serverpod-offline-sync/docs/row-ownership.md

# Row ownership and space isolation

Status: implemented.

## Summary

Synced rows are physically shared on the server while their CRDT metadata is
tracked per user. Ownership is therefore implicit: nothing in the data model
prevents two users from tracking the same `(table, uuidRowId)` pair, but the
storage model cannot actually represent that state. Deletes survive it because
visibility is per-user metadata; field values cannot, so concurrent updates by
different users silently overwrite each other across tenant boundaries.

This document makes ownership explicit. The core decisions:

> **Every synced domain row is globally identified by `(table, uuidRowId)` and
> owned by exactly one space.** `spaceId` is not part of row identity; it is
> ownership and isolation metadata on a globally identified row. Two spaces
> never materialize or sync the same `(table, uuidRowId)`. Cross-space UUID
> reuse is a terminal ownership collision, not valid multi-tenant state. Sync
> fails on that collision, records a durable violation, and does not advance
> the sender past the rejected change. Sharing is modeled as shared spaces that
> multiple identities subscribe to — never as two spaces aliasing one row.

> **Every synced table declares `spaceId`, identically on server and
> client.** The column is a hard requirement, never `scope=serverOnly`, so
> both databases run one schema shape and the package has exactly one code
> path. "Single user" is a usage pattern (a database that happens to contain
> one space), not a mode: there are no mode flags, no schema branching, and
> no per-mode test matrix.

> **The sync layer owns the column: stamp if null, assert if set — and the
> value leaves the way it came in.** On every insert, a null `spaceId` is
> stamped from the effective space; a non-null one must equal it or the
> operation throws. Stamped values are stripped back to null on returned
> rows and on space-scoped reads, so the database-local value never leaks into
> endpoint payloads; explicitly provided values are kept on the way out.
> One rule serves both personalities: users who never want to think about
> tenancy never do, and users who pass it explicitly get verified
> explicitness and keep seeing what they manage. After insert the column is
> immutable.

The package contract for a synced table is two structural requirements, both
validated at `initialize()` with paste-ready error messages:

```yaml
fields:
  id: UuidValue?, defaultPersist=random_v7
  ### Owner space of this row. Maintained by the CRDT sync layer.
  spaceId: int?
```

No model inheritance is needed: like the UUID primary key requirement, the
contract is structural (a reserved column name with a required type, checked
by `CrdtSchemaRegistry`), not nominal (a base class), which is the only
option anyway since Serverpod models cannot extend models from an imported
module. The same constraint shapes the implementation: with no common
interface to call, the package stamps the field through dynamic dispatch —
`(row as dynamic).copyWith(spaceId: …)` (or the plain field setter, since
generated fields are mutable). The field name is a compile-time constant, so
the serialization manager's `patchTableRow` would be the wrong tool here: its
serialize–patch–deserialize round-trip costs 50–200× a dynamic call and is
justified only where field names are data-driven, as in the merge path.

## Goals and constraints

1. **One ownership model, one schema shape, one code path.** Earlier drafts
   explored an optional column and then schema-presence modes; both were
   rejected because every mechanism (CRUD, merge, sync, unique and FK
   invariants) would need building and testing per branch. Requiring the
   column everywhere deletes the mode system, the schema-projection
   generator feature it needed, and the per-mode test matrix. The cost — an
   int column (~1–2 bytes per row in SQLite) and a visible read-only field
   on domain models — is accepted.
2. **No external model inheritance.** Serverpod does not let application
   models extend models from an imported module, so a base class cannot
   carry shared columns into user tables. Requirements are enforced by
   runtime validation, and generic field access by dynamic dispatch or
   `patchTableRow`.
3. **Roles are not hard-coded.** Nothing in the package assumes the server
   is multi-user or the client single-user. Multi-account devices and
   single-tenant servers are the same code path with different space counts.
4. **Application code cannot get tenancy wrong.** The package must verify
   any user-supplied space against the effective space regardless (a tenant
   id from application code can never be trusted), so correctness never
   rests on call-site discipline.
5. **Multi-device per user keeps working unchanged.** Devices of one user
   share the user's space and merge into one chain.
6. **Sharing must not be foreclosed.** Collaborative data is a core
   local-first use case; the design must leave a clean path to it.

