forms: draw a schema-stated closed set as a picker, and gate on it

[Yaraslaut]

Closes #386.

schemaJson<A>() describes a glz::enumerated enum class member completely — glaze emits a closed oneOf of const alternatives, each carrying its own title — but DynamicForm read none of it. A three-value set drew a TextField, and ready was true for role = "Emperor".

Two independent halves

resolveProp discarded the alternatives. It collapsed any oneOf/anyOf to its first non-null branch. That path was added for the nullable-$ref shape (#189), where the branches differ only in nullability and collapsing is right. Applied to a closed set, whose branches differ in value, it dropped every alternative but the first — and merged that branch's const onto the field as if the schema had pinned the value.

fields never looked at const/enum. isChoice keyed solely on x-optionsAction, so an enum member fell through to the text field and fieldJsonLiteral returned whatever was typed.

The fix

enumChoices() recognises "every branch bar {"type":"null"} carries a const" as a closed set — distinct from the nullability shape, where one branch has no const and the property falls back, so #189's case still collapses as before. The bare JSON-Schema enum keyword is read the same way. The set draws with the combo box Choice already uses (no options action, no fetch — values and labels are both already in the schema), taking the field_ objectName from the hidden TextField exactly as the boolean's CheckBox does. fieldJsonLiteral refuses an out-of-set value, so the gate means what it says, and resetFields clears a combo by currentIndex because it has no writable text.

Two shipped rungs were rendering free-text boxes for fully enumerated members today: pastebin::CreatePaste's visibility/editability, and bookmarks::CreateBookmark's visibility.

Spec/code disagreement, resolved in the spec's favour

docs/spec/forms/forms.md:819 claimed glaze does not emit oneOf. It does, and that sentence was the stated rationale for treating oneOf as nullability-only — so it was load-bearing, not cosmetic. The spec gains a "Closed sets" section and loses the false claim.

The issue's framing was also slightly too broad: the oneOf-of-consts shape appears only for an enum declaring a glz::meta/glz::enumerate. That was measured directly from a schemaJson<>() dump and is now reflected in both the spec and the code comments. An enum without a glz::meta gets a different, worse treatment — filed separately as #392 rather than folded in here.

Verification

src/qt/forms/tests/tst_DynamicFormEnumChoice.qml, 17 cases. Every schema in it is pasted verbatim from a real schemaJson<>() dump (kanban SetMemberRole, pastebin CreatePaste) rather than hand-written to match the implementation.

Non-vacuity, measured — DynamicForm.qml reverted to its parent commit and rebuilt, new test left in place:

enum PASS: 3   enum FAIL: 16
FAIL!  : ...test_an_enum_renders_a_selection_control_not_a_text_field() 'verify()' returned FALSE
FAIL!  : ...test_a_value_outside_the_set_leaves_the_form_not_ready() Compared values are not the same
FAIL!  : ...test_choosing_an_option_serialises_the_enum_name_glaze_reads() Uncaught exception: Cannot assign to non-existent property "currentIndex"

16 of 19 entries fail; the 3 that pass are initTestCase, cleanupTestCase and one preservation guard. With the fix, 19/19 pass. One case originally passed against the unfixed code — it read const off the field descriptor, which never carried it — so it was rewritten to assert on resolveProp's own output and re-measured. The resetFields guard was reverted alone as well: Uncaught exception: Cannot assign to non-existent property "text".

Regressions: ladder_pastebin_tests 769 assertions, ladder_bookmarks_tests 837, ladder_kanban_tests 1240 — all pass.

Not verified: the HTML renderer (morph::render). This change is confined to the Qt/QML renderer; whether the HTML one has the same gap is untested either way.

On the QML suite total: it reports 250 passed, 6 failed. The 6 are module "Lightweight.Migrations" is not installed, from _deps/lightweight-src/src/tools/dbtool-gui/tests/qml/ — a vendored dependency's own QML tests that the harness scans. They fail identically before this change and are unrelated to it.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!