feat(engine): Stage the delete path; retire the inline-delete residual (#308)

* test(engine): pin zero-row cascade delete must not drift an edge table (red)

A delete <Node> cascades a delete_where into every incident edge type. The
inline delete_where (Dataset::delete) advances Lance HEAD even when zero edges
match, but the cascade records the new version only if deleted_rows > 0 — so a
node with no incident edges leaves edge:Knows HEAD>manifest drift, which trips
the next strict write's ExpectedVersionMismatch and repair refuses it.

Red today: edge:Knows manifest=v5, Lance HEAD=v6. Goes green when delete moves
to the staged two-phase path (iss-950, Lance 7.0 DeleteBuilder::execute_uncommitted),
where a 0-row delete commits no Lance version and the deleted_rows>0 gate becomes
correct by construction.

* fix(engine): a zero-row delete must not advance Lance HEAD

Lance's Dataset::delete commits a new version even when the predicate matches
nothing (build_transaction always emits Operation::Delete), so a node delete
that cascades a delete_where into an incident edge type with no matching edges
advanced that edge table's Lance HEAD while the cascade skipped record_inline
(gated on deleted_rows > 0) — leaving HEAD>manifest drift that wedged the next
strict write and that repair refused as suspicious/unverifiable.

Use Lance 7.0's two-phase DeleteBuilder::execute_uncommitted to read
num_deleted_rows before committing: a no-match delete now advances nothing (no
version, no drift) and the existing deleted_rows>0 gate is correct by
construction. Non-zero deletes commit the staged transaction with
skip_auto_cleanup + affected_rows (parity with the prior inline path).

First step of the staged-delete migration (iss-950); turns the
node_delete_with_no_incident_edges_leaves_no_edge_table_drift regression green.

* feat(engine): stage_delete two-phase primitive (MR-A step 0)

Add TableStore::stage_delete (Lance 7.0 DeleteBuilder::execute_uncommitted),
the two-phase analogue of stage_merge_insert: writes deletion files without
advancing Lance HEAD, returns Option<StagedWrite> (None on 0 rows = true no-op),
carrying the deletion-vector updated_fragments as new_fragments and the
superseded originals as removed_fragment_ids so combine_committed_with_staged
makes the deletion visible to in-query reads.

No affected_rows is threaded: like stage_merge_insert's Operation::Update commit,
the staged delete relies on OmniGraph's per-table write queue + manifest CAS, not
Lance's per-dataset conflict resolver (commit_staged is a single attempt).

Flip the two residual guards to the staged path: staged_writes.rs now asserts
stage_delete does NOT advance HEAD and that a staged delete is read-your-writes
visible (the deletion-vector RYW proof D2 retirement depends on); the
lance_surface_guards delete guard pins execute_uncommitted's UncommittedDelete.

No behavior change yet (callers still use delete_where); Step 1 wires them.

* feat(engine): TableStorage::stage_delete + migrate merge delete path (MR-A step 1a)

Add stage_delete/Option<StagedHandle> to the TableStorage trait (delegates to
TableStore::stage_delete). Migrate the two branch_merge delete sites
(three-way RewriteMerged + adopt delta) from the inline delete_where residual to
stage_delete + commit_staged — identical in shape to the stage_merge_insert +
commit_staged pair above each. HEAD still advances within the merge sequence
(via commit_staged), under the unchanged SidecarKind::BranchMerge Phase-B
confirmation; the _pre_delete/_pre_index failpoints fire by position, unchanged.

merge_truth_table, branching, composite_flow green.

* feat(engine): migrate all delete sites to staged path, retire inline delete (MR-A step 1b/1c)

Routes every delete through the staged write path so delete never advances
Lance HEAD inline — the last inline-commit residual on the mutation path is
gone. `MutationStaging` now accumulates delete predicates (`record_delete`)
alongside pending write batches; at end-of-query `stage_all` combines a
table's predicates into one `(p1) OR (p2) …` `stage_delete` (a deletion-vector
transaction, no HEAD advance) and `commit_all` commits it through the same
`commit_staged` path as inserts/updates. Deletes are now ordinary staged
entries: one sidecar pin at `expected + 1`, no inline special-casing.

Migrated callers (all 5): the 3 mutation.rs sites (delete-node, cascade,
delete-edge) and the 2 merge.rs sites (already on stage_delete in step 1a).
`affected_edges`/`affected` move from post-inline-commit `deleted_rows` to a
committed `count_rows` at record time — exact under D₂, bounded by the cascade
working set. A predicate matching zero rows stages nothing (the staged
equivalent of the old "skip record_inline on 0 deleted rows"), so the zero-row
edge-table drift class stays closed by construction.

