Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@protos/index_old.proto`:
- Around line 32-38: Clarify the use_seeds documentation to distinguish absent
on-disk values, which represent disabled seeds for old datasets and are treated
as false by create_seed_writer, from creation-time defaults applied by index
creation logic. Document that newly created indexes may explicitly set the field
to true for variable-length or wider fixed-width types, while present false
disables seeds.

In `@python/python/benchmarks/zone_map_seeds.py`:
- Around line 71-117: Public benchmark helpers gen_int_batches, gen_fsb_batches,
gen_large_binary_batches, and run_scenario lack required API documentation.
Either rename them with a leading underscore and update all call sites, or add
docstrings containing usage examples; for run_scenario, also cross-reference
IndexConfig and Dataset.optimize.optimize_indices().
- Around line 118-119: zone_map_seeds should not recursively delete
deterministic, caller-owned paths under --tmpdir. Update the benchmark setup to
create a unique run directory beneath the user-supplied temporary directory,
pass that directory to the scenarios, and remove or limit the shutil.rmtree call
in the dataset setup so only the benchmark-owned run directory can be deleted.
- Around line 102-124: Update run_scenario to declare a dict[str, float] return
type, and annotate the timings variable as dict[str, float] so the
phase-duration mapping is consistently typed.

In `@rust/lance-index/src/scalar.rs`:
- Line 52: The public seed module lacks a linked usage example demonstrating its
main workflow. Update the module documentation for `seed` to include a doc link
and concise example using `IndexSeedWriter` with `FragmentSeed`, showing how to
construct and use them.

In `@rust/lance-index/src/scalar/registry.rs`:
- Around line 282-328: Document the public hooks create_seed_writer and
update_from_seeds with a concise implementation example in the surrounding trait
documentation. Show how create_seed_writer returns an IndexSeedWriter that
observes writes and serializes seed data, and how update_from_seeds consumes
FragmentSeed values to update the reference index or returns None for scan
fallback.

In `@rust/lance-index/src/scalar/seed.rs`:
- Around line 17-59: Add runnable Rust documentation examples for the exported
SEED_META_KEY_PREFIX, IndexSeedWriter, and FragmentSeed contract. Demonstrate
generating the schema metadata key, implementing or invoking a writer through
observe_batch and finish, and later harvesting a FragmentSeed while inspecting
metadata_value and bytes; ensure the examples compile as doctests and use
suitable minimal test data or mocks.

In `@rust/lance-index/src/scalar/zonemap.rs`:
- Around line 142-146: Add linked Rust doc examples for the public zone-map seed
APIs: document rows_per_zone on ZoneMapIndex, ