These two integration tests append an execution-default to the contract at
runtime (a tags.updated_at @updatedat default; a user_tags.created_at onCreate
default). After the ref gained a required namespace, the structural validator
rejects the namespace-less refs. Add namespace: "public".

Signed-off-by: Alexey Orlenko's AI Agent <robot@aqrln.net>

The flat sql-builder write path called applyMutationDefaults with no
namespace, so a multi-namespace contract with two same-named tables could
apply the wrong namespace's execution default (or none at all). The ORM
path already threaded it everywhere; only the lane did not (F01).

Forward `namespace: namespaceId` from buildParamValues/buildSetExpressions,
and make MutationDefaultsOptions.namespace required so a forgotten namespace
is a type error rather than a silent degrade to table-name-only matching
(F02) — the runtime matcher now filters on namespace unconditionally.

Regression test: two same-named `users` tables in `public`/`auth` with
per-namespace execution defaults; each insert/update now picks up its own
namespace's default and not the collision twin's.

Signed-off-by: Alexey Orlenko's AI Agent <robot@aqrln.net>

The guard's remediation advice told users to write the related model (e.g.
`Role`) directly, but the unfillable required column lives on the junction
table (`user_roles`), not the related model. Writing `Role` would never
populate the junction payload. Point the advice at the junction `through.table`
instead, and pin the wording in the create/connect guard tests.

Signed-off-by: Alexey Orlenko's AI Agent <robot@aqrln.net>

The cast-smuggling section listed `blindAs`, which does not exist in
`@prisma-next/utils/casts` — the only exported helpers are `blindCast` and
`castAs`. Reference the real helpers so the forbidden-pattern enumeration is
accurate.

Signed-off-by: Alexey Orlenko's AI Agent <robot@aqrln.net>

…resolution

Table names are unique per namespace, not globally — the storage IR and runtime
already key tables by namespace, and the low-level buildSqlContractFromDefinition
path already accepts same-named tables in different namespaces. The DSL/emit path
did not:

- The table-uniqueness check keyed `tableOwners` by bare table name, so two models
  mapping `public.users` and `shadow.users` were rejected. Key it by
  `(namespace, table)`.
- M:N junction resolution lost the junction's namespace: the lowered `through`
  node carried only a table name, and `buildThroughDescriptor` re-derived the
  namespace by bare table name (last-wins). With two `user_roles` junctions this
  emitted the wrong `through.namespaceId` for both. Carry the junction's namespace
  on the `through` node and resolve `through.namespaceId ?? defaultNamespaceId`.

Value-identical for existing single-namespace contracts (no fixture drift).

Signed-off-by: Alexey Orlenko's AI Agent <robot@aqrln.net>

…ate-disconnect

Two type-level corrections to the nested-write mutators, mirroring the runtime:

- The junction-payload required-column gate (`HasRequiredJunctionPayload`) matched
  execution defaults by `(table, column)` only, dropping the junction namespace
  even though `ExecutionMutationDefault.ref` is namespace-scoped (runtime twin:
  the sql-builder/applyMutationDefaults namespace fix). In a multi-namespace
  contract an execution default for `public.user_roles.token` could make
  `shadow.user_roles.token` look optional, wrongly enabling create/connect on the
  shadow relation. Thread the junction's namespace through the payload-field
  helpers and include `ref.namespace` in `HasExecutionCreateDefault`.
- `MutationCreateInput` typed junction `disconnect` as available, but `createGraph`
  rejects any nested disconnect during `create()`. Add a create/update context
  axis to `RelationMutator` so disconnect is unavailable in create input; it stays
  available in update input.

Tested against a new emitted two-namespace fixture (`junction-namespaces`) where
`public`/`shadow` share the `user_roles` junction but only `public` carries the
execution default.

Signed-off-by: Alexey Orlenko's AI Agent <robot@aqrln.net>

`withMutationScope` only opened a transaction when the runtime exposed a
top-level `transaction()`, but no runtime does — `transaction()` lives on a
connection (`connection().transaction()`). So every ORM mutation, including
multi-statement nested-write graphs, ran without a transaction: a failure after
the first write left a partial write behind. Open a connection and run the whole
graph inside its transaction (falling back to a bare scope only when the runtime
exposes no connection, e.g. unit-test stubs).

The integration harness is reworked to match: the `@prisma/dev` server is
PGlite-backed (one concurrent connection), so the wrapper now drives a single
`pg.Client` (not a pool) through the direct driver, and its `connection()` /
`transaction()` keep recording executions so the per-test execution-count
assertions still see reads and writes. New regression test: a nested M:N tag
create that fails (duplicate pk) after the parent insert now rolls back the
parent and the junction.

Signed-off-by: Alexey Orlenko's AI Agent <robot@aqrln.net>

…n emitted fixture

The execution-default junction-payload test hand-cloned the emitted contract and
mutated `storage`/`execution` directly (dropping `user_tags.created_at`'s storage
default and pushing an execution default by hand) — exactly the pattern the
no-contract-data-patching rule forbids. Replace it with a dedicated emitted
fixture (`execution-defaulted-tags`) authored through the DSL, where
`user_tags.created_at` is NOT NULL with an execution-time onCreate default and no
storage default. The test now builds its runtime/context from that emitted
contract; the scenario comes from a shape the authoring surface can produce.

Signed-off-by: Alexey Orlenko's AI Agent <robot@aqrln.net>

The sql-orm-client integration runtime now drives a single long-lived client so
transactions hold one connection on the single-backend PGlite server. The dev
server's default 1s idle timeout reaps that client during brief idle windows
under full-suite load (a pool would reconnect; a lone client cannot), causing
intermittent "connection terminated" flakes. Raise the idle timeout to 30s for
these tests — no test holds the connection idle anywhere near that long.

Signed-off-by: Alexey Orlenko's AI Agent <robot@aqrln.net>

…main

Rebasing onto main picked up TML-2916 (un-namespaced PG models default to
`public`, dropping the spurious empty `__unbound__` storage slot). Re-emit the
slice's own fixtures (`junction-namespaces`, `execution-defaulted-tags`) so they
match the new emitter output; storage hashes update accordingly. Lockfile records
the `@prisma-next/sql-errors` dependency at the 0.14.0 workspace version.

Signed-off-by: Alexey Orlenko's AI Agent <robot@aqrln.net>

…ack version

The supabase example contract embeds the supabase extension pack version, which
the 0.14.0 release bump left at the stale `0.13.0` — `fixtures:check` re-emits it
to `0.14.0`. Re-emit so the committed contract matches; no schema change.

Signed-off-by: Alexey Orlenko's AI Agent <robot@aqrln.net>

Signed-off-by: Alexey Orlenko's AI Agent <robot@aqrln.net>