Retired scaffolding now that no caller remains:
- `MutationStaging.inline_committed` + `record_inline` → `delete_predicates` +
  `record_delete`; `StagedMutation.inline_committed`/`paths` fields and all the
  `commit_all` inline handling (queue keys, sidecar pins with the
  `record_inline` table_version special-case, the inline recheck loop).
- `open_table_for_mutation`'s post-inline-commit reopen branch (deletes no
  longer advance HEAD mid-query, so a second touch reopens at the pinned
  version like any write).
- `InlineCommitResidual::delete_where` + its `TableStore` impl, the orphaned
  `TableStore::delete_where`, and `DeleteState`. `InlineCommitResidual` now
  carries only `create_vector_index` (Lance #6666 still open).

D₂ stays for now: staged-delete read-your-writes doesn't yet compose into the
pending accumulator (insert-then-delete on one table), so mixed
insert/update/delete in one query is still rejected at parse time. Retiring D₂
is step 2. Doc comments updated to match across exec/, storage_layer, db/.

Tests (all green): writes, consistency, validators, end_to_end, composite_flow,
merge_truth_table, maintenance, recovery, staged_writes, forbidden_apis,
lance_surface_guards, changes, point_in_time (286), plus failpoints (63).

* docs: delete is a staged write, not an inline-commit residual (MR-A step 1)

Update the docs that described `delete` as the inline-commit residual now that
MR-A routes it through `stage_delete`. Always-loaded surfaces (AGENTS.md rule
4 / capability matrix, invariants.md Invariant 4 / truth matrix / known gaps)
plus the dev write-path docs (writes.md, execution.md incl. its mutation
sequence diagram, architecture.md) now state: deletes accumulate as predicates
and stage like inserts/updates, no inline HEAD advance; `InlineCommitResidual`
carries only `create_vector_index` (Lance #6666). The parse-time D₂ rule is
documented as retained — not because delete inline-commits, but because
staged-delete read-your-writes is not yet wired into the pending accumulator
(MR-A step 2). lance.md's 7.0 audit note marked MR-A as landed.

* docs: D₂ is a deliberate boundary, not temporary scaffolding (MR-A close-out)

After MR-A staged the delete path, D₂ (a mutation query is insert/update-only
OR delete-only) was left framed as temporary — "until Lance ships two-phase
delete" / "retire in step 2". Lance shipped that and we used it for the
inline-commit fix; D₂'s original justification is gone. It now stands for a
different, permanent reason: keeping a query to one kind keeps its
read-your-writes unambiguous and each table to one version per query. Retiring
it would buy single-commit mixed atomicity (cheap workaround: split, or a
branch) at the cost of an in-query delete view, pending pruning, edge
id-resolution, and two-commit-per-table ordering in the hot mutation path —
complexity not worth earning. Decision: keep D₂ as a deliberate boundary.

Reframes the now-stale wording everywhere, no logic change:
- The D₂ parse-time error message no longer promises "this restriction lifts
  when Lance exposes a two-phase delete API"; it states the boundary and points
  to a branch+merge for one atomic commit.
- `enforce_no_mixed_destructive_constructive` doc, AGENTS.md, invariants.md
  (Invariant 4 / truth matrix / removed from the known-gaps), writes.md,
  architecture.md, lance.md, and the user mutations doc (which wrongly said
  deletes "commit through a different path" — both stage now).
- Swept remaining stale `delete_where` mentions left from the Step-1 migration:
  the merge.rs "swap when upstream ships" comments (already swapped), the
  forbidden_apis / table_ops residual notes, the staged_writes vector-index
  guard doc (was "same as stage_delete's absence" — stage_delete now exists),
  and test comments/assert messages in recovery/maintenance/writes/failpoints.
  Genuinely-historical records (dated Lance audit, rfc-013, bug-case-fix) left.

Verified: engine builds warning-free; check-agents-md OK; writes/maintenance/
recovery/staged_writes/forbidden_apis all green. Closes MR-A.

* test(engine): overlapping delete predicates must not double-count affected_* (red)

Reproduces a reporting regression from the staged-delete migration flagged in
PR #308 review. Because deletes now stage (instead of inline-committing), two
delete statements in one query both scan the same unchanged committed snapshot;
counting each predicate independently over-reports `affected_*` when they
overlap. The old inline path committed each delete before the next ran, so it
counted distinct.