## Problem: ownership is implicit

The server holds one physical domain row per `(table, uuidRowId)`, while CRDT
metadata (`crdt_data_rows` and its children) is keyed per
`(spaceId, tblId, uuidRowId)`. Before this design was implemented, that
metadata shape let multiple spaces track the same row even though the domain
table could only materialize one copy. Symptoms, in increasing severity:

1. **Tombstone masking.** A tombstone recorded by one user (or for another
   table) used to hide unrelated rows sharing a UUID. Fixed by scoping the
   visibility predicates — but the fix required defining what an *unscoped*
   query means ("hidden only when hidden for every tracking user"), an
   aggregate query shape, and a covering index we could not afford on
   clients. Complexity that exists only because multiple trackers can.
2. **Update leakage is unrepresentable.** Field values materialize into the
   single shared domain row. If two users' chains track it, each chain has
   its own LWW history, but there is only one cell to materialize into.
   Whoever merges last wins, across tenant boundaries, with no record of the
   conflict. This is not a policy gap — divergent per-user values cannot be
   stored at all.
3. **Merge-insert overwrites other users' rows.** `_applyMergeInsertForMissingRow`
   treats "missing" as "no CRDT row *for the merging user*" and writes the
   domain row update-then-insert. A client that syncs an insert reusing
   another user's row UUID overwrites that user's data through the normal
   sync path. (Updates and deletes are already safe: they no-op when the
   merging user has no tracker.)
4. **Unique-conflict resolution crosses tenants.** Domain unique indexes are
   global on the server and the flag policy (see
   `unique-constraint-invariants.md`) picks the winner by HLC with no notion
   of ownership. One user's insert can cause another user's newer row to have
   its unique value rewritten.

All four share one root cause: the metadata says who owns a row, but nothing
enforces that the answer is unique, and the domain storage assumes it is.

"Leave it up to users" is rejected explicitly: symptom 2 means the storage
model cannot represent what multi-space tracking promises, so allowing it is
not granting flexibility — it is permitting silent cross-tenant data
corruption. Applications that need shared data use shared *spaces* (see
`shared-spaces.md`), which preserve the one-chain-per-row property that makes
LWW merging coherent.

## The `spaceId` column

### Contract

`spaceId` is a reserved column name, required on every synced table:

- **It is not part of row identity.** The domain primary key remains the
  required `id: UuidValue`, and `(table, uuidRowId)` remains globally unique
  in the physical domain table. A row with the same UUID in a different space
  cannot be represented as a second domain row; it is treated as a corrupt or
  hostile ownership collision, the merge fails, and the violation is recorded.
- The value references `offline_sync_spaces.id` — the space that owns the row. Today
  a space is a user; the name is chosen deliberately so shared spaces can be
  added later without renaming storage again.
- **The value is database-local.** It is a normalization of the space UUID
  into the local `offline_sync_spaces` table; the client's int for a space is not
  the server's int for the same space. The sync layer re-maps it on merge
  and never ships it as data, and application code cannot ship it by
  accident either, because space-scoped reads return it stripped (see *Keeping
  the value off the wire*). This is also why the field is nullable: a model
  outside the database honestly carries "not mine to say" instead of a
  meaningless local id.
- `scope=serverOnly` is **rejected**: it would reintroduce asymmetric
  schemas and with them the mode branching this design exists to delete.
  The client-side registry sees the column missing and says exactly that.
- The column is **owned by the sync layer** and immutable after insert.
  The value round-trips the way it came in: callers that never provide it
  never receive it (stamped values are stripped from returned rows, and
  space-scoped reads return null), explicit callers keep seeing what they passed,
  and unscoped admin reads always return it populated (see *Keeping the
  value off the wire*).

### Write semantics: stamp if null, assert if set

On every local insert (and batch insert), the database proxy resolves the
effective space — transaction space, else persistent space, else the
operation throws, exactly as recording already requires — and applies one
rule per row:

- `spaceId == null` → the row is stamped with the effective space before the
  delegate insert (one write, not an UPDATE after the fact). Stamping uses a
  dynamic call, since no base interface can declare the field; this is the
  per-row hot path of every insert, which is why the heavier `patchTableRow`
  stays out of it. In-place mutation is preferred (generated fields are
  mutable, so no per-row allocation); dynamic `copyWith` is the fallback
  only if rows were ever immutable.
- `spaceId != null` → it must equal the effective space, or the operation
  throws. Explicit callers get verification for free.

Consequences of the rule:

- **Cross-space writes cannot succeed.** A model carrying another space's id
  trips the assertion; a model carrying null is stamped with the acting
  space and then the space-scoped `WHERE` filter (which includes
  `spaceId = <space>`) matches no foreign row, so the update or delete fails
  with "no rows" instead of touching another tenant's data.
- **The column never appears in update SET lists**, exactly like `id`
  (`_updateRowWithoutRecording` already filters `columnName != 'id'`).
  A tampered model cannot attempt to move a row across spaces.
- **Merges stamp from the merging space** (each merge acts in exactly one
  space), and incoming payloads have the field stripped by
  `_sanitizeMergeRowData` — it is not portable data.

