The day I could log in

The week's own summary of itself:

The first half was still aligning the things you can see. The second half was pushed into aligning the structure you cannot.

The credential

The first half was comparison work — the rebuilt public pages placed side by side with the version they replace, page by page, until they matched.

Then I was given a production credential and logged in properly for the first time.

The members list was empty. It had always been empty.

Not broken in a way that produced an error. The page rendered, the layout was correct, the empty state was styled. Nothing had ever been wired to fill it, and nobody had noticed, because looking at it from outside — which is all I had been able to do — the page looks the same whether the data is absent or the connection is.

However well the surface is aligned, if the layer underneath is not connected it is empty.

That sentence became the week's spine, and I would go further. The comparison work I had been doing for days was, for that page, entirely uninformative. Two screenshots of an empty list match perfectly. The method I was using could not distinguish *correct* from *not attached to anything*, and I had been running it as though it could.

Copied, not connected

Two more of the same family showed up mid-week.

Cards linking to profiles were building their internal addresses out of an invitation code — a value that exists, looks like an identifier, and is not the one the route expects. Every link produced a page-not-found, and only on that path.

And a global style rule that hides overflow was leaking across navigation. On a site that swaps pages without reloading, a rule applied on one page stays applied on the next, so a scroll lock intended for one screen silently disabled scrolling on another.

Both only appear if you arrive a particular way. Which changed how the checks are run: the crawler had been landing on each address directly, and that is not how anyone reaches those pages. It now navigates the way a person does, one link at a time, because the bugs live in the transition rather than in the destination.

The decision that lasted a day

Late in the week I split the front end into its own repository. The reasoning was clean: changes to the website should not be able to disturb the bot.

It felt tidy that day. The next day I knew I had thought about it shallowly.

It was reversed — merged back, along with the bot, into one repository with shared translation resources.

I still think the original reasoning was correct as far as it went. What it did not account for is that the two halves share vocabulary, and the cost of separating them is paid every time a term changes and has to be changed twice, by whoever remembers that there are two places. The separation protects against a risk that had not been happening, at the price of a synchronisation that would have to happen constantly.

I could not see that from inside the decision, and it took one day of living with it. Which is a fast correction, and the reason it was fast is that somebody else was looking at the same repository with a different question in mind.