…hing

Done-criterion 1 of the SDK maturity program asked for this and phase 0.4
scheduled it; neither happened, and I closed the program out without
checking. bomly-cli#454 recorded it.

The sentence said detectors.EnsureNode deliberately merges nothing. It was
stale in two ways and contradicted its own paragraph even when written --
twenty words earlier the same passage says scope, relationship and
locations union everywhere, and something that unions scope is merging.

The helper is detectorkit.EnsureNode in the SDK now, wrapping
Graph.InsertNode, which folds by identity. detectors.EnsureNode does not
exist in this repository at all, so a reader following the ADR looked for
something that was not there.

Amended rather than rewritten, in the convention the document already
uses for its ADR-0041 correction. The claim this ADR actually rests on
survives intact and is now stated plainly: folding unions usage facts,
because two records at one identity are two witnesses of one use, and it
still never reconciles contradicting assertions -- records that disagree
about origin stay distinct nodes and no tiebreak picks a winner.

Closes #454

Co-Authored-By: Claude Opus 5 <[email protected]>

Codex found that my correction repeated the defect it was fixing. I
carried the "contradicting records stay distinct nodes" sentence forward
from the original text and cited preserveContradictingOccurrences without
checking that either still held. Neither does.

That mechanism was deleted with ADR-0041 -- along with foldWitness,
resolutionKey, originKey and renameNode -- and enrichment.go says so at
the site it used to live. Identity is minted from the canonical package
URL now, so two records of one package have one identity and cannot be
given two.

Verified rather than reasoned: inserting two nodes of one identity with
different artifact URLs yields one node carrying both origins.

The principle survives and the amendment now says which part. No tiebreak
ever picks a winner over a contradiction -- that holds more directly than
before, because the disagreement is published as two origins on one node
rather than split into two nodes and reconciled afterwards. What moved is
enforcement: from splitting identity to accumulating assertions on one
identity.

Co-Authored-By: Claude Opus 5 <[email protected]>

CodeRabbit caught the amendment contradicting itself one sentence apart:
it claimed nothing is discarded and then described MergeOrigins dropping
what cannot be published. That is the same self-contradiction the original
paragraph had, which is what this amendment exists to fix.

MergeOrigins drops three things, and naming them is more useful than
hedging: an entry that fails validation and cannot be published, an exact
duplicate keyed on normalized form rather than input bytes, and a
superseded entry.

Supersession is the one worth stating precisely, because it is the only
one that touches a disagreement. It is narrow by design -- same
repository, same artifact URL, superseded entry has no revision, survivor
does -- so it collapses a floating reference into the pinned revision a
lockfile later supplied. Two genuinely different places both survive. The
SDK's own comment names why: that disagreement is the shape of a
dependency-confusion signal, not noise to tidy away.

So the claim narrows to the one that is true and is the point: it never
chooses between two origins that disagree.

Co-Authored-By: Claude Opus 5 <[email protected]>