omnigraph/docs/dev/writes.md

# Direct-Publish Write Path

> History: the Run state machine and `__run__<id>` staging branches were
> removed in MR-771 (shipped v0.4.0). Writes now go directly to the target
> table; this document specifies that direct-publish path.

`mutate_as` and `load` prepare against one immutable branch-authority token,
write **directly to the target table**, and call
`ManifestBatchPublisher::publish` once at the end. The token is
`(Lance branch identifier, exact optional graph_head, accepted schema identity)`;
the exact `graph_head` check protects validation dependencies on tables the
write does not touch. Publisher row-level CAS on `__manifest` is the visibility
fence. Process-local branch/table queues reduce retries but are not distributed
authority.

## What this means in practice

- No `RunRecord`, no `_graph_runs.lance`, no `_graph_run_actors.lance`.
- No `omnigraph run *` CLI subcommands and no `/runs/*` HTTP endpoints.
- No `__run__<id>` staging branches; `__run__*` is no longer a reserved
  name. The branch-name guard was removed in MR-770, and any stale
  `__run__*` branch on an upgraded graph is swept off `__manifest` by the
  v2→v3 internal-schema migration on first read-write open. (The inert
  `_graph_runs.lance` bytes remain until a `delete_prefix` primitive lands.)
- Cancelled mutation futures leave **no graph-visible state** unless the
  manifest publish already completed. Before that point they can leave
  reclaimable uncommitted Lance files, or sidecar-covered committed table
  effects that the next quiesced recovery rolls forward/compensates. A
  first-touch named-branch write can also leave a target table ref, but it is
  never created without ownership: the schema-v3 sidecar is durable first and
  names that `(table_path, target ref)`. Reclaim and `cleanup` treat any
  matching pending sidecar as a hard stop. Quiesced full recovery accepts both
  logical crash shapes — sidecar durable with no ref yet, or an exact untouched
  ref at the inherited version. Because Lance creates a branch dataset before
  writing its authoritative `BranchContents`, the first shape may still contain
  a clone-only tree; recovery force-reclaims that absent-ref tree idempotently.
  It removes an exact untouched ref before deleting the empty intent. If several
  pending intents claim one ref, a no-effect intent discards only itself while
  any competitor remains; the last no-effect survivor cleans an untouched ref,
  or the effect-owning survivor recovers normally. `cleanup` remains the
  backstop for genuinely unclaimed legacy/stale refs; see the fork-reclaim note
  in [invariants.md](invariants.md).

## Mutation/load coarse OCC (RFC-022 first adapter)

Mutation and load use a closed prepare → effect → publish attempt:

1. run the branch-aware recovery barrier, then capture the target branch's native Lance
   `BranchIdentifier`, exact `graph_head:<branch>` (including absence on a
   fresh branch), accepted schema identity, and table snapshot;
2. run the complete validator and prepare every effect outside the effect gate.
   Existing-table transactions may stage reclaimable files here. A first-touch
   named-branch table retains its batch/predicate and pre-mints the transaction
   identity instead: Lance branch-local files cannot be staged until its target
   ref exists;
3. acquire the schema gate, branch gate, then sorted table queues; re-check for
   a relevant sidecar armed since step 1, then revalidate the token. Any
   unresolved relevant intent returns typed `RecoveryRequired` before effects;
4. on a pre-effect mismatch, discard the complete attempt. Append/Insert/Merge
   reprepare with a bounded retry; strict Update/Delete/Overwrite return typed
   `ReadSetChanged`;
5. arm a schema-v3 recovery sidecar. For each deferred first-touch table, create
   its target ref, stage branch-local files on that ref, and bind the staged
   transaction to the pre-minted UUID. Then commit every planned transaction
   with zero transparent conflict retries, confirm exact transaction UUIDs and
   table updates, and publish the pre-minted lineage intent under the same token.

The publisher checks the exact head and native branch identity on every CAS
attempt. It never reparents a validation-sensitive intent after contention. A
mismatch after any physical effect returns `RecoveryRequired` and leaves the
sidecar intact; it is not an ordinary retry loop.

This adapter preserves the documented single-writer-process support boundary.
The native branch identifier detects delete/recreate ABA but is not a Lance
conditional-ref fence, and destructive recovery remains unsafe beside a live
foreign process.

