the rows were deleted, deletion deadlines and releases were lost for good. The rule from the status sync now holds here too: not answering is not a no.
Cleaning up several quality profiles at once forgot the ownership record for all of them, including the ones the instance refused to delete. Those stayed in Radarr but counted as foreign here.
Sixteen error messages reached everybody in German, whatever language they had chosen, among them the very first contact with Nexview: an expired invitation link. All of them are named now, so the interface can say them in the reader's language; the German sentence stays as the fallback. Media-server timeouts join them in this release.
An approver was told "new feedback" and then could not read it. The bell led to an address that required an administrator, so the page said "there is no unanswered feedback". Reading is now open to approvers; answering is not.
Three pages showed their empty text when loading had failed: your own requests, favourites, and a child's wishes. Somebody with twelve running requests read "you have not requested anything yet".
The bell sent two kinds of notification to the wrong page. Where a click leads was answered in two places that disagreed; there is now one list, and it covers all 26 kinds with no fallback.
A crash while drawing wiped the whole interface, navigation included. The likeliest trigger is an update with a tab left open. There is a net now, with one visible way out.
The heart and the bin swallowed every failed click. They report it themselves, where the eye already is, and a screen reader says it too. Deleting from the cleanup list failed silently for the same reason.
A child was shown TMDB's technical text. The reason goes to the log instead.
A link to a conditional settings sub-tab always landed on "General". On the first render the configuration is not loaded yet, so every sub-tab with a condition dropped out of the visible list and the fallback jumped away before the configuration arrived. ?unter=radarr kept working, which is what hid the cause.
"Library touched" read "0 %" above "2 of 0 titles". Without a stock the share is not zero, it is unknown, and a dash says that where a zero asserts something.
The health check in docker-compose.yml ignored NEXVIEW_PORT, so following our own instructions produced a container that runs and counts as unhealthy. ESLint checked generated output and therefore always failed.
Under the hood
Backups are now checked by what lives in the data directory, not by a list of parts somebody remembered to name. data/trash/ could have gone missing unnoticed: it had not been forgotten, it had never been asked about. Every name is now either archived or excluded with a written reason.
Five new guards, each proven against a mutation: every literal t('a.b') exists in both languages (2,204 keys), no new error answer without a code, all 26 notification kinds know where their click leads, the v1 promise covers nested fields (252 instead of the top level), and a media server that only answers to small requests still delivers every title.
Thanks
@ldoctoru for reporting #7, with screenshots and a log excerpt, which is what turned "Jellyfin is not syncing" into something that could be measured. If you run into something, please open an issue; that is how this list gets shorter.
