* test(inkless:retention): add earliest_batch_timestamp write-path benchmark

Add a schema-agnostic write-path benchmark (@tag("benchmark"), excluded
from CI) that drives only the commit/delete Job APIs and never references
the new column, so the same test compiles and runs both before and after
the column exists.

It is committed first, ahead of the feature, on purpose: its parent is
the pre-column state, so the maintenance-cost A/B is reproducible by
checkout alone -- run it on this commit for the "before" numbers and on
the feature commit for the "after" -- with no cherry-pick or separate
worktree. Shapes: steady-state append (recompute guard skips),
commit-into-empty, and delete-advance.

* feat(inkless:retention): track earliest batch timestamp per log

Add logs.earliest_batch_timestamp: the effective timestamp (see
batch_timestamp) of the batch at log_start, i.e. the oldest retained
batch of the partition. This lets a future retention check decide "time
retention deletes nothing" from a single column instead of scanning the
whole partition, for the common steady-state case where the oldest batch
is still within retention. No reader uses the column yet; this only adds
and maintains it.

The column is maintained under the existing logs FOR UPDATE in every
write path that changes the oldest batch, sourced from the batches row
(so it matches what a coalesced row stores), never from the request:
- delete_records_v1 recomputes it when a delete advances log_start, or
  NULLs it when the log becomes empty;
- commit_file_v1 and commit_file_v2 populate it when a partition in the
  commit has no value yet. This covers empty->non-empty (the just
  committed batch is the oldest) and lazily backfills pre-existing NULL
  logs; an IS NULL guard keeps steady-state appends free;
- prune_batches_below_highest_tiered_offset_v1 (cross-tier pruning also
  removes the oldest batches and advances log_start) recomputes it when
  the prune deleted something, or NULLs it when the log becomes empty.

NULL means "unknown, must scan", so there is no backfill over the batches
table at migration time. Both commit functions are maintained because
batch coalescing defaults off, making commit_file_v1 the default
production commit path.

Regenerates jOOQ sources (new column + timestamp_nullable_t domain, and
the schema-version stamp bump) and updates existing tests for the new
LogsRecord constructor arity. Adds LogsEarliestBatchTimestampTest, which
covers the invariant across every write path (commit into empty,
CreateTime vs LogAppendTime, delete/prune advance recompute, delete-all
and prune-all reset to NULL, no-op delete/prune, and commit backfill of a
pre-existing NULL log); it lives with the feature because it references
the new column and asserts the maintenance this commit introduces.

* test(inkless:retention): add earliest_batch_timestamp write-path benchmark

Add a schema-agnostic write-path benchmark (@tag("benchmark"), excluded
from CI) that drives only the commit/delete Job APIs and never references
the new column, so the same test compiles and runs both before and after
the column exists.

It is committed first, ahead of the feature, on purpose: its parent is
the pre-column state, so the maintenance-cost A/B is reproducible by
checkout alone -- run it on this commit for the "before" numbers and on
the feature commit for the "after" -- with no cherry-pick or separate
worktree. Shapes: steady-state append (recompute guard skips),
commit-into-empty, and delete-advance.

* feat(inkless:retention): track earliest batch timestamp per log

Add logs.earliest_batch_timestamp: the effective timestamp (see
batch_timestamp) of the batch at log_start, i.e. the oldest retained
batch of the partition. This lets a future retention check decide "time
retention deletes nothing" from a single column instead of scanning the
whole partition, for the common steady-state case where the oldest batch
is still within retention. No reader uses the column yet; this only adds
and maintains it.

The column is maintained under the existing logs FOR UPDATE in every
write path that changes the oldest batch, sourced from the batches row
(so it matches what a coalesced row stores), never from the request:
- delete_records_v1 recomputes it when a delete advances log_start, or
  NULLs it when the log becomes empty;
- commit_file_v1 and commit_file_v2 populate it when a partition in the
  commit has no value yet. This covers empty->non-empty (the just
  committed batch is the oldest) and lazily backfills pre-existing NULL
  logs; an IS NULL guard keeps steady-state appends free;
- prune_batches_below_highest_tiered_offset_v1 (cross-tier pruning also
  removes the oldest batches and advances log_start) recomputes it when
  the prune deleted something, or NULLs it when the log becomes empty.

NULL means "unknown, must scan", so there is no backfill over the batches
table at migration time. Both commit functions are maintained because
batch coalescing defaults off, making commit_file_v1 the default
production commit path.

Regenerates jOOQ sources (new column + timestamp_nullable_t domain, and
the schema-version stamp bump) and updates existing tests for the new
LogsRecord constructor arity. Adds LogsEarliestBatchTimestampTest, which
covers the invariant across every write path (commit into empty,
CreateTime vs LogAppendTime, delete/prune advance recompute, delete-all
and prune-all reset to NULL, no-op delete/prune, and commit backfill of a
pre-existing NULL log); it lives with the feature because it references
the new column and asserts the maintenance this commit introduces.