### Branch-merge authority fence (adapter bridge)

Branch merge still uses its writer-specific multi-commit table effects and
confirmation sidecar; it has not yet been converted to the RFC-022 exact-effect
adapter. It does, however, join the closed control boundary needed by this first
slice: after the strict recovery barrier it acquires the root-shared schema gate
and the sorted source/target branch gates, performs the final sidecar check,
loads one operation-local catalog from the accepted contract, captures both graph
heads plus the base/source/target snapshots, and holds those gates through table
effects and manifest publication. Planning stays outside table queues. Before
Phase A, merge acquires the conservative all-catalog table envelope for both
source and target, re-lists sidecars, and compares fresh source/target manifest
versions with the captured snapshots. A stale warm handle catalog or coordinator
snapshot is never accepted as that revalidation.

The source snapshot is a captured merge input, not authority that the target
manifest CAS can arbitrate. The current process-local source gate is a stronger
same-process fence around that capture, including delete/recreate ABA, but the
semantic contract is still "merge the captured source commit." A later source
advance does not invalidate an otherwise prepared target publish. Claiming
"latest source at target publish" would instead require a cross-process source
fence held through the target CAS.

That fence prevents a same-process target delete/recreate from reusing the branch
name underneath a merge plan. The race test deliberately recreates a target with
the same name and numeric Lance version but a different `BranchIdentifier`, so
version-only checking cannot accidentally satisfy it. This is a process-local
bridge, not a cross-process conditional-ref primitive and not a substitute for
the later full branch-merge read-set/reprepare adapter. `sync_branch` joins the
same root schema gate before replacing a handle's coordinator, so it cannot
overwrite merge's temporary target coordinator or change a native control's
active-branch authority mid-operation.

### Branch-delete orphaning exception

Branch deletion runs the healer first and then holds schema, the target branch,
and every accepted-catalog table gate through the native ref removal. An
unresolved sidecar scoped to that target does not permanently block deletion:
once those gates prove its in-process owner is no longer live, removing the
manifest branch makes its physical effects unreachable. The next write/open
records the orphan-discard recovery audit and deletes the sidecar. A
`SchemaApply` sidecar remains graph-global and blocks deletion. This exception
is specific to removing the authority that made the intent reachable; create,
merge, mutation, and load still reject relevant unresolved ownership.

### Native graph-branch control recovery

Graph branch create/delete do not use the graph-visible table-effect sidecar or
emit graph lineage. Their sole logical authority is Lance `BranchContents` for
the `__manifest` dataset, and Lance mutates that authority in two physical
phases:

- create shallow-clones `tree/{branch}` before writing `BranchContents`;
- delete removes `BranchContents` before reclaiming that tree.

Under the schema/branch/table control gates, create validates the name before
the clone and rejects a live graph name that is a physical path ancestor or
descendant of another live name. It then force-reclaims any absent-ref same-name
tree and performs at most two native attempts. An ambiguous result is accepted
only when fresh metadata has
the captured parent branch/version/incarnation plus exactly one new identifier
element and the target opens. Foreign or broken authoritative refs are never
deleted. Delete captures the exact target identifier; after an ambiguous error,
an absent ref is logical success, the same identifier preserves the original
error, and a different identifier is a typed delete/recreate conflict. Derived
tree cleanup is retried best-effort.

There is deliberately no branch-control sidecar: within the supported
single-writer-process topology, an absent ref makes a same-name tree unreachable
garbage; the path-prefix-disjoint namespace is what makes Lance's recursive
force cleanup exact. Same-name create is therefore the targeted reconciler.
First-touch data-table
forks remain sidecar-owned because they are physical effects of a graph-visible
mutation/load. Lance does not expose conditional ref create/delete, so this
classifier is not advertised as a cross-process branch-control fence.

Legacy prefix-overlap recovery is the one first-touch case that does not prove an
entire nested tree unreachable. If a Full sweep finds an ancestor first-touch
target with a live path-child, it keeps the sidecar. Open may complete for
leaf-first deletion only when the sidecar owns no physical table effect. A mixed
attempt that owns an effect plus an untouched fork must roll back as one recovery
outcome, so open fails closed while the child blocks fork cleanup. After an
existing handle or an offline Lance-level branch tool removes the child, a later
Full sweep reclaims the untouched fork, compensates the owned effect, and retires
the intent.

