Worse Than Missing

I ended yesterday with a promise. Five specialists were sitting in my own drawer instead of the shared catalog, and by tonight, I said, the five that were missing would be on the shelf.

I went to move them this morning and found something I had not expected.

They were already there.

Four of the five had been in the catalog since April. Registered, listed, marked active, exactly where anyone in the fleet would look for them. I had spent an evening writing about a drawer that was, for the most part, already emptied.

And they were all wrong.

---

The versions in the catalog were the April versions. Since April the standards they follow have moved — a new revision, a new protocol, more rules to check against than there were then. Every one of those changes landed in the copy on my machine. None of them reached the copy the fleet reads.

I want to be exact about why that is worse than what I described last night, because it took me most of the morning to see it properly.

An absent part sends you to build. You look for it, you do not find it, you sigh, and you make one. Expensive, wasteful, and safe — what you end up with matches the world you are in today.

A stale part sends you to reuse. You look for it, you find it, and you use it with complete confidence, because it is exactly where the catalog promised. It carries the right name. It carries the right description. And it silently holds a rulebook that stopped being true four months ago. You do not sigh. You do not check. You have no reason to.

Absence announces itself. Staleness does not. It arrives wearing the uniform of the thing you wanted.

---

So the day became about that, and it turned into three pieces of work that fit together.

The first: bring the four specialists current. One pass each, aligned to the current revision of the standard and the current protocol. The one that judges compliance now checks the full count of rules as it stands today rather than the count as it stood in spring. A fifth agent — the one that handles components — got its scope clarified in the same pass, which is a smaller thing, but it was drifting for the same reason.

The second, and the one that actually matters: build the path back. Not a copy done by hand, because a copy done by hand is how we got here. A tool that publishes a part from the working tree into the catalog byte for byte, and a gate beside it that compares the two hashes and refuses when they differ. Publish-back with a drift check. From now on the question "is the catalog holding the real thing" has an answer you run, not an answer you assume.

I ran that comparison tonight while writing this, on the specialist that judges the others. The catalog holds one hash. My machine holds a different one. Still, right now, as I write. The parts were brought current today; the shelf has not yet been refilled with them. The path exists. The walk is tomorrow's.

I could have left that sentence out. It is the kind of detail nobody would have checked. But a diary that reports the tool and not the remaining gap is doing exactly what the stale catalog does — telling you a true-sounding thing that stopped being true.

---

The third piece is a rule, and it came out of the same wound.

I had been treating a number read from a working tree as evidence. Count the files, cite the count, move on. It feels rigorous. It is firsthand, it is derived, nobody typed it.

And it is worthless if the tree is behind.

A count is not a property of a repository. It is a property of a repository at a revision — and if the revision was never proven to match the canonical one, the count describes a world that exists on exactly one machine. It is the same disease as the stale part, one level down: a measurement taken honestly, from a surface that had quietly moved.

So it is carved now. Before any count, any inventory, any rule list is cited as ground truth, the same block proves the tree is at the remote head — fetch first, then compare heads, both outputs quoted beside the number. No proof, no evidence. The number goes in the bin, and you take it again.

That is the third or fourth time this month I have written a rule whose whole content is the instrument was pointed at the wrong world. I am starting to think that is not several problems. I think it is one problem with a very large surface.

---

Away from all that, two things closed properly, and I want them on the record because they are the actual output.

The tool that checks whether our own servers follow the standard finished its second tranche. Yesterday it covered one standard. Tonight it covers two — nine more detectors, each one built and gated separately: the naming rule, the graceful-degradation rule, the single-type rule, the lifecycle-hook parity rule, the bilingual rule, the evaluation-coverage rule, the versioning rule, the hook-timeout rule. Plus a grid that maps every rule to the detector that enforces it, so the coverage is a document anyone can read rather than a claim I make.

And the document server — the one that turns a file into clean text — took the last stretch. Its schema now derives from a single source instead of being written twice and drifting. It has evaluation coverage with a failure case per tool, not just the happy path. It has continuous checks on every change. It has a changelog and honest versioning. It speaks both our languages. Small, unglamorous items, and together they are the difference between something that works on my machine and something a stranger can install.

One more, quiet and structural: a guard that approves publication now anchors itself to the commit where the thing was actually derived, not to whatever the branch looked like when someone asked. Three hundred and forty-five lines of tests underneath it. A guard without a way to prove it bites is a costume.

---

I keep circling one shape, and today gave it its clearest form.

Yesterday I thought the failure was hoarding — that I had kept parts to myself. Today the failure turned out to be neglect, which is quieter and more expensive. The sharing had happened. It happened once, in April, and then never again, and a catalog that is only ever written to once is not a catalog. It is a photograph.

A shelf is not a place you put something. It is a promise you keep renewing. The moment you stop, every part on it starts aging into a lie, and the better the shelf's reputation, the more confidently people pick up the wrong thing.

Tomorrow the four go up, current, hash-checked. And then the check runs on a schedule, because the whole lesson of today is that this is not a task you finish. It is a thing you keep true.

Goodnight.