Introduce parser, AST, and metadata plumbing for `TTL ... EXPORT TO db.table`.
No background scheduler or per-part TTL info yet — those land in follow-up
commits. The clause is recognised, round-trips through `SHOW CREATE TABLE`,
and the resulting `TTLDescription` is collected into `TTLTableDescription`'s
new `export_ttl` list (exposed via `StorageInMemoryMetadata::getExportTTLs`).

Validation in `TTLTableDescription::getTTLForTableFromAST`:
  * reject two `EXPORT` clauses to the same destination,
  * reject `EXPORT` TTL on a table with no partition key.

The destination-specific override of `TTLDescription::result_column`
(`"_export_" + db + "." + table`) is required so that future per-part TTL
info (keyed by `result_column`) keeps separate clocks per destination.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Stores per-part TTL info under `MergeTreeDataPartTTLInfos::export_ttl`,
keyed by `TTLDescription::result_column`. The map is:

  * populated at write time in `MergeTreeDataWriter` for every TTL
    returned by `getExportTTLs`,
  * recomputed during `MATERIALIZE TTL` and merge-time TTL recompute via
    `TTLCalcTransform` / `TTLTransform` (a new `TTLUpdateField::EXPORT_TTL`
    finalizes into the right map),
  * serialized in JSON under the `"export"` key (mirroring the
    `recompression` entry),
  * propagated across merges through the existing `update` aggregation,
  * surfaced through `hasAnyNonFinishedTTLs` and `checkAllTTLCalculated`
    so old parts that predate the TTL are flagged for `MATERIALIZE TTL`.

Adds the partition-wide helper `getPartitionExportTTLMax`: returns the
max `export_ttl.max` across all parts of a partition, or `nullopt` if
any part is missing the entry (with optional `missing_parts_out` for
the scheduler to log). Deliberate: no on-the-fly evaluation — the user
runs `ALTER TABLE ... MATERIALIZE TTL` to backfill, same UX as moves
and recompression TTLs.

Also pulls `export_ttl` into `hasAnyTableTTL` / `hasOnlyRowsTTL` and the
`getColumnDependencies` TTL-column-set walk.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Align the TTL syntax with `ALTER ... EXPORT PARTITION TO TABLE`: the
keyword is now `EXPORT TO TABLE <db.table>` instead of `EXPORT TO
<db.table>`. Parser, formatter, exception messages, and tests updated to
match.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds an `ExportOrigin` enum (`alter` | `ttl`) to the manifest body so
manifests submitted manually (`ALTER ... EXPORT PARTITION`) can be told
apart from manifests submitted by the upcoming TTL scheduler. Surfaced
as `system.replicated_partition_exports.export_origin
Enum8('alter' = 0, 'ttl' = 1)`. Existing manifests in ZooKeeper that
don't carry the field read back as `alter` for backwards compatibility.

`ttl`-origin manifests are skipped by manifest-TTL eviction: the
background cleanup in `ExportPartitionManifestUpdatingTask` and the
overwrite path in `StorageReplicatedMergeTree::exportPartitionToTable`
both refuse to consider them expired. The existing
`export_merge_tree_partition_force_export` setting still overrides via
the unchanged gate.

The write site keeps the default `ExportOrigin::alter`; the TTL
submitter writes `ExportOrigin::ttl` in a follow-up commit.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds a new query-level setting `export_merge_tree_partition_mark_as_ttl`
(default false). When set on `ALTER ... EXPORT PARTITION`, the resulting
manifest is written with `export_origin = ttl` (same as what the TTL
scheduler will write in a follow-up). The TTL scheduler always sets this
implicitly when it submits.

Enforces the "at most one ttl-origin manifest per (src, dest)"
invariant at submission time: when a ttl-origin manifest is being
created, scan siblings under `<zk_root>/exports/` for an existing
ttl-origin marker at a different `partition_id`. If found at `P_old`,
reject the submission as a back-fill (`new < P_old`) unless
`export_merge_tree_partition_force_export` is set; otherwise
best-effort `tryRemoveRecursive` of the old marker before creating the
new one. Same-key collisions continue to be handled by the existing
block.

A plain `alter` over a ttl marker at a different partition is allowed
without friction — alter manifests coexist with the ttl marker, and
the TTL scheduler will filter by `export_origin = ttl` when reading
its own state.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Extracts the readable-vs-insertable column diff and the partition-key
AST compare from `StorageReplicatedMergeTree::exportPartitionToTable`
into `ExportPartitionUtils::verifyExportDestinationCompatibility`, and
calls it from `TTLTableDescription::getTTLForTableFromAST` for every
`TTLMode::EXPORT` clause when not attaching. The destination is resolved
through `DatabaseCatalog::getTable`, matching the manual
`ALTER ... EXPORT PARTITION` flow (throws `UNKNOWN_TABLE` if missing).

The check is skipped under `is_attach=true` because the destination
table may not yet be loaded at server startup; submission-time
validation in `exportPartitionToTable` still covers that path.

Iceberg destinations skip the partition-key AST compare here; the
existing `verifyIcebergPartitionCompatibility` runs against the runtime
iceberg metadata at submission time.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Introduces `TTLExportScheduler`, a per-`StorageReplicatedMergeTree`
background driver that submits partition exports for tables with
`TTL ... EXPORT TO TABLE db.table`. The scheduler is stateless across
restarts: it reads the latest `export_origin = ttl` manifest from
ZooKeeper on every tick and acts on its status — no manifest →
submit the smallest eligible partition; PENDING → wait; COMPLETED →
walk forward to `partition_id > completed`; FAILED → resubmit with
`force_export=1` after per-partition exponential backoff; KILLED →
idle with a `LOG_WARNING` carrying the recovery recipe.