## Read-your-writes within a multi-statement mutation

A `.gq` query with multiple ops (e.g. `insert Person … insert Knows …`)
must observe earlier ops' writes when validating later ops (referential
integrity, edge cardinality). After MR-794 step 2+ this is implemented
via an in-memory `MutationStaging` accumulator in
[`crates/omnigraph/src/exec/staging.rs`](../../crates/omnigraph/src/exec/staging.rs),
shared by both `mutate_as` and the bulk loader:

- On the first touch of each table, the pre-write manifest version is
  captured into `expected_versions[table_key]` (the publisher's CAS
  fence at end-of-query).
- Each insert/update op pushes a `RecordBatch` into the per-table
  pending accumulator. Lance HEAD does **not** advance during op
  execution.
- Read sites (validation, predicate matching for `update`) consume
  `TableStore::scan_with_pending`, which scans committed via Lance
  and applies the same SQL filter to the pending batches via DataFusion
  `MemTable`. Same-query writes are visible to subsequent reads.
- Blob-bearing updates use the materializing variant: Lance reads only matched
  committed blob payloads as binary, the engine normalizes them back to the
  logical blob schema, and the full rows join the same pending-shadow union.
  Rewriting matched blob bytes costs I/O proportional to those bytes, but makes
  correctness independent of whether a physical index selects a different
  Lance merge plan.
- At end-of-query, `MutationStaging::stage_all` prepares exactly one staged
  transaction per touched table and `commit_all` commits it (concatenating accumulated
  batches; merge-mode dedupes by `id`, last-write-wins), and the publisher
  publishes the manifest atomically across all touched sub-tables. Existing
  tables stage before gate acquisition; a first-touch named-branch table stages
  after sidecar + fork under the gates so its uncommitted files live in the
  correct Lance branch tree. Cross-table conflicts surface as typed read-set or
  manifest conflicts.
- **Deletes stage too (MR-A).** Lance 7.0's
  `DeleteBuilder::execute_uncommitted` (#6658) makes delete a two-phase op,
  so deletes no longer inline-commit. Each delete records a predicate in
  `MutationStaging.delete_predicates`; at end-of-query `stage_all` combines a
  table's predicates into one `stage_delete` (a deletion-vector transaction,
  no HEAD advance) committed through the same `commit_staged` path as writes.
  A predicate matching zero rows stages nothing — no inline residual, and the
  zero-row drift class is closed by construction. The parse-time D₂ rule
  (below) still prevents inserts/updates from coexisting with deletes in one
  query.

This upholds the manifest-atomic mutation and read-your-writes invariants
tracked in [docs/dev/invariants.md](invariants.md).

### D₂ — parse-time mixed-mode rejection

A single mutation query is either insert/update-only or delete-only.
Mixed → rejected at parse time with a clear error directing the user to
split the query. This is a deliberate boundary, not a temporary limitation.
Inserts/updates accumulate as pending batches and deletes as predicates, and
both stage correctly; keeping a single query to one kind means read-your-writes
within that query stays unambiguous (a read never reconciles pending inserts
against same-query delete predicates) and each touched table commits at most one
version. Compose mixed operations by issuing separate atomic mutations (writes,
then deletes), or a branch + merge for one atomic commit. Allowing mixing would
instead require an in-query delete view, pending pruning, and per-table
two-commit ordering in the hot mutation path — complexity this boundary
deliberately avoids.

### MR-793 status (storage trait two-phase invariant) — partial

MR-793 hoists the staged-write pattern into a `TableStorage` trait
surface with sealed-trait enforcement and opaque `SnapshotHandle` /
`StagedHandle` types — see `crates/omnigraph/src/storage_layer.rs`.
The trait is the canonical surface for new engine code; existing call
sites still use the inherent `TableStore` methods (mechanical migration
deferred to a follow-up cycle — tracked).

