Filed from objectui#8058 (dev seat, session session_01611D6ZaRaMmwTNQmSbk8MH), which surfaced this while fixing its own half. ⛔ Unassigned; grading and domain:* are triage's.
What objectui#8058 closed, and the half it did not reach
objectui#8058 taught MetadataFieldsPage's READ DOOR (toDesignerField) to find a lookup's target under either spelling, so a field whose target survives only as the pre-objectui#6041 referenceTo keeps its relationship through an edit-and-save. That fix reaches the designable half only — the fields that go through toDesignerField at all.
Two sites carry the identical defect and have no read door in front of them.
1. MetadataFieldsPage's carried-through half — measured
objectui#8060 split the stored fields map: a field whose stored type this designer cannot author (master_detail, vector, the other 20) is preserved and re-emitted straight from the stored document as carryOver(keep.raw) — which strips referenceTo — with no toDesignerField in the path. So the strip takes the target and assertRelationshipTargetPresent refuses the whole save.
Measured on this branch, with objectui#8058's read-door fix already applied, stored document { parent_id: { type: 'master_detail', label: 'Parent', referenceTo: 'invoice' } }, author relabels an unrelated name field:
PROBE-RESULT puts=[] || error=[MetadataFieldsPage] cannot save the field `parent_id`:
a `master_detail` field needs a `reference` naming the object it links to, and this
one has none. ... Pick the target object, or change the field to a non-relationship
type.
⭐ This half is worse than the one objectui#8058 fixed, and the difference is the repair path. A designable lookup refused this way has a control on screen — the author re-picks the target. A preserved field is rendered read-only on this page by design (objectui#8060), so "Pick the target object" names a control that does not exist here. Every later save of that object is refused from this page and the author has no way out of it from this page.
2. metadata-admin's readFields — the metadataAdminFieldsReadDoor site
packages/app-shell/src/views/metadata-admin/previews/object-fields-io.ts strips the same key on read and its docblock states the same invariant:
Nothing is lost on the way out: where the spec has a spelling for the concept it is a SEPARATE key (reference, system) that is NOT stripped and rides through untouched, which is what lets the designer read it back.
That holds only when the draft carries the spec spelling. A draft holding the target only as referenceTo has it stripped on load with nothing left behind. ⚠️ Not measured by this seat — stated as the shape to measure first, not as a finding.
MetadataService's metadataServiceCarryOver site takes an already-built payload rather than a stored document, so its read path is a third thing to check rather than an assumed instance.
The declared claim this falsifies
packages/types/src/internal/retired-field-keys.ts, the referenceTo tombstone:
A rename: the spec spells the lookup target reference ("Did you mean referenceTo -> reference?"). The strip loses nothing — every write path re-emits the designer's target under reference, and the read door's writers never emit the retired spelling.
objectui#8058 made that true for one site's designable half by adding the read it depends on. It is still written as unconditional, and for site 1 above it is measurably false. Class (b) — violation of a declared contract, the contract being the registry entry's own stated invariant. The same sentence is what a future reader uses to skip the check, which is how objectui#8058 got there.
The shape of the repair, for whoever grades this
⛔ Not "read both spellings in three more places". The registry already knows this key is a pure rename with a specEquivalent. Canonicalizing at the strip — rewrite referenceTo onto reference when the spec key holds nothing, instead of deleting it — would close the class at every site at once and make the tombstone's sentence true as written. That is a decision about the registry's contract, not a local patch, which is why it is filed rather than ridden along on objectui#8058.
Related
objectui#8058 (the designable half) · objectui#6041 (the rename) · objectui#6043 (the formula precedent, and why it does not transfer to this key) · objectui#6527 (the tombstone registry) · objectui#8060 (the preserved half exists because of it) · objectui#7714 (the guard that makes this loud)
Generated by Claude Code
Filed from objectui#8058 (dev seat, session
session_01611D6ZaRaMmwTNQmSbk8MH), which surfaced this while fixing its own half. ⛔ Unassigned; grading anddomain:*are triage's.What objectui#8058 closed, and the half it did not reach
objectui#8058 taught
MetadataFieldsPage's READ DOOR (toDesignerField) to find a lookup's target under either spelling, so a field whose target survives only as the pre-objectui#6041referenceTokeeps its relationship through an edit-and-save. That fix reaches the designable half only — the fields that go throughtoDesignerFieldat all.Two sites carry the identical defect and have no read door in front of them.
1.
MetadataFieldsPage's carried-through half — measuredobjectui#8060 split the stored
fieldsmap: a field whose storedtypethis designer cannot author (master_detail,vector, the other 20) is preserved and re-emitted straight from the stored document ascarryOver(keep.raw)— which stripsreferenceTo— with notoDesignerFieldin the path. So the strip takes the target andassertRelationshipTargetPresentrefuses the whole save.Measured on this branch, with objectui#8058's read-door fix already applied, stored document
{ parent_id: { type: 'master_detail', label: 'Parent', referenceTo: 'invoice' } }, author relabels an unrelatednamefield:⭐ This half is worse than the one objectui#8058 fixed, and the difference is the repair path. A designable
lookuprefused this way has a control on screen — the author re-picks the target. A preserved field is rendered read-only on this page by design (objectui#8060), so "Pick the target object" names a control that does not exist here. Every later save of that object is refused from this page and the author has no way out of it from this page.2. metadata-admin's
readFields— themetadataAdminFieldsReadDoorsitepackages/app-shell/src/views/metadata-admin/previews/object-fields-io.tsstrips the same key on read and its docblock states the same invariant:That holds only when the draft carries the spec spelling. A draft holding the target only as⚠️ Not measured by this seat — stated as the shape to measure first, not as a finding.
referenceTohas it stripped on load with nothing left behind.MetadataService'smetadataServiceCarryOversite takes an already-built payload rather than a stored document, so its read path is a third thing to check rather than an assumed instance.The declared claim this falsifies
packages/types/src/internal/retired-field-keys.ts, thereferenceTotombstone:objectui#8058 made that true for one site's designable half by adding the read it depends on. It is still written as unconditional, and for site 1 above it is measurably false. Class (b) — violation of a declared contract, the contract being the registry entry's own stated invariant. The same sentence is what a future reader uses to skip the check, which is how objectui#8058 got there.
The shape of the repair, for whoever grades this
⛔ Not "read both spellings in three more places". The registry already knows this key is a pure rename with a
specEquivalent. Canonicalizing at the strip — rewritereferenceToontoreferencewhen the spec key holds nothing, instead of deleting it — would close the class at every site at once and make the tombstone's sentence true as written. That is a decision about the registry's contract, not a local patch, which is why it is filed rather than ridden along on objectui#8058.Related
objectui#8058 (the designable half) · objectui#6041 (the rename) · objectui#6043 (the
formulaprecedent, and why it does not transfer to this key) · objectui#6527 (the tombstone registry) · objectui#8060 (the preserved half exists because of it) · objectui#7714 (the guard that makes this loud)Generated by Claude Code