`submit` classifies outcomes as `Submitted | Transient | Failure` so
ZK CAS races and `UNKNOWN_TABLE` (destination dropped post-DDL) do
not bump backoff, while genuine submission errors do.

Adds three table-level settings used by the scheduler:
`export_merge_tree_partition_ttl_poll_interval_seconds` (default 5),
`export_merge_tree_partition_ttl_min_backoff_seconds` (default 1),
`export_merge_tree_partition_ttl_max_backoff_seconds` (default 60).

The scheduler is not yet wired into the background task pool; that
follows in a separate commit.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Declare the scheduler and its background task next to the other
`export_merge_tree_partition_*` task holders. Under the
`allow_experimental_export_merge_tree_partition` server gate:

- Construct the scheduler and create the `TTLExport` task, logging any
  exceptions from `run` via `tryLogCurrentException`.
- `ReplicatedMergeTreeRestartingThread::tryStartup` activates the task
  alongside the other export tasks; `partialShutdown` deactivates it.
- `alter` calls `ttl_export_task->schedule` when any `MODIFY TTL`
  command is in the alter so newly added EXPORT TTLs take effect
  immediately.
- `TTLExportScheduler::run` reschedules itself with
  `export_merge_tree_partition_ttl_poll_interval_seconds * jitter25`
  on the polling path. Early returns (shutdown, readonly, no EXPORT
  TTL, experimental gate off) intentionally skip the reschedule —
  deactivation, the `alter` hook, and server-level config drive those
  paths instead.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ookeeper

In `exportPartitionToTable`, the cross-partition swap used to delete the
existing ttl-origin marker before the `tryMulti` that creates the new
manifest. Any throw between the delete and the multi (parts.empty,
pending mutations, iceberg compatibility, or a ZK error mid-multi) left
the scheduler with no ttl marker at all, so the next tick of
`TTLExportScheduler` would treat the table as fresh and restart from the
oldest expired partition.

Now:

  - The sibling cache walk collects every stale ttl-origin entry for the
    destination, not just the first one encountered. The freshest by
    `create_time` is used for the back-fill check; the others are kept
    for cleanup. This makes the walk deterministic (no break-on-first-
    match in unspecified iteration order) and self-healing for any
    stragglers a previous cleanup may have missed.
  - The `tryRemoveRecursive` of those stale markers runs after the
    `tryMulti` succeeds. Failures during validation or in the multi
    itself now leave the existing marker intact, so the scheduler keeps
    its high-water mark.
  - The post-multi cleanup is best-effort; a ZK error there at worst
    leaves dead nodes, which the next ttl submission will reap.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Add `export_merge_tree_partition_mark_as_ttl` to the Antalya 26.3 session
block, and the three TTL EXPORT scheduler MergeTree settings
(`export_merge_tree_partition_ttl_poll_interval_seconds`,
`export_merge_tree_partition_ttl_min_backoff_seconds`,
`export_merge_tree_partition_ttl_max_backoff_seconds`) to the 26.3 MergeTree
block, so `02995_new_settings_history` passes.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The previous representation flattened `dst_database` and `dst_table` into a
single `destination_name` string joined by a dot. That made the boundary
ambiguous for quoted table names that legitimately contain a dot — e.g.
``TTL ... EXPORT TO TABLE `a.b` `` round-tripped through the formatter as
`` `a`.`b` `` (database `a`, table `b`), changing semantics on metadata
reattach and sending exports to the wrong destination via
`QualifiedTableName::parseFromString` in the scheduler and the DDL-time
schema-compat check.

Mirror the convention already used by `ASTAlterQuery` for the matching
`ALTER ... EXPORT PARTITION ... TO TABLE` syntax: store `destination_database`
and `destination_name` as two separate strings on `ASTTTLElement` and
`TTLDescription`. The parser fills both via `parseDatabaseAndTableName`; the
formatter emits `backQuoteIfNeed(db) + "." + backQuoteIfNeed(table)` when
qualified, else just the table. Both `parseFromString` call sites use the
fields directly.

Per-part `export_ttl` map is keyed by `TTLDescription::getExportKey`, which
uses `backQuoteIfNeed` for both parts so qualified `(a, b)` and unqualified
single-table `a.b` produce distinct keys.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

`TTLExportScheduler` is only constructed and activated by
`StorageReplicatedMergeTree`; the base `MergeTreeData::exportPartitionToTable`
throws `NOT_IMPLEMENTED`. Without a DDL-time guard, `CREATE TABLE ... ENGINE
= MergeTree ... TTL ... EXPORT TO TABLE` and the matching `ALTER` succeeded
but no background exporter ever submitted manifests — silent no-op.

Throw `NOT_IMPLEMENTED` at parse time:
- CREATE: in `registerStorageMergeTree.cpp` right after `getTTLForTableFromAST`,
  using the local `replicated` bool.
- ALTER: in `MergeTreeData::checkAlterIsPossible`, using virtual
  `supportsReplication`. The check can't live in `checkTTLExpressions`
  because that method is also called from `MergeTreeData`'s constructor,
  where virtual dispatch returns the base-class value.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Replace the full `std::map<String, DataPartsVector>` build + two lookups
with a single pass that fills two `DataPartsVector`s for the new and the
existing-marker partitions directly. `getPartitionExportTTLMax` returns
`nullopt` for an empty vector, so the `find`/`end` guards collapse into the
function call. No semantic change.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>