The same effective-space requirement applies to every CRUD entrypoint, not
only inserts:

- `insert`/`insertRow`/batch insert: stamp-or-assert before the delegate write;
- reinsert of a tombstoned row: only the same owning space can revive it;
  a foreign-space row id is a UUID ownership collision, not a reinsert target;
- `update`/`updateRow`: the space-scoped `WHERE` includes `spaceId = <space>`, and
  `spaceId` is removed from the SET list even if it appears in `columns`;
- `updateWhere`/`updateById`: any `ColumnValue` for `spaceId` is rejected (or
  dropped before SQL generation, but rejection gives better caller feedback),
  and the space-scoped `WHERE` still includes `spaceId = <space>`;
- `delete`/`deleteRow`/`deleteWhere`: deletes only see rows owned by the
  effective space, so a foreign row id returns "no rows" instead of recording a
  tombstone for another owner;
- future `upsert`: must first resolve ownership of the conflicting row and may
  update only when the existing row's `spaceId` equals the effective space.

### Keeping the value off the wire

The wire discipline is enforced by construction and respects the caller's
choice: **the field leaves the package the way it came in.**

- **Stamped values stay internal.** When a model comes in with `spaceId`
  null, the proxy stamps the effective space for the database write, and
  the rows handed back by that operation carry null again. The same applies
  to space-scoped reads, which have no input model to respect. Stripping applies to
  every synced model instance returned by the proxy, including nested
  `IncludeObject` and `IncludeList` result graphs. Callers who never provide
  the value never observe it, and cannot serialize the database-local int into
  endpoint payloads even by accident.
- **Explicit values are respected.** When a model comes in with `spaceId`
  set (and verified by the assert arm), the rows returned by that operation
  keep it — the package does not hide a value the caller knowingly manages.
  In batch operations the rule applies per row.
- **Unscoped (admin) reads keep it populated.** Within a space-scoped read the
  value is redundant — it always equals the acting space — but for admin
  reads it is the only per-row ownership information, and those results are
  consumed server-side.

Mechanically this stays on the cheap path: rows returned by the delegate are
fresh, package-owned instances, so stripping is a per-row in-place setter
with no allocation, applied at the same proxy layer that injects the
visibility predicates. The caller's own input instance is never left
observably mutated by a stamp: the implementation either stamps a copy bound
for the database or restores the field after the write.

This mirrors mature row-level-security practice, where the tenant column is
populated from session context (`DEFAULT current_setting(...)`) precisely so
application code cannot get it wrong, rather than passed through call sites.

A convenience accessor exposing the effective space's normalized id (e.g. on
the transaction or the database) may be added for explicit callers and
custom filtering; it is not required by any flow, and its documentation
carries the database-local warning above.

### Sync-layer exclusions

`spaceId` is infrastructure that happens to live in the domain row. Keyed by
the reserved name, it is excluded from:

- `crdt_schema_columns` registration and field-level CRDT metadata (it has no
  HLC history; it is never merged as a field) — which also keeps it out of
  the synced-schema hash symmetrically on both sides;
- update recording (`afterUpdate` ignores it) and merge payload sanitization;
- outbound sync insert payloads and outbound sync update payloads. The sender
  must not serialize `spaceId` from the domain row; the receiver stamps the
  merge-local normalized space instead.

### Validation

`CrdtSchemaRegistry` extends the existing UUID primary key check:

- every synced table must have a nullable `int` column named `spaceId`
  (missing or mistyped → error with the paste-ready yaml, including the
  "remove `scope=serverOnly`" hint when the server schema has the column and
  the client schema does not);
- every non-primary unique index on a synced table must include `spaceId`
  (missing → error), except a foreign-key-only unique index whose referenced
  tables are also synced. Ordinary global unique indexes are forbidden for now
  because the package does not yet define deterministic cross-space conflict
  resolution.