`delete Person where name = "Alice"` then `delete Person where age > 29` over
the standard fixture (Alice 30, Charlie 35) removes 2 distinct nodes and 3
distinct edges, but the buggy per-statement counting returns 3 nodes / 6 edges.
RED at this commit (asserts left=3, right=2).

* fix(engine): dedup overlapping delete predicates when counting affected_*

Count each delete statement against the committed snapshot MINUS the predicates
a prior delete statement on the same table already recorded:
`(pred) AND NOT ((prior1) OR (prior2) …)`. Summed over statements this is
inclusion-exclusion — `Σ |pₙ \ (p₁ ∪ …)| = |p₁ ∪ p₂ ∪ …|` — exactly the distinct
count the combined `(p1) OR (p2)` staged delete removes. Works for nodes and
edges alike with no edge identity needed; the node ID scan uses the same
exclusion so a later statement also doesn't re-cascade already-deleted nodes.
The ORIGINAL predicate is still what gets recorded (the staged delete removes
the union); only the count uses the exclusion. The common single-delete path is
unchanged (`prior` empty → filter is just the base predicate).

New helper `dedup_delete_filter` + `MutationStaging::recorded_delete_predicates`.
Turns the red regression test green (2 nodes / 3 edges); writes (33),
end_to_end, validators, maintenance, recovery, composite_flow, merge_truth_table,
consistency, changes, and failpoints (63) all stay green.

* test(engine): delete dedup must not drop NULL-column rows (red)

Follow-up to the overlapping-delete fix flagged in PR #308 review (Greptile P1):
the `(base) AND NOT (prior)` exclusion breaks under SQL three-valued logic. If a
prior delete predicate references a NULLable column, a later statement's
matching row whose column is NULL makes `prior` evaluate to UNKNOWN, `NOT
UNKNOWN` is UNKNOWN, and the row is filtered out of the scan — even though the
prior delete never matched it. That drops it from `deleted_ids`, skipping its
cascade (orphaned edges) or, if it is the only match, leaving the node
undeleted. A data bug, not just a miscount.

Data: Charlie(age 35), Zoe(age NULL); Knows Zoe→Charlie. `delete Person where
age > 30` then `delete Person where name = "Zoe"`. Under the buggy `NOT`, Zoe's
scan `(name='Zoe') AND NOT (age>30)` is UNKNOWN → Zoe survives. RED at this
commit (Person count left=1, right=0).

* fix(engine): NULL-safe delete dedup — exclude only definitely-matched prior rows

Change `dedup_delete_filter` from `(base) AND NOT (prior)` to
`(base) AND ((prior) IS NOT TRUE)`. `IS NOT TRUE` keeps both FALSE and UNKNOWN
rows, so a prior predicate that evaluates to SQL UNKNOWN (a NULL in a referenced
column) no longer drops a row this statement legitimately matches — only rows a
prior predicate matched as definitely TRUE are excluded from the count/scan. The
distinct-count semantics are unchanged for non-NULL data.

Turns the red NULL-dedup test green (Zoe deleted, her edge cascaded), and the
overlapping-dedup + writes/end_to_end/validators/maintenance/recovery/
composite_flow/consistency suites stay green.

* docs(engine): note dedup_delete_filter's load-bearing dependency on D₂

Self-review follow-up: the overlapping-delete dedup assumes the committed
snapshot is invariant across a query's statements, which holds only because D₂
forbids mixing writes with deletes (so a delete-touched table has no pending
writes). Make that dependency explicit at the function so a future D₂ relaxation
is forced to revisit the dedup. Comment-only.