Three writers have been migrated onto staged primitives:

* **`ensure_indices`** (`db/omnigraph/table_ops.rs::build_indices_on_dataset_for_catalog`)
  — scalar indices (BTree, Inverted) use `stage_create_*_index` +
  `commit_staged`. Which index a `@index`/`@key` property gets is dispatched by
  type via `node_prop_index_kind` (enum + orderable scalar → BTree, free-text
  String → Inverted/FTS, Vector → vector). Vector indices stay inline (residual
  — Lance `build_index_metadata_from_segments` is `pub(crate)` in 6.0.1;
  companion ticket to lance-format/lance#6658 needed). This build is
  existence-gated (it creates a *missing* index over current fragments); folding
  fragments appended afterward into an *existing* index is `optimize`'s
  `optimize_indices` pass — an inline-commit residual, not a staged write (Lance
  exposes no uncommitted index-optimize), covered by the optimize recovery
  sidecar (see [maintenance.md](../user/operations/maintenance.md)).
* **`branch_merge::publish_rewritten_merge_table`**
  (`exec/merge.rs`) — merge_insert now uses `stage_merge_insert` +
  `commit_staged`; its deletes use `stage_delete` + `commit_staged` (MR-A).
* **`schema_apply` rewritten_tables** (`db/omnigraph/schema_apply.rs`)
  — rewrites use `stage_overwrite` + `commit_staged`, including empty-table
  rewrites via a zero-fragment Lance `Operation::Overwrite`.

A defense-in-depth integration test (`tests/forbidden_apis.rs`) walks
engine source and fails if non-allow-listed code calls Lance's
inline-commit APIs directly. The trait surface itself is the primary
enforcement (sealed + only-callable-via-trait once call sites land);
the grep test catches type-system bypass attempts.

The "finalize → publisher residual" described below applies equally to
the migrated writers — Lance has no multi-dataset atomic commit
primitive, so the per-table commit_staged → manifest publish gap is
the same drift class. Closing it requires either upstream Lance
multi-dataset commit OR the omnigraph-side recovery-on-open reconciler
described in `.context/mr-793-design.md` §15 (deferred to MR-795).

### Inline-commit residuals live on `InlineCommitResidual`, not `db.storage()` (MR-793 acceptance §1, by construction)

MR-793's acceptance criterion §1 ("`TableStore` (or successor) public API has no method that performs a manifest commit as a side effect of writing") holds **by construction** after MR-854. `db.storage()` (`&dyn TableStorage`) exposes only staged primitives + reads; the inline-commit writes Lance cannot yet stage live on a separate `InlineCommitResidual` trait reached via `Omnigraph::storage_inline_residual()`. A new engine writer cannot couple a write with a Lance HEAD advance through the default surface — it would have to name the residual accessor explicitly. The dead legacy methods (trait `append_batch` / `merge_insert_batches`, inherent `merge_insert_batch{,es}`, `create_{btree,inverted}_index`) were removed; appends/merges and scalar index builds all use the `stage_*` primitives.

One method remains on `InlineCommitResidual`, named honestly at its call site:

| Residual method | Inline-commit reason | Closes when |
|---|---|---|
| `create_vector_index` | Vector indices take Lance's "segment commit path"; `build_index_metadata_from_segments` is `pub(crate)` (Lance [#6666](https://github.com/lance-format/lance/issues/6666) still open) | Lance #6666 lands and `stage_create_vector_index` joins the staged surface |

`delete_where` used to be the second residual. Lance 7.0's
`DeleteBuilder::execute_uncommitted` ([#6658](https://github.com/lance-format/lance/issues/6658))
made delete a staged write, so MR-A migrated it to `TableStorage::stage_delete`
and removed `InlineCommitResidual::delete_where`. The parse-time D₂ rule is
retained as a deliberate boundary (constructive XOR destructive per query) — see
the D₂ section above.

The `tests/forbidden_apis.rs` guard still catches direct `lance::*` inline-commit misuse outside the storage layer; the trait split makes the staged-only default a type-system guarantee on top of it.

### `LoadMode::Overwrite` uses staged Lance `Overwrite`

The bulk loader's Append, Merge, and Overwrite modes all use the
staged-write path described above. `LoadMode::Overwrite` accumulates
replacement batches in memory, validates node/edge constraints, referential
integrity, and edge cardinality before any Lance HEAD movement, stages
each touched table with Lance `Operation::Overwrite`, then runs
`commit_staged` under the normal `SidecarKind::Load` recovery sidecar
before publishing `__manifest`. `OMNIGRAPH_LOAD_CONCURRENCY` applies to the
fragment-writing stage only; the commit and manifest publish run while holding
the root-shared schema → branch → sorted-table gates. Empty-table overwrite is
represented as a valid zero-fragment Lance `Overwrite` transaction, not as
truncate-then-append.