## Unique indexes

Per-space uniqueness is declared by the application as an ordinary composite
index, which now exists identically on both databases:

```yaml
indexes:
  person_space_name_unique_idx:
    fields: spaceId, name
    unique: true
```

Ordinary global unique indexes are rejected until the package has a
deterministic cross-space arbitration policy and a way to record the
losing-side release in that space's CRDT chain. "Whoever syncs first" is not
an acceptable CRDT rule. The only permitted global unique indexes are
foreign-key-only indexes: every index element must be a column, and the
indexed columns must be composed only of complete foreign-key column sets whose
referenced tables are also synced. Under the current `flag` unique-conflict
policy, every indexed FK column in that exception must be nullable. A required
unique FK has no safe loser value: `null` violates the column, a synthetic UUID
violates the FK constraint, and keeping the original value violates uniqueness.
This covers optional Serverpod relation indexes, including optional one-to-one
and composite FK relation indexes, without forcing duplicate per-space indexes
for include-query performance. Those columns point at globally unique synced row
ids, and FK enforcement already repairs or rejects cross-space references, so
they are not treated as global application-level arbitration. The model must
still declare space-scoped composite unique indexes for application uniqueness. On a
deployment that happens to contain one space, a composite unique with a constant
`spaceId` accepts exactly the same rows as a plain unique; nothing degrades.

The unique-conflict resolver discovers indexes from the table definitions,
uses `spaceId` as part of the lookup predicate for space-scoped indexes, and releases
only releasable non-`spaceId` columns. For allowed FK-only global unique
indexes, nullable FK columns may be released to `null`; required FK columns are
rejected during schema initialization. Same-space application conflicts are
handled by space-scoped composite indexes, and cross-space references are handled by
FK repair. Rows in different spaces never form a conflict group under the
required composite index. The flag policy (see `unique-constraint-invariants.md`)
is otherwise unchanged.

## Enforcement mechanics

1. **Local writes: stamp-or-assert** (see *Write semantics*), with the
   effective space required exactly as it already is for CRDT recording.
2. **Merge-insert ownership check.** `_applyMergeInsertForMissingRow` treats
   an existing domain row with another `spaceId` as a UUID ownership collision.
   This is a terminal sync error, not a recoverable conflict policy. The check
   is atomic with the physical write:
   1. Look up the domain row by `(table, uuidRowId)` before writing.
   2. If it exists with the merging space, follow the same-space
      update/reinsert path.
   3. If it exists with another space, roll back the merge work for the
      incoming change, record a durable integrity violation with denormalized
      operation, table, row, space, node UUID, and HLC fields, then fail the
      sync session.
   4. If it does not exist, insert with `spaceId` stamped from the merging
      space.
   5. If the insert races another space and hits the domain primary key,
      re-read the row and apply the same ownership decision before any CRDT
      metadata survives the transaction.

   Step 5 carries an implementation constraint: today the merge records the
   space's CRDT row (`_upsertMergeRow`) — and unique-conflict resolution and
   FK-safe handling, which may mutate neighboring rows — *before* the domain
   write where the collision surfaces. The insert application must therefore
   run inside a savepoint rolled back on collision (or be reordered
   domain-write-first), so that a rejected insert leaves behind no CRDT row,
   no released unique values, and no FK attempts. The durable violation is
   written after that rollback, outside the rolled-back merge work, so the
   diagnostic record survives even though the bad change does not.

   The durable record is sparse exception state, not a second ownership index.
   `offline_sync_integrity_violations` stores one deduplicated row per observed
   violation key: `(type, operation, domainTableName, uuidRowId,
   ownerSpaceUuid, incomingSpaceUuid)`. All other stored fields are
   denormalized scalars so the record remains useful after merge rollbacks and
   on databases that do not share the same local CRDT metadata rows. There are
   no foreign keys into CRDT metadata tables (`crdt_schema_tables`,
   `crdt_data_rows`, `crdt_nodes`): relations would often target rows that
   never committed or no longer exist when the violation is persisted outside
   the rolled-back merge transaction.

   `type` and `operation` are serialized enum names. The current types are
   `ownershipCollision` and `missingDomainRow`; operations distinguish merge
   and outbound insert/update/delete paths. Stored alongside the key:
   `uuidNodeId` and `hlcDatetime` / `hlcCounter` for the rejected change's
   causal context; optional `crdtDataRowId` as a same-database hint to the
   local metadata row id when one was known (for example outbound sync with a
   stale tracker); `firstSeenAt`, `lastSeenAt`, and `occurrences`.
   `ownerSpaceUuid` is nullable because a corrupt physical row can exist with
   a null owner;
   `incomingSpaceUuid` is required and globally locatable. `crdtDataRowId` is
   often null on merge-insert collisions and is not portable across devices.
   The record intentionally stores neither the rejected payload nor a payload
   hash. The sync response returns the violation identifier and checkpoints do
   not advance past the rejected change, so retrying the same corrupt client
   state fails deterministically until the application resolves it.