* Preserve staged write commit metadata

AGENTS.md

@@ -32,7 +32,7 @@ OmniGraph is a typed property-graph engine built as a coordination layer over ma
 - **Languages**: a `.pg` schema language and a `.gq` query language, both Pest-based, with a typed IR.
 - **Multi-modal querying**: vector ANN (`nearest`), full-text (`search`/`fuzzy`/`match_text`/`bm25`), Reciprocal Rank Fusion (`rrf`), and graph traversal (`Expand`, anti-join `not { … }`) in one runtime.
 - **Branches and commits across the whole graph**: Git-style — every successful publish appends to a commit DAG; merges are three-way at the row level.
-- **Atomic per-query writes**: `mutate_as` and `load` accumulate insert/update batches into an in-memory `MutationStaging.pending` per touched table; one `stage_*` + `commit_staged` per table runs at end-of-query, then `ManifestBatchPublisher::publish` commits the manifest atomically with per-table `expected_table_versions` CAS. A mid-query failure leaves Lance HEAD untouched on staged tables — no drift, no run state machine, no staging branches. Deletes still inline-commit; D₂ at parse time prevents inserts/updates and deletes from coexisting in one query.
+- **Atomic per-query writes**: `mutate_as` and `load` accumulate insert/update batches into an in-memory `MutationStaging.pending` per touched table; one `stage_*` + `commit_staged` per table runs at end-of-query, then `ManifestBatchPublisher::publish` commits the manifest atomically with per-table `expected_table_versions` CAS. A mid-query failure leaves Lance HEAD untouched on staged tables — no drift, no run state machine, no staging branches. Deletes stage through the same path (MR-A: `stage_delete` via Lance 7.0 `DeleteBuilder::execute_uncommitted`), so they no longer advance Lance HEAD inline. D₂ at parse time is a deliberate boundary — one mutation query is constructive (insert/update) XOR destructive (delete) — so read-your-writes within a query stays unambiguous and each table commits at most one version; compose mixed operations via separate mutations, or a branch for single-commit atomicity.
 - **HTTP server**: Axum + utoipa OpenAPI, bearer auth (SHA-256 hashed, optional AWS Secrets Manager). Cedar policy enforcement is engine-wide — every `_as` writer calls `Omnigraph::enforce(action, scope, actor)`, so HTTP, CLI, and embedded SDK consumers all hit the same gate. **Cluster-only boot** (RFC-011): the server always boots from a cluster directory (`--cluster <dir | s3://…>`, RFC-005) and serves N graphs (N ≥ 1) under multi-graph routes (`/graphs/{graph_id}/...` + read-only `GET /graphs` enumeration); there are no single-graph flat routes and no positional-URI boot. Per-graph + server-level Cedar policies. Runtime add/remove (`POST /graphs`, `DELETE /graphs/{id}`) is not exposed — operators run `cluster apply` and restart.
 - **CLI** with two-surface config (RFC-007/008): the team-owned cluster directory (`cluster.yaml`) plus the per-operator `~/.omnigraph/config.yaml` (servers, clusters, credentials, actor, profiles, aliases, defaults). Graphs are addressed via `--store`/`--server`/`--cluster`/`--profile`/operator defaults (RFC-011). Multi-format output (json/jsonl/csv/kv/table).
 