### Open-time recovery sweep

The staged-write rewire eliminates one drift class **by construction at
the writer layer**: an op that fails before pushing to the in-memory
accumulator (validation errors, missing endpoints, parse-time D₂
rejection) leaves Lance HEAD untouched on every staged table. This is
the case the `partial_failure_leaves_target_queryable_and_unblocks_next_mutation`
test pins.

A second, narrower drift class — the **finalize → publisher window** —
is closed across one open cycle by the open-time recovery sweep:

`MutationStaging::stage_all` prepares the table transactions and `commit_all`
runs their independent HEAD advances before the publisher commits the manifest. Lance has
no multi-dataset atomic commit, so the per-table `commit_staged` calls
are independent operations: if commit_staged on table N+1 fails *after*
commit_staged on tables 1..N succeeded, or if the publisher's CAS
pre-check rejects *after* every commit_staged succeeded, tables 1..N
are left at `Lance HEAD = manifest_pinned + 1`.

**Recovery protocol** (lifecycle of every staged-write writer —
`MutationStaging::commit_all`, `schema_apply::apply_schema_with_lock`,
`branch_merge_on_current_target`, `ensure_indices_for_branch`,
`optimize_all_tables`):

1. **Phase A**: writer writes a sidecar JSON to
   `__recovery/{ulid}.json` BEFORE its first independently durable physical
   effect (including a first-touch Lance branch ref) or HEAD-advancing commit
   (`commit_staged`, or `compact_files` for `optimize_all_tables`,
   which advances the Lance HEAD via a reserve-fragments + rewrite
   commit rather than a staged write). The
   sidecar names every `(table_key, table_path, expected_version,
   post_commit_pin)` it intends to commit + the writer kind +
   actor_id.
   For a first-touch named-branch Mutation/Load table, Phase A is followed by
   target-ref creation and branch-local `stage_*`; the sidecar already carries
   its pre-minted transaction identity.
2. **Phase B**: writer's per-table `commit_staged` loop runs.
   - **Phase-B confirmation:** a `BranchMerge` writer
     advances each table's HEAD by *several* commits (append → upsert →
     delete), so a bare "HEAD moved" is ambiguous — it could be a complete
     publish or one crashed mid-sequence. After the whole per-table loop
     finishes, the writer re-writes the sidecar stamping each pin's
     `confirmed_version` with the exact achieved version, then proceeds to
     Phase C. Schema-v3 Mutation/Load sidecars also confirm: each table must
     match the staged Lance transaction's `(read_version, uuid)`, and the
     sidecar records the exact `SubTableUpdate` plus original lineage intent.
     This is the commit point of the recovery WAL: a crash *after* confirmation
     rolls forward only when the captured branch token still matches; a crash
     *during* Phase B (sidecar still unconfirmed) rolls back. Remaining legacy
     writers don't confirm — their drift is derived state (index coverage,
     compaction) that a partial roll-forward never corrupts.
3. **Phase C**: publisher commits the manifest.
4. **Phase D**: writer deletes the sidecar.

> **Phase letter convention.** Throughout the recovery code, log
> messages, failpoint names (e.g. `branch_merge.post_phase_b_pre_manifest_commit`),
> and the per-writer integration tests, "Phase A/B/C/D" refers
> exclusively to the four-step lifecycle above. The per-table
> staged-write contract (`stage_*` then `commit_staged`, two steps)
> is referred to by those API verbs — never by phase letters — so a
> reader of `recovery.rs`, `failpoints.rs`, or this document only
> encounters phase letters in the per-writer context.

A failure between Phase A and Phase D leaves the sidecar on disk. The
next `Omnigraph::open` (gated on `OpenMode::ReadWrite`) runs the
recovery sweep in `crates/omnigraph/src/db/manifest/recovery.rs`:

