Four days of missing blog posts are back, and won't go quiet againFour days of missing blog posts are back, and won't go quiet again
September 2, 2026 · A tracking bug silently blocked four days of posts here; today it was found, fixed twice over, and made to text me if it happens again.This blog silently skipped four days of posts due to a small bug. Today it caught up, fixed the bug two ways, and will text me if it happens again.
The blog notices when it can't publish
Four days of posts here went missing (2026-08-30 through 09-02), and every scheduled run reported success. Today found the cause and closed it from three directions, and all four missing days published today as a batch.
The actual fault was content/watch-room.json, the file behind the /watch page. The server rewrites it on every playback update, so as a tracked file it was dirty within seconds of anyone opening that page. The publish job refuses to write a post while anything under content/ or public/ is uncommitted. That is a sound guard, since publishing over a dirty checkout can silently discard a post that was just written. The file being tracked at all was the real bug. It still lives on disk, since the server reads it at startup, but git no longer has an opinion about its contents, so it can't trip that guard again.
Tracing the guard turned up a second bug. A blocked day is supposed to stay due so a later run retries it. Instead the code stamped every attempt as done, published or not, so each blocked day quietly dropped off the schedule instead of coming back. Now it stays due until it actually publishes.
Besides that, the failure mode used to be invisible. A blocked run said so only in a log file nobody reads. Now it sends a Telegram message naming the day and the files at fault, deduplicated so one ongoing blockage doesn't turn into dozens of pings.
A tournament bracket now shows the whole finish, not just the winner
Reckreation's bracket page used to draw one thing: a champion name, or nothing at all for a round robin that never crowns one. The engine now computes a full ordered finish for a settled bracket (every place the bracket actually declares, with entrants who tied for a place sharing one row instead of one being arbitrarily printed above the other), and champion is just whoever holds first place alone. The bracket page was switched from a single champion prop to the full finish list, so there's one place the answer lives rather than two copies that could quietly disagree. A drawing that settles no order at all, like an unfinished round robin, still draws no block, rather than a heading over nothing.
This blog is supposed to write itself a post every day, from that day's work. For four days, August 30th through September 2nd, it didn't, and it never said why; every run just quietly reported success. Today's work tracked down the cause and fixed it, and all four missing days were published today.
The cause was a small technical accident. A different part of this website, a shared page for watching videos together, constantly rewrites one of its own files. That file got mixed into the same set the blog checks before publishing, so the check saw unsaved changes every single time and refused to write. The fix was to stop the blog from watching that file at all.
Two more problems turned up along the way. First, when the blog refused to publish a day, it was marking that day as done anyway instead of leaving it for the next run to retry, so those days weren't just delayed, they were skipped for good. That's fixed: a blocked day now keeps trying until it actually gets published. Second, a blocked day used to fail silently, recorded only in a log nobody was reading. Now it sends me a phone message naming which day got blocked and why.
Separately, a tournament bracket tool got an upgrade. Instead of only ever showing who won (or nothing, if the tournament wasn't finished), it now shows the full standings (first place, second, and so on, including ties), pulled from one single source, so the winner it shows can't quietly disagree with the standings it shows.