fix(ui-react): stop editing an existing model input from failing at save (#91)

[mosoriob]

Fixes #91. Found on #76, walking the signed-in form against TACC's live catalog.

Fault 1 — the save hard-fails (blocker)

Editing any existing input failed:

expected one of the values ['_PLACEHOLDER'] for type
'modelcatalog_configuration_input_update_column', but found 'is_optional'

Role user holds insert/select/delete on modelcatalog_configuration_input and no
update permission (graphql_engine/metadata/tables.yaml:3088). Under that role Hasura
emits the update-column enum with _PLACEHOLDER alone, so
on_conflict.update_columns: [is_optional] is rejected before it runs.

InsertConfigurationInputJunction is replaced by SetConfigurationInputOptional —
delete_by_pk + insert_one in one document, so both root fields run in a single
Hasura transaction and the row is never left missing. This is the same delete-then-insert
the API layer already uses for junction updates.

The mutation is now called only when is_optional actually changed. The old code fired
it for every row in toUpdate, so editing a label, a format, a description — or the
standard variable and unit, which is the SOW workflow — was enough to fail.

Not TACC-specific: it came from this repo's metadata, so it reproduced everywhere.
No metadata change is needed, so nothing has to be applied to TACC's Hasura.

Why codegen never caught it. The committed SDL snapshot is the admin schema, where
is_optional is an updatable column. Codegen and the offline tests validate against that,
never against the user role the app actually runs as.

Fault 2 — a successful save rendered stale (minor)

ConfigurationDetail reads the same GetConfiguration cache entry the form does, and no
mutation wrote to the cache, so a successful save looked as though nothing happened until a
manual reload. The form now refetches before onSaved. A failed refresh is swallowed — it
must not report a successful save as failed.

Verification

• 760 tests pass; lint 0 errors; npm run build clean; Prettier clean.
• Both new assertions were checked red against the pre-fix code: making the junction
  write unconditional fails the label-edit test, and removing the refetch fails the
  refetch test.
• Types regenerated with codegen against the committed SDL snapshot. The graphql.ts diff
  is +21/-15, confined to the renamed mutation.

Still unverified: the new document has not been run under a real Tapis token. Both root
fields are individually proven under the user role (#76 wrote and deleted junction rows),
but the composite mutation needs one live save at localhost:3000 signed in as
mosorio@portals to close #91.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
monorepo	Ready	Preview	Aug 8, 2026 11:56pm

[mosoriob]

Verified live against TACC, signed in as mosorio@portals. Full detail on #91.

The core claim was settled with a zero-write probe — Hasura validates before it executes, so
sending each document with a bogus id separates the two outcomes cleanly:

document	result under the user role
old, update_columns: [is_optional]	validation-failed — the exact _PLACEHOLDER error
new, delete_by_pk + insert_one	constraint-violation (foreign key, from the fake id)

So the new document is accepted where the old one never could be. Nothing was written.

End-to-end through the form on simpleModelAnnotated, using a throwaway input rather than the
client's existing row:

• Add → saved, appeared with no manual reload (Fault 2 fixed).
• Edit the label → saved clean. Operations sent: UpdateConfiguration,
  UpdateDatasetSpecification, GetConfiguration — no junction mutation, since is_optional
  did not change. This is the save that used to hard-fail.
• Toggle Optional → SetConfigurationInputOptional ran with no errors and is_optional: true
  was confirmed in the database.
• Cleanup → both rows deleted; the catalog is back to its exact prior state.

CI note: the first Tests run failed on ConfigurationForm.test.tsx:236 (Configuration name is required), a pre-existing flake in the async validation assertion — not a test from this PR, and
green in the other workflow's Tests job. Rerun.