- For each sidecar in `__recovery/`, compare every named table's
  Lance HEAD to the manifest pin. Classify per the all-or-nothing
  decision tree (RolledPastExpected / NoMovement / UnexpectedAtP1 /
  UnexpectedMultistep / IncompletePhaseB / InvariantViolation). For a
  `BranchMerge` sidecar, a moved HEAD with no `confirmed_version` classifies
  as `IncompletePhaseB` (a partial multi-commit publish) and forces roll-back;
  with a `confirmed_version`, roll-forward targets exactly that version.
  Schema-v3 Mutation/Load additionally requires `EffectsConfirmed`, the exact
  Lance transaction identity at the confirmed version, the original immutable
  manifest delta, and a matching captured authority token. A changed token is
  rollback-only; an unknown/foreign effect is refused rather than adopted.
  An Armed first-touch intent with no owned transaction is deferred by live
  roll-forward-only healing because another handle may still own it. Quiesced
  full recovery tolerates an absent target ref (either crash before clone or a
  clone-only tree with no `BranchContents`) and force-reclaims that absent-ref
  target idempotently. If an authoritative ref exists, recovery removes it only
  when it is exactly unchanged and no other pending sidecar claims it. With
  competing claims, the current no-effect sidecar discards itself without
  touching the ref; the final survivor owns cleanup/recovery.
  During partial rollback, no-effect refs are removed before the rollback
  outcome is published so a retry cannot strand them. If a legacy live
  path-child blocks that cleanup, rollback returns an error and read-write open
  fails closed; only a sidecar proven to own no table effect may defer cleanup
  while returning an open handle.
- If any table is `InvariantViolation` (Lance HEAD < manifest pinned —
  should be impossible), **abort** with a loud error and leave the
  sidecar on disk for operator review.
- Otherwise, if every table is `RolledPastExpected`, **roll forward**:
  a single `ManifestBatchPublisher::publish` call extends every pin
  atomically. `SchemaApply` sidecars are eligible only when schema-state
  recovery promoted the matching staging files in the same recovery pass;
  otherwise full open-time recovery rolls them back and refresh-time
  recovery leaves them for the next read-write open.
- Otherwise **roll back**: per-table `Dataset::restore` to the
  manifest-pinned table version, then a single `ManifestBatchPublisher::publish`
  of the restored HEAD — symmetric with roll-forward, so `manifest == HEAD`
  after recovery (no residual drift). This convergence is what lets a
  failed-then-retried schema apply succeed instead of failing one version higher
  each iteration. The audit row's `to_version` records the logical
  rolled-back-to version (`manifest_pinned`); the manifest is published at the
  restore commit (`manifest_pinned + 1`, same content).
- After a successful roll-forward or roll-back, an internal
  `_graph_commit_recoveries.lance` row records `recovery_kind`,
  `recovery_for_actor` (the original sidecar's actor), `operation_id`, and
  exact per-table outcomes. A v3 roll-forward publishes the interrupted
  writer's fixed lineage intent, including its original actor; rollback and
  legacy recovery commits use `actor_id = "omnigraph:recovery"`. Ordinary
  commit history is therefore not a complete recovery enumeration, and the
  CLI currently has no public query for the recovery-audit table.
- Sidecar deleted as the final step.

Triggers for the residual: transient Lance write errors during finalize
(object-store retry budget exhaustion, disk full); persistent publisher
contention exceeding `PUBLISHER_RETRY_BUDGET = 5` retries.