@@ -250,7 +250,7 @@ omnigraph policy explain --cluster ./company-brain --graph knowledge --actor act
 | Columnar storage on object store | ✅ Arrow/Lance | URI normalization, S3 env-var plumbing |
 | Per-dataset versioning + time travel | ✅ | `snapshot_at_version`, `entity_at`, snapshot-pinned reads across many tables |
 | Per-dataset branches | ✅ | **Graph-level** branches (atomic across all sub-tables), lazy fork, system branch filtering |
-| Atomic single-dataset commits | ✅ | **Multi-table publish via three layers**, NOT a single Lance primitive: (1) per-table Lance `commit_staged` for the data write, (2) `__manifest` row-level CAS via `ManifestBatchPublisher` for cross-table ordering, (3) the open-time recovery sweep for the residual gap between (1) and (2). All three layers ship; the five migrated writers (`MutationStaging::finalize`, `schema_apply`, `branch_merge`, `ensure_indices`, `optimize_all_tables`) write a `__recovery/{ulid}.json` sidecar before Phase B and delete it after Phase C. The next `Omnigraph::open` (gated on `OpenMode::ReadWrite`) runs the sweep in `db/manifest/recovery.rs`: classify, decide all-or-nothing per sidecar, roll forward via single `ManifestBatchPublisher::publish` or roll back via `Dataset::restore` followed by a manifest publish of the restored version (so both directions converge to `manifest == HEAD` — no residual drift), and record an audit row in `_graph_commit_recoveries.lance` (queryable via `omnigraph commit list --filter actor=omnigraph:recovery`). The write entry points (`load_as`, `mutate_as`, `apply_schema_as`, `branch_merge_as`) and `refresh` additionally run an in-process roll-forward-only heal (serialized against live writers via the per-table write queues), so a long-lived server converges on its next write without restart; only rollback-eligible sidecars still defer to the next read-write open (a future background reconciler's goal). Engine writes route through a sealed `TableStorage` trait (`db.storage()`) exposing only `stage_*` + `commit_staged` + reads; the inline-commit residuals (`delete_where`, `create_vector_index`) are split onto a separate sealed `InlineCommitResidual` trait reached via `db.storage_inline_residual()` (MR-854), so the default surface cannot couple a write with a HEAD advance — §1 holds by construction. `delete_where` and `create_vector_index` stay inline until upstream Lance ships a public two-phase API ([#6658](https://github.com/lance-format/lance/issues/6658), [#6666](https://github.com/lance-format/lance/issues/6666)); `LoadMode::Overwrite` uses Lance `Overwrite` staged transactions. |
+| Atomic single-dataset commits | ✅ | **Multi-table publish via three layers**, NOT a single Lance primitive: (1) per-table Lance `commit_staged` for the data write, (2) `__manifest` row-level CAS via `ManifestBatchPublisher` for cross-table ordering, (3) the open-time recovery sweep for the residual gap between (1) and (2). All three layers ship; the five migrated writers (`MutationStaging::finalize`, `schema_apply`, `branch_merge`, `ensure_indices`, `optimize_all_tables`) write a `__recovery/{ulid}.json` sidecar before Phase B and delete it after Phase C. The next `Omnigraph::open` (gated on `OpenMode::ReadWrite`) runs the sweep in `db/manifest/recovery.rs`: classify, decide all-or-nothing per sidecar, roll forward via single `ManifestBatchPublisher::publish` or roll back via `Dataset::restore` followed by a manifest publish of the restored version (so both directions converge to `manifest == HEAD` — no residual drift), and record an audit row in `_graph_commit_recoveries.lance` (queryable via `omnigraph commit list --filter actor=omnigraph:recovery`). The write entry points (`load_as`, `mutate_as`, `apply_schema_as`, `branch_merge_as`) and `refresh` additionally run an in-process roll-forward-only heal (serialized against live writers via the per-table write queues), so a long-lived server converges on its next write without restart; only rollback-eligible sidecars still defer to the next read-write open (a future background reconciler's goal). Engine writes route through a sealed `TableStorage` trait (`db.storage()`) exposing only `stage_*` + `commit_staged` + reads; the sole inline-commit residual (`create_vector_index`) is split onto a separate sealed `InlineCommitResidual` trait reached via `db.storage_inline_residual()` (MR-854), so the default surface cannot couple a write with a HEAD advance — §1 holds by construction. `delete` migrated to the staged path in MR-A (`stage_delete` via Lance 7.0 `DeleteBuilder::execute_uncommitted`, [#6658](https://github.com/lance-format/lance/issues/6658)); `create_vector_index` stays inline until upstream Lance ships a public two-phase API ([#6666](https://github.com/lance-format/lance/issues/6666)); `LoadMode::Overwrite` uses Lance `Overwrite` staged transactions. |
 | Compaction (`compact_files`) + reindex (`optimize_indices`) | ✅ | `omnigraph optimize` orchestrates over all node/edge tables, bounded concurrency; per table runs `compact_files` **then Lance `optimize_indices`** (folds appended/rewritten fragments back into existing indexes — incremental merge, not retrain) and **publishes the resulting version to `__manifest`** (so the manifest tracks the Lance HEAD — required for reads to observe the work and for schema apply / strict writes to pass their HEAD-vs-manifest precondition), under the per-`(table, main)` write queue with `SidecarKind::Optimize` recovery coverage spanning both ops; **commits even with no compaction work if index coverage is stale**; **refuses on an unrecovered graph**; **skips uncovered HEAD > manifest drift** with `DriftNeedsRepair`; **skips blob-bearing tables** (reported via `TableOptimizeStats.skipped`, not silent; reindex is skipped for them too today), gated on `LANCE_SUPPORTS_BLOB_COMPACTION` until the upstream blob-v2 compaction-decode bug is fixed (see [docs/dev/invariants.md](docs/dev/invariants.md) Known Gaps) |
 | Repair uncovered drift | — | `omnigraph repair` explicitly classifies uncovered table `HEAD > manifest` drift: verified maintenance drift (`ReserveFragments`/`Rewrite`) can be published with `--confirm`; suspicious or unverifiable drift requires `--force --confirm`. Sidecar-covered crash residuals still recover automatically on open. |
 | Cleanup (`cleanup_old_versions`) | ✅ | `omnigraph cleanup` with `--keep` / `--older-than` policy |