What approves a connector

crates/connect/README.md said, for eleven phases, that a connector appearing in the registry without a decision record fails a test. It did not. The test it named — registry::tests::every_connector_approved_in_the_ontology_is_present — compared CONNECTORS against a hand-written array of 21 keys sitting six lines above it, with a comment attributing each key to the record that approved it.

A hand-written array cannot check the property the sentence claimed, and the gap was not hypothetical. Four connectors — dew-scholarship-reports, dew-school-improvement, ohio-auditor and lsc-catalog — were in the array, in the registry, and in no decision record that names them. The array had been extended alongside each one, so every addition passed.

Issue #129 found the sentence and pull request #137 corrected it — it now says the test cannot do this, and names the four. That was the honest move and it was not the fix. A README that documents its own gap still leaves the gap, and the sentence that was there first was the one worth making true.

Prose mention cannot stand in for approval, and the distance is not small. tax-abstract is named in seven of the 29 records here and approved by one; dew-foundation in seven, approved by one. A record cites the connectors it reasons about far more often than it declares one. Any test reading these files as text would have to treat a citation as an approval, which would pass the four this is meant to catch and would also make the-channel-with-no-line — a record about tax-casino that discusses tax-abstract and ohio-bills at length — look like the approving record for three.

Declare it, in the record, in a field. A decision record that approves one or more connectors carries a connectors: sequence of registry keys. Thirteen records carry it and they account for all 21 connectors, each named exactly once.

The field is data and not prose, and it is chosen for the reason Revision in web/src/lib/corpus.ts is structured rather than found in prose: the failure mode is asymmetric. A record that forgets the field fails the test loudly; a record whose prose happens to mention a key stays silent, which is the direction that has already gone wrong once.

It costs the site nothing. readDecision reads fields off DECISION_SECTIONS, an allowlist, so an unlisted field is ignored rather than rendered as an unheaded card — and ontology has carried a non-prose corpus_depth since genesis, so this is a shape these files already have.

The test derives the list and checks it both ways. Every connector in the registry is declared by some record, every declared connector is in the registry, and no connector is declared twice. The third is not redundant: without it, two records claiming the same key would let a dropped connector pass on the count.

And the four get records. “Scholarship reports connector”, “School improvement connector”, “Auditor reports connector” and “Catalog connector”, each written from the commit that wired the connector and each opening by saying it was written after the fact. A retroactive record dated as though it were contemporaneous would be a worse defect than the one being closed.

The README’s original sentence is true now, and it is the test that makes it so. A twenty-second connector added without a record fails cargo test -p connect on the line that names it, before review.

Four records exist that record no deliberation. They reconstruct a decision from the commit that made it, and they are weaker documents than the nine written at the time — there is no rejected alternative in them that was actually rejected in front of anyone. Each says so in its own first paragraph. This is the cost of the fix and it is not recoverable: the reasoning that was not written down at the time is gone, and what can be salvaged is the reasoning the commit message preserved.

The registry comment shrinks from 34 lines to a reference. The attribution it carried — which record approved which key — moves into the records themselves, where it can be wrong in a way something notices.

Scan the records’ prose for each key. Rejected on the counts above: seven mentions to one approval for two different connectors, and no rule over text distinguishes them. A test that passes for the wrong reason is worse than the array, because the array at least did not claim to have checked.

Put the approval list in one file — a manifest of key to record. Rejected: it is the array with a different extension. The property worth holding is that the approval lives in the document that made it, so that adding a connector and writing its record are the same act. A separate manifest is a third place to forget.

Delete the claim instead, and let the array be an array. Considered, and it is what #137 did as far as it went. Rejected here because the claim describes a discipline this repository actually wants and has mostly kept — 17 of 21 connectors did have a record, written before the code. The four are the exception, and the cheaper repair of the sentence would have made the exception permanent.

Require a record before a connector may be Wired rather than before it may exist. Rejected: it would exempt exactly the connectors most in need of a stated reason. ofcc-projects has been Declared since genesis behind a policy blocker, and the argument for keeping it declared is the most contested thing in proposals.