   Re-keying the incoming row to a synthetic UUID was considered (non-lossy,
   consistent with the unique flag policy) but rejected: row identity re-keying
   breaks the incoming chain's own foreign key references to that UUID, which
   would require alias tracking across the whole merge. UUIDv7 collisions
   between honest clients are negligible; a collision is a bug or an attack,
   and failing the sync is the only response that preserves both the existing
   owner's data and the sender's still-unaccepted local change.
3. **Foreign keys stay within a space.** A synced row may only reference
   synced rows of its own space. Locally,
   `_assertVisibleForeignKeyTargets` gains a space-equality check next to
   its visibility check; in merges, a foreign-space target is treated
   exactly like an invisible target, so the existing attempt/override
   machinery (see `foreign-key-invariants.md`) repairs the reference instead
   of linking across spaces. Cascade deletes and SET NULL projections
   therefore never cross a space boundary. References to non-synced tables
   are unaffected.
4. **No duplicated metadata lookup index.** An earlier draft added a unique
   index on `crdt_data_rows (tblId, uuidRowId)` as defense-in-depth. It is
   dropped because the hot CRDT paths need the existing space-scoped index
   `(spaceId, tblId, uuidRowId)`, and adding the unscoped index duplicates
   most of the key space on every row. Instead, every consumer of CRDT row
   metadata is required to prove ownership against the physical row: reads key
   visibility through `domain.spaceId`; outbound sync serializes a domain row
   only when `domain.spaceId` equals the collecting space; merge, FK, and
   unique helpers reject metadata that attempts to materialize or mutate a
   foreign-owned row. A stray foreign tracker is therefore inert by consumer
   discipline, while an observed attempt to use it is a durable sync violation.
   Deployments that need a stronger offline audit can run an explicit scan for
   duplicate `(tblId, uuidRowId)` trackers; the package does not pay that
   index cost on every database by default.

## Read path

Ownership on the row collapses every visibility predicate into an
index-covered probe; the aggregate (`GROUP BY`/`MIN`) form, its full-scan
cost (~95 ms per point lookup at 1M CRDT rows in the SQLite benchmark), and
the covering index it wanted never ship.

- **Space-scoped query** (a `transactionForUser` read, or the persistent default
  space): rows are filtered by `spaceId = <space>` — the isolation promise,
  and the central point of the design — plus the existing correlated
  `NOT EXISTS` visibility probe on `(spaceId, tblId, uuidRowId)`.
- **Unscoped query** (no transaction space, no persistent space — admin
  reads): no isolation filter; the visibility probe is keyed by the row's
  own column — `NOT EXISTS (… WHERE c.spaceId = t."spaceId" AND tblId = ?
  AND uuidRowId = t.id AND visibility > …)` — fully covered by the existing
  unique index. A row is visible while its owner sees it.

On a database containing one space, the two views return the same rows; no
code anywhere branches on space count. Rows with a null `spaceId` — created
behind the sync layer's back (raw SQL, seed fixtures) — are visible only to
unscoped reads; seeding flows must stamp spaces, which ties into the
reference-table open question.

Multi-space deployments additionally gain Postgres row-level security as an
option, since policies can reference `spaceId` directly — the package's
isolation filter and RLS become two enforcements of one predicate.

## Sync path

Sync also consumes CRDT metadata, so it follows the same row-ownership rule
as reads and writes:

- **Outbound inserts** collect `crdt_data_rows` for one space, then fetch the
  physical domain row only if `domain.spaceId` equals that same space. A
  stale or hostile tracker for a foreign-owned row fails the sync and records
  a durable violation before any payload is produced.
- **Outbound updates** fetch the column value only after the field's CRDT row
  is proven to match the domain row owner. `spaceId` itself is never emitted
  as an update because it has no field metadata.
- **Outbound deletes** are space-scoped by CRDT metadata; a foreign tracker can at
  worst try to produce a tombstone for a row the collecting space never owned,
  so delete collection also verifies the domain owner when the domain row still
  exists and fails with a durable violation on mismatch.
- **Inbound inserts and updates** sanitize any incoming `spaceId` before merge
  and stamp from the merge space. Incoming row UUIDs that already belong to
  another space fail the sync session as ownership collisions, record durable
  violations, and do not advance checkpoints past the rejected change.

This is the audit rule for all future sync helpers: no path may serialize,
materialize, resolve conflicts for, or project a synced domain row through
CRDT metadata unless the domain row's `spaceId` equals the effective CRDT
space, except explicit unscoped admin reads.

