# @platform-modules/fields

## 0.1.0

### Minor Changes

- 1fd8651: Make the D1 atomicity capability truthful and stop advertising an interactive transaction the driver does not have.

  `@platform-modules/db` — `createD1Client` no longer exposes a `transaction` member or `TransactionIdentity`, and is no longer assignable to `TransactionalDatabase`. Cloudflare D1 has no interactive transaction callback API, so the previous shape was a false claim that failed at runtime with `Failed query: begin`; D1 gets a typed atomic batch seam over exactly one real `binding.batch(...)` call instead. Genuinely transactional adapters (Postgres) keep the callback-minted transaction, exact identity semantics and rollback.

  The transaction handle splits in two: `Transaction<S>` is now the dialect-neutral mutation handle (`execute()` only), while the Postgres query-builder surface (`select`/`insert`/`update`/`delete`) lives on `PostgresTransaction<S>`. Callers that use query builders inside a transaction must type against `PostgresTransaction<S>`.

  `withTransactionIdentity` becomes public API. It was marked internal, yet every package's real-Postgres harness needs it to produce a value satisfying `TransactionalDatabase`, and three packages had resorted to importing it through a relative path into `db/src` — which breaks the moment these packages are consumed as published `dist`. Consumers that build their own driver client now have a supported way to brand it.

  Every downstream consumer is migrated: transaction parameters re-annotated, PG test harnesses branded through the public helper or rebuilt on the first-party adapters, and `PgliteTransactionalDatabase` now declares the `$client` the driver actually attaches.

  Two latent defects surfaced and are fixed, both the same shape — code reading one result shape through a cast that hid the mismatch from the compiler:

  - `@platform-modules/content` — `taxonomy.ts` read `result.rows` behind an `as { rows?: unknown[] }` cast, so `isInSubtree` always returned false and the depth-cap query never matched. `moveTerm` would accept a parent inside its own subtree (creating a cycle) and allow a subtree past the depth cap, raising no error.
  - `@platform-modules/affiliate` — `maturity-sweep.ts` cast every `execute()` result to `{ rows }`. Its opening probe short-circuits the sweep when empty, so on a handle returning the other shape the sweep silently promoted nothing and moved no money.

  Both now normalize both shapes rather than casting to one: `execute()` yields a plain row array inside a platform-wrapped transaction and the driver's `{ rows }` envelope on an unwrapped handle, and consuming code must tolerate either.

  Breaking at the seam, released as minor: these packages are pre-1.0.0 and stay 0.x until the deliberate public release.
- ee63372: Generate `uuid` and `createdAt`/`updatedAt` column defaults client-side so inserts work on SQLite/D1 as well as Postgres.

  `.defaultRandom()` and `.defaultNow()` declare a *database-side* default, and drizzle inlines that default's SQL into the INSERT. Under the SQLite dialect that emits `values (gen_random_uuid(), …, now())` — neither function exists in SQLite, so every insert into an affected table fails at runtime with `no such function`. On mod-cms, where D1 is the default database, that meant creating a content entry, a revision, a taxonomy term, a form submission, a field value or a translation was impossible; the audit insert failed the same way but is swallowed by `logAudit`, so audit logging silently did nothing.

  The columns now use `$defaultFn` alone. `$defaultFn` binds the value as a parameter from the runtime, which both dialects accept.

  `.defaultRandom().$defaultFn(…)` does **not** work — the SQL default takes precedence and the function never runs. `@platform-modules/i18n-content` carried that combination on `translationValue.id` and was still emitting `gen_random_uuid()`; it is corrected here.

  Hand-written DDL is unchanged, so Postgres columns keep their `DEFAULT gen_random_uuid()` / `DEFAULT now()` and any direct-SQL insert path behaves exactly as before. No package in this repo generates DDL from these schemas.

  One behavioral note at the seam: timestamps now come from the application clock rather than the database clock. Keyset pagination over `createdAt` assumes a monotonic clock, so a writer with backward clock skew can insert a row that sorts behind an already-served page boundary and is not returned by the listing.

### Patch Changes

- d515db0: Add the recursive typed EAV migration floor with immutable field definitions, migration journals/checkpoints, stable node identity, and reversible D1/Postgres parity proof.
- Updated dependencies [1fd8651]
  - @platform-modules/db@0.4.0

## 0.0.1

### Patch Changes

- Updated dependencies [df32439]
  - @platform-modules/db@0.3.0