fix(formdesigner): bind form_uid as UUID, matching the column 004 creates

[Juan2coder]

The bug

packages/parrot-formdesigner/migrations/004_form_uid_uuid_type.sql retypes form_schemas.form_uid and form_data.form_uid to native UUID. The storage layer never followed: it still called str() on every form_uid before binding, each site marked # TASK-2008: form_uid column is VARCHAR(36) until migrated.

The package ships a migration that breaks its own data access. Apply 004 — the migration parrot itself publishes — and asyncpg rejects every bind:

invalid input for query argument $1: '001308f6-...' ('str' object has no attribute 'bytes')

The failure mode is the bad kind. FormRegistry.load_from_storage() catches per row and logs a warning (registry.py:1070), so the app boots looking healthy with an empty registry.

Observed on a real deployment (FieldSync staging, 2026-08-02): 64 forms loaded before applying 004, Loaded 0 forms from storage after.

The fix

The column is UUID, so bind a UUID.

• services/storage.py — _create_table_sql declares form_uid UUID NOT NULL; save / load / delete bind form.form_uid directly. Module docstring updated.
• services/submissions.py — same for the form_data DDL, its ADD COLUMN IF NOT EXISTS, and the insert.
• api/handlers.py — the str -> UUID normalisation in the descriptor merge is kept. It is inert against a UUID column and still covers a deployment that has not applied 004. Only its now-false comment is corrected.

Tests: updated, not deleted

Four tests failed. They were pinning the old contract and were right to fail — so they now pin the new one:

• The DDL assertion becomes form_uid UUID NOT NULL.
• save / load / load-with-version assert isinstance(arg, uuid.UUID) as well as equality. Equality alone would still pass if the code found another way to bind a string.
• test_delete_by_form_uid passed the literal "test-uid-004" — not even a UUID — to a method annotated form_uid: uuid.UUID. It only passed because delete() stringified it and the fake connection type-checks nothing. It now uses a real UUID.

Verification

Package suite, this branch	18 failed / 1790 passed / 80 errors
Same suite, clean dev	18 failed / 1790 passed / 80 errors
Delta	empty — the pre-existing failures are untouched
End to end vs a real 004-migrated Postgres	list_forms() -> 59, load() resolves 59 of 59. Before: 0

Note on compatibility

This is a clean cut: after this change the code requires 004 to have been applied. That matches the direction FEAT-393 already took in Python (FormSchema.form_uid is uuid.UUID), and 004 is already published. A deployment on an unmigrated column needs to run 004 — which was always the intent, since initialize() does not retype.

Worth flagging separately: migrations/README.md documents only steps 001-003. 004, 005 and 006 exist in the directory but appear in no execution order, so anyone following the README lands exactly in the state this PR fixes.

🤖 Generated with Claude Code