## API surface

The package is unreleased, so renames land directly, with no aliases and no
data migration. Public API stays user-first (that is the semantics
developers expect to write); internals and storage are named after what they
actually key — spaces:

- **`transactionForUser(userId, fn)`** keeps its name and shape: it resolves
  the user's UUID to the local space and pins the transaction to it — writes
  stamp and assert against it, reads are isolated to it. The optional
  `spaceId` parameter is reserved for shared spaces (see *Future*).
- **`persistentUserId`** remains the way to pin a default space outside
  explicit transactions — the standard single-user client configuration.
- **`mergeChanges(spaceId: …)`** keys the chain being merged — the parameter
  is renamed from `userId` to what it always was.
- **Internals and storage** adopt space naming in one sweep: `crdt_users` →
  `offline_sync_spaces` (`CrdtUser` → `OfflineSyncSpace`, `uuidUserId` → `uuidSpaceId`),
  `crdt_data_rows.userId` → `spaceId` (matching the domain column),
  `userForTransaction`, the effective-user helpers, and the HLC manager keys.
- **Unscoped reads** remain supported as admin reads.

Consequences for application code, stated as contract:

1. **Every synced table declares `id: UuidValue` and `spaceId: int?`.**
   Startup fails with a paste-ready error otherwise.
2. **`id` remains globally unique per synced table.** The same
   `(table, uuidRowId)` cannot exist in two spaces. If merge observes that
   UUID already owned by another space, it records a durable ownership
   violation and fails the sync.
3. **`spaceId` round-trips the way it came in** — left null, the sync layer
   stamps it internally and returns it null (space-scoped reads return it null
   too); provided explicitly, it must match the effective space or the
   operation throws, and it stays visible on the way out. The column is
   immutable after insert.
4. **Space-scoped reads are isolated to their space.** Reading across spaces is an
   explicit unscoped (admin) read.
5. **Merge-inserts colliding with another space's row fail the sync**, never
   apply over the owner's data, and never advance the sender past the rejected
   change.
6. **Foreign keys on synced tables target same-space rows.** Cross-space
   references are repaired by the existing FK override machinery instead of
   being linked.
7. **Application unique indexes on synced tables include `spaceId`.** Global
   unique indexes are rejected at startup until a deterministic cross-space
   policy exists, except FK-only indexes whose referenced tables are synced.

## Implementation plan

Each phase lands independently green; later phases depend on earlier ones
only where stated. No generator or Serverpod fork changes are required.

- **Phase 1 — naming, contract, validation.** The user→space renames in
  internals and storage (above), done first so later phases build on the
  final names. `CrdtSchemaRegistry`: the `spaceId` presence/type check with
  the `serverOnly` hint, and the global unique-index rejection. Reserved-name
  exclusions: `crdt_schema_columns` registration, field-level merge, update
  recording, inbound merge sanitization, and outbound sync serialization.
  Test models gain `spaceId`.
- **Phase 2 — write-path enforcement.** Stamp-or-assert in the proxy insert
  paths (dynamic dispatch on the known field); `spaceId` rejected from
  `updateWhere`/`updateById` column values and excluded from model-update SET
  lists; ownership assertion on update, delete, reinsert, and upsert; atomic
  merge-insert ownership check with durable fail-fast collision handling,
  including primary-key race handling; merge stamping from the merging space;
  FK same-space assertion and merge repair routing; unique resolver limited
  to space-scoped indexes, with ordinary global indexes rejected by validation.
- **Phase 3 — read paths.** Isolation filter for space-scoped reads; row-keyed
  visibility probe for unscoped reads; stripping of `spaceId` from space-scoped
  root and include-graph read results; outbound sync owner checks before
  serializing inserts, updates, or deletes; delete the `GROUP BY`/`MIN` form
  and the server-only-index TODO from `tombstone.dart`.
- **Phase 4 — tests and benchmarks.** The plan below, plus updating
  `TombstoneSpaceBenchmark` to ownership-conformant seeding (one tracker per
  row); its current second-tracker fixtures survive only as the enforcement
  scenario.

### Test plan

Because no code branches on space count or deployment role, mechanisms are
tested **once**, not per combination:

- **Mechanism suites (run once).** All existing CRUD, tombstone visibility,
  unique-conflict, FK-invariant, merge, and sync suites continue on the
  uniform model; test models declare `spaceId` and are otherwise unchanged
  (the stamp rule means call sites do not change).
- **Enforcement tests.** Stamping on insert; assert-on-mismatch throwing for
  insert and update; immutability of the column; round-trip symmetry (null
  in → null out, explicit in → kept, per row in batches); stripping on
  space-scoped root and included read results and retention on unscoped reads; the
  caller's input instance left unmutated; merge-insert collision fails sync,
  records a durable violation, and covers the primary-key race regression;
  stale foreign tracker fails before producing outbound insert/update/delete
  payloads and records a durable violation (the multi-user fixtures in
  `tombstone_space_test.dart` become these tests); FK cross-space repair;
  schema validation rejects ordinary global unique indexes and accepts
  FK-only unique indexes whose referenced tables are synced.
- **Combination smoke tests (one each, not per mechanism).** A two-space
  database proving read isolation and per-space uniqueness on SQLite; one
  `withServerpod` sync roundtrip proving the wiring end to end. These exist
  to prove deployment shapes compose, not to re-verify mechanisms.