**Long-running servers**: the write entry points (`load_as`,
`mutate_as`, `apply_schema_as`, `branch_merge_as`) and
`Omnigraph::refresh` run roll-forward-only recovery in-process
(`recovery::heal_pending_sidecars_roll_forward`) — the common
Phase B → Phase C residual closes on the next write, without a
restart and without an explicit refresh. The heal lists `__recovery/`
(one `list_dir`; empty in the steady state) and, per sidecar, acquires
schema → branch → sorted-table gates that overlap the writer's guarded
sidecar lifetime. RFC-022 mutation/load writers hold the complete order. Branch
merge now holds schema plus source/target branch authority for its whole attempt
and then the all-catalog source/target table envelope; other legacy adapters
serialize through their existing table or schema gate until their own adapter
slices land. The
manager is shared by every
`Omnigraph` handle for one canonical local root identity (relative, absolute,
and symlink aliases converge; object-store/custom schemes stay opaque), so this
also serializes a refresh or separately-opened handle against a live writer instead of rolling its
in-flight sidecar forward from under it (a sidecar whose queues can be
acquired belongs to a writer that finished or died; an existence
re-check after the wait skips the finished case). Lock order is
schema → branch → sorted tables → coordinator, matching the writer effect path.
Enrolled mutation/load attempts and branch merge perform one additional
`list_dir` after acquiring their authority gates; that final check closes the
pre-gate recovery TOCTOU without moving mutation/load validation or staged-file
construction under the gate.
Pinned by the four
`tests/failpoints.rs::*_after_finalize_publisher_failure_heals_without_reopen`
tests (load, mutation, schema apply, branch merge). The maintenance
entries need the heal for more than liveness: without it, a schema
apply re-plans rewrites from the manifest pin and orphans the drifted
Phase-B commit (dropping its rows), and a branch merge publishes the
drift as an unattributed side effect — both while the stale sidecar
lingers to misclassify later.
Sidecars that would require a `Dataset::restore` (mixed / unexpected
state) are deferred to the next `OpenMode::ReadWrite` open. Full open-time
recovery uses the same root-scoped ordered gates and post-wait existence check,
so it cannot Restore/delete under a live writer owned by another handle in the
same process. Restore remains unsafe across processes because Lance's
`check_restore_txn` accepts
the restore against in-flight Append/Update/Delete commits and
silently orphans them (pinned by
`tests/staged_writes.rs::lance_restore_loses_to_concurrent_append_via_orphaning`).
When such a deferred sidecar blocks a write, the commit-time drift
guard says so explicitly ("a pending recovery sidecar requires
rollback — reopen the graph read-write") instead of pointing at
`omnigraph repair`, which refuses while a sidecar is pending.
`cleanup` refuses pending sidecars at entry as well, before orphan reconciliation
or version GC: v3 ownership and compensation recovery may need the retained
Lance transaction/version history, so garbage collection cannot outrun the
recovery barrier.
Continuous in-process recovery for the rollback path is the goal of a
future background reconciler. `ensure_indices` does not heal at entry itself;
it is an explicit maintenance/reconciliation call, separate from mutation,
load, and schema apply, and its strict preconditions fail loudly on drift.

For enrolled mutation/load, the publisher rechecks the attempt's exact native
branch identity and `graph_head` as well as the touched-table versions. A
concurrent graph commit anywhere on the target branch therefore invalidates the
prepared authority instead of silently reparenting it. Before effects, an
insert-only mutation or Append/Merge load fully reprepares with a bounded retry; strict
Update/Delete/Overwrite returns `ReadSetChanged`; after any effect, any later
error returns `RecoveryRequired` and leaves the fixed v3 intent durable. Legacy
writers still arbitrate only their explicit touched-table expectations until
their adapters are enrolled.

**Sidecar I/O failure semantics** (all sidecar I/O goes through the
backend-generic `StorageAdapter`; the contracts below are pinned by the
storage-fault failpoints `recovery.sidecar_{write,delete,list}` /
`recovery.record_audit` and their tests in `tests/failpoints.rs` and
`tests/recovery.rs`):

- **Phase A put fails** (S3 PutObject / fs write): the writer aborts
  before its first HEAD-advancing commit — no sidecar, no drift,
  nothing to recover; a transient fault never wedges later writes.
- **Phase D delete fails** (S3 DeleteObject): swallowed with a warning —
  the write already published, so failing the caller would report an
  error for a durable write. The stale sidecar is consumed by the next
  write's entry heal (or the next open) via the stale-sidecar
  audit-recovery path, recorded as `RolledForward`.
- **`__recovery/` list fails** (S3 ListObjectsV2): loud at every
  consumer — the write-entry heal fails the write, the open-time sweep
  fails the open. Silently skipping recovery would be consumer
  tolerance of drift.
- **Corrupt / unparseable sidecar**: refused loudly by heal and open
  alike; the file stays on disk for operator inspection (read-only
  opens still work — the sweep is skipped there).
- **Audit append fails after a roll-forward publish**: that recovery
  attempt errors and keeps the sidecar; re-entry sees the
  already-published manifest, records exactly one `RolledForward`
  audit row, and deletes the sidecar (the retry tolerance documented
  on `record_audit`).

Backend notes (the adapter is one implementation over `object_store`
for every backend): local writes stage through `name#<digits>` temp
files that the backend filters from listings and refuses to address —
crash residue of that shape is invisible to the sweep, harmless, and
reclaimed by `delete_prefix`/manual cleanup. Storage errors are
backend-wrapped text without a typed NotFound discriminant — callers
that need missing-vs-error (the cluster store) probe `exists()` first.
`exists()` itself is object-store semantics everywhere: only objects
(or non-empty prefixes) exist, and a permission failure is a loud
error, not a silent `false`.

