Edit nested object/array properties in the namespaced property forms (#2027)

[mstomar125]

Summary

Follow-up to #2009. The recursive nested (object/array) property editor only reached GenericSchemaForm, which persists flat at properties[<field>]. The other two forms share the same field dispatch but write namespaced, so the flat modal couldn't be dropped in — it would have written nested edits to the wrong place, which is exactly why #2009 deferred them.

What changed

The modal is now store-agnostic. nested-property-modal.svelte no longer imports the store; it takes the current value and an onPersist callback from whichever form owns it. Each form supplies its own reader/writer, so all three share one editor and one modal while persisting to their own namespace:

Form	Persists to
GenericSchemaForm	properties[<field>] (flat — unchanged)
TaskSchemaForm	properties.task[<field>]
SchemaPropertyForm	properties[<nodeType>][<field>] (through its existing flat→nested migration, titleTemplate recompute and replaceProperties: true)

Both property forms now render leaves through the shared SchemaFieldLeaf instead of their own copies of the five-case dispatch. That removes both fall-throughs in one step — the "Boolean field (not yet implemented)" stub and "Unknown field type" — and is most of the ~360 deleted lines. Only the assignee combobox stays hand-rolled, having no shared equivalent.

Two behaviors were preserved deliberately: the enum default fallback (via a leafValue helper that resolves through the existing getEnumValue), and close-the-popover-on-date-pick, which moved into the shared leaf — so it now also fixes the generic form and the nested editor, where the calendar previously stayed open.

The summary button is shared too (nested-field-trigger.svelte), so a nested field looks identical in every form.

Tests

property-form-field-dispatch.test.ts is the crux: it renders each real form, clicks through to the modal, edits, and asserts the exact persisted properties bag — proving each namespace and that sibling keys survive ({ task: { estimate: 3, address: { street: '1 Main' } } }, { invoice: { total: 10, address: … } }, flat for generic), plus boolean rendering/persistence. Also new: 3 modal-contract tests and 3 nestedFieldSummary unit tests. #2009's tests stay green.

Full suite 4568 passed; lint + svelte-check clean.

Closes #2027.

🤖 Generated with Claude Code

[mstomar125]

APPROVE (self-review, mstomar125)

Adversarial review focused on the risk in a diff that deletes ~360 lines: silently dropped behavior. It went through the old and new versions of both property forms field by field and confirmed the substitutions are faithful — date parse/YYYY-MM-DD storage byte-identical, parseFloat || 0, placeholders, the enum default via leafValue, titleTemplate recompute, replaceProperties: true, the flat→nested migration, and the assignee combobox all intact; parsePropertyValue and the comboboxOpen pre-init loop genuinely dead once the leaf owns its popover state. Namespaces verified on both read and write for all three forms, and sibling survival is structural (deepMergeProperties + a complete namespace per caller), not incidental.

It found one major, fixed in c3d2351a: getPropertyValue documented a flat fallback it never implemented. That was harmless for a leaf (a blank input the user retypes) but destructive for a nested object — an un-migrated node opened an empty editor, and the first edit persisted that empty object over the real one under replaceProperties, taking keys the user was never shown. Now reads the flat shape, matching the docblock and the bridge the task form already had.

Also addressed: the empty-enum guard this form used to carry is restored at the call site (not in the shared leaf, since the other two forms never had it and may legitimately clear a value), and the shared enum trigger humanizes an undeclared stored value so the control stops disagreeing with the collapsed header that already humanized it.

Three new tests cover the migration path — real key count instead of an empty editor, a never-displayed sibling surviving the migrating write, and replaceProperties asserted rather than assumed. I verified all three fail against the previous read path and pass with the fix, so they genuinely lock the behavior in.

Known gap, non-blocking: no leaf-level test pins the date/number/enum onChange payloads directly — the equivalence there was established by reading old vs new rather than by assertion.

Full suite 201 files / 4571 tests; lint + svelte-check clean; gate green on both pushes.