## Shared spaces

`offline_sync_spaces` is the entity everything is keyed by: sync, locking, HLC
chains, and visibility. Sharing becomes an indirection in front of it, not a
change to row ownership:

- an authenticated identity maps to one personal space plus any number of
  shared spaces;
- a shared list/document/workspace is a space that several identities
  subscribe to and sync;
- every row still has exactly one owning space (its `spaceId`), so merging
  stays one chain per row and nothing in this document's enforcement
  changes.

This is the standard shape of sharing in local-first systems, and it is the
reason the column and concept are named `space` rather than `user`.
Implementation details and current limits live in `shared-spaces.md`; the rest
of this section summarizes why the ownership rules remain compatible.

### Users authenticate, spaces act

The two concepts separate cleanly:

- a **user** is an authentication identity — what endpoints and the sync
  handshake verify;
- a **space** is the unit of replication and ownership — what transactions
  act in, what chains and tombstones are keyed by, what `spaceId`
  references.

Two conventions bridge them with no storage migration:

1. **A user's personal space has the user's own UUID.** This is already true
   (the space UUID keys the chain today), so every deployment without
   sharing is the degenerate case "each user has exactly one space: their
   personal one". Shared spaces are new `offline_sync_spaces` rows whose UUID is no
   user's id.
2. **Membership** is a server-side relation, `offline_sync_space_members(spaceId,
   userUuid, role?)`, managed by application code (invitations and roles are
   app domain). Personal-space membership is implicit; no row is stored for
   it.

### The transaction API

`transactionForUser` keeps its name and user-first semantics — that is what
developers expect to write — and gains an optional space:

```dart
db.transactionForUser(userId, fn);                  // acts in the personal space
db.transactionForUser(userId, fn, spaceId: listId); // acts in a shared space
```

- Without `spaceId`, the space resolves to the user's personal space — by
  convention 1, simply `userId` itself, which is exactly today's behavior.
- With `spaceId`, the `userId` argument is **not dead weight**: the package
  asserts `membership(userId, spaceId)` before acting. The parameter pair
  reads as "authenticated as `userId`, acting in `spaceId`, verified
  member" — the same predicate row-level security policies mirror. The
  caller remains responsible for `userId` being authenticated, exactly as
  today.
- A transaction acts in **exactly one space**. Cross-space writes need
  separate transactions; this is deliberate honesty about CRDT semantics,
  since two spaces' chains replicate independently and remote replicas can
  never observe a cross-space write atomically anyway.

### Reads: membership-wide; writes: space-pinned

A user-space-scoped read in a sharing world naturally means "rows in any space I am
a member of": the isolation filter becomes `spaceId IN (<member spaces>)`
instead of `spaceId = <space>`. The visibility probe is the row-keyed form —
the same probe as unscoped admin reads — so sharing simplifies the predicate
matrix rather than growing it: reads differ only in their membership filter
(one space, a user's spaces, or none for admin).

This is also exactly the shape of a Postgres row-level security policy
(`spaceId IN (SELECT spaceId FROM offline_sync_space_members WHERE userUuid =
current_setting(...))`), which is the standard practice for multi-tenant
isolation and on the Serverpod backlog; the package's application-level
filter and RLS become two enforcements of one predicate.

### Sync and clients

- A device syncs one chain per space it holds: each sync call enumerates the
  authenticated user's member spaces, verified server-side against
  `offline_sync_space_members` at the start of every sync cycle; the membership table
  itself is not synced. `OfflineSyncSpaceRole` is enforced as described in
  `shared-spaces.md`: reads stay membership-wide, writes are role-gated, and the
  server records `unauthorizedWrite` for rejected inbound changes.
- Every database is already sharing-capable schema-wise — the column is
  universal — so a sharing-enabled client simply holds several spaces' rows.
  One database file per space remains a valid alternative layout.
- Checkpoints stay per `(space, node)`: `crdt_nodes` stores the stable replica
  identity, while `offline_sync_space_nodes` stores each space's checkpoint for that
  node.