## Conflict shape

For mutation/load, a changed authority detected before effects is
`ManifestConflictDetails::ReadSetChanged { member, expected, actual }`.
Retryable Insert/Merge/Append attempts handle this internally by fully
repreparing; strict writes surface **409 Conflict** with structured
`read_set_conflict` details. A changed authority discovered after a physical
effect, or any unresolved overlapping intent found at the synchronous recovery
barrier, is `OmniError::RecoveryRequired { operation_id, … }`, mapped to **503
Service Unavailable** with structured `recovery_required`; retry only after the
sidecar has been resolved. Legacy, not-yet-enrolled writers may still surface
`ExpectedVersionMismatch` and `manifest_conflict`.

## Commit actor history

`actor_id` lands in the graph commit lineage — the `graph_commit` rows in
`__manifest`, written in the publish CAS (RFC-013 Phase 7; previously
`_graph_commits.lance`). Ordinary commit/actor history is queried via
`omnigraph commit list`. Crash-recovery actions additionally live in the internal
`_graph_commit_recoveries.lance` table described above; that exact recovery log
does not yet have a public CLI query.

## Storage versioning (no in-place migration)

`db/manifest/migrations.rs` is the single place the on-disk `__manifest` shape is
reconciled with what the binary expects. Storage is **strict-single-version** (the
strand model): this binary reads exactly ONE internal-schema version
(`MIN_SUPPORTED == CURRENT == 4`), so there is no in-place migration.

- **Graph creation** stamps `omnigraph:internal_schema_version` at CURRENT, so a
  fresh graph always opens.
- **`Omnigraph::open`** (both read-write and read-only) reads main's stamp before
  the coordinator reads any branch state and calls `refuse_if_stamp_unsupported`:
  a stamp *below* CURRENT is refused with a rebuild-via-export/import message; a
  stamp *above* CURRENT is refused with "upgrade omnigraph". The publisher
  re-checks the stamp on its write path against the branch it targets, with no
  object-store writes, so the check is safe under a read-only open.
- The stamp + `refuse_if_stamp_unsupported` floor is the only seam a future
  in-place migration would re-introduce (re-add a dispatcher and lower
  `MIN_SUPPORTED`). Until a concrete graph demands it, that machinery is
  deliberately absent — see [versioning.md](versioning.md) (the compatibility
  policy) and [the upgrade guide](../user/operations/upgrade.md) (the rebuild
  recipe).

The stamp history (v1 PK-less, v2 unenforced-PK, v3 `__run__*` sweep, v4 lineage
in `__manifest` with the commit-graph tables retired) is recorded on the
`INTERNAL_MANIFEST_SCHEMA_VERSION` doc-comment; only v4 is served. An earlier-stamped
graph is rebuilt via export/import, not migrated in place.

## Mid-query partial failure: closed by MR-794

The pre-MR-794 design had a known limitation: a multi-statement `.gq`
mutation where op-N inline-committed a Lance fragment and op-N+1 then
failed left the touched table at `Lance HEAD = manifest_version + 1`,
blocking the next mutation with `ExpectedVersionMismatch`.

MR-794 (step 1 + step 2+) closed this for inserts/updates **by
construction at the writer layer**: insert and update batches accumulate
in memory; no Lance HEAD advance happens during op execution; one
`stage_*` + `commit_staged` per touched table runs at end-of-query, and
only after every op succeeded. A failed op leaves Lance HEAD untouched
on the staged tables, so the next mutation proceeds normally with no
drift to reconcile.

The cancellation case (future drop mid-mutation) inherits the same
guarantee — the in-memory accumulator evaporates with the dropped task
and no Lance write was ever issued.

Delete-touching mutations now inherit the same guarantee (MR-A). Deletes
accumulate as predicates and stage via `stage_delete` at end-of-query, so a
delete cascade that fails mid-way advances no Lance HEAD — the same
"untouched on failure" property as inserts/updates. The old narrow inline
window (and the retry/`cleanup` workaround it required) is gone. The
parse-time D₂ rule keeps inserts/updates from coexisting with deletes in one
query as a deliberate boundary (see the D₂ section above), so a mutation is
always purely constructive or purely destructive.