### What sharing does not change

Every enforcement rule in this document survives unchanged: `(table,
uuidRowId)` stays globally identified, rows have exactly one owning space,
merge-insert ownership checks compare spaces, unique conflict groups form
within a space, ordinary global unique indexes stay rejected, and
stamp-or-assert applies with the transaction's resolved space. Sharing adds a
membership layer in front of spaces; it does not touch row ownership.

## Resolved during review

Decisions that were open questions in earlier drafts, kept here so they are
not relitigated by accident:

- **Optional vs. required `spaceId`:** required, on every synced table and
  in both schemas. The optional-column draft and the schema-presence-mode
  draft were both rejected: each created a second ownership architecture (or
  a second schema shape) that every mechanism would have to support and
  test. Requiring the column everywhere deleted the mode system, the
  index-projection generator feature, and the per-mode test matrix, at the
  cost of ~1–2 bytes per row and a visible read-only field on domain models.
- **Global UUID row identity:** retained. Serverpod's generated domain model
  and ORM expect one primary key column, so the package does not introduce
  composite `(spaceId, id)` identity or a surrogate physical row id. A
  duplicate `(table, uuidRowId)` in another space is an ownership collision,
  not a valid second row.
- **`scope=serverOnly` on the column:** rejected (would reintroduce
  asymmetric schemas); validation forbids it.
- **Stamp-or-assert as the single write path:** chosen over required
  explicit passing (ceremony without added security — the package must
  verify any supplied value anyway, and the persistent-user client pattern
  has no transaction object in hand) and over dual documented paths (two
  mental models). Explicit callers still get verified explicitness through
  the assert arm.
- **Nullability:** nullable, permanently. The int is database-local, so a
  non-nullable field would force clients to serialize a meaningless local
  id into endpoint payloads; null honestly means "not mine to say". The
  sync layer guarantees the value on every synced row.
- **Wire transport prevented by construction, respecting the caller:** the
  value leaves the way it came in — stamped values are stripped from
  returned rows and space-scoped reads, including include graphs; explicitly
  provided (and verified) values are kept on the way out; unscoped admin
  reads keep the value as the only per-row ownership information server-side;
  outbound sync never serializes it. Mechanism: in-place dynamic mutation on
  package-owned instances (preferred over `copyWith`, the fallback only
  under immutability; `patchTableRow` is reserved for the merge path, where
  field names are data-driven and its 50–200× cost is justified), never
  leaving the caller's input observably mutated.
- **Unscoped-read semantics:** an unscoped read is an admin view; a row is
  visible while its owner sees it. Space-scoped reads are isolated to their space
  — confirmed as the central goal of the design.
- **Fail vs. skip vs. re-key on colliding inserts:** fail and record a
  durable violation. Skipping can silently consume a client's rejected change
  and advance sync state without any recovery policy; re-keying row identity
  would break the incoming chain's own FK references.
- **Metadata backstop index:** dropped only with the consumer audit rule:
  every read, write, sync, merge, FK, and unique path that consumes CRDT row
  metadata must prove `domain.spaceId` equals the effective metadata space
  before materializing domain data.
- **Audit surface:** a durable `offline_sync_integrity_violations` table is part
  of the terminal sync-integrity contract. Session logs may mirror it, but
  they are not the source of truth.
- **No compatibility aliases:** the package is unreleased, so renames
  (storage and API parameters) land without aliases. The storage migration
  preserves existing durable violation rows by mapping prior operation strings
  to the new enum names.

## Open questions

1. **Globally shared reference tables.** Read-only data synced to every user
   would violate "every row has one owning space" and is out of space; if
   the use case materializes, it likely becomes a third table kind rather
   than a relaxation of the contract. Related: rows seeded behind the sync
   layer's back carry a null `spaceId` and are visible only to admin reads.
2. **`spaceId` naming collisions.** The name is reserved by convention; an
   application with a pre-existing, semantically different `spaceId` column
   on a synced table has no escape hatch other than renaming its column.
   Acceptable for now; revisit if it bites real users.
3. **Membership revocation semantics (sharing).** When a user loses membership
   of a space, the server stops streaming it at the next sync cycle. The device
   may still store the space's data and unsynced local changes, though
   membership-wide reads no longer expose revoked rows. Purge-vs-keep policy,
   and whether revoked unsynced changes are surfaced or dropped, need their own
   design.
4. **Roles within a space (sharing).** Resolved in `shared-spaces.md`.
   `offline_sync_space_members.role` is enforced for writes through the closed
   `OfflineSyncSpaceRole` enum: the server rejects and records unauthorized inbound
   writes, the client rejects local writes from projected `readOnly` roles, and
   reads remain membership-wide. The remaining cleanup question for demoted or
   revoked unsynced local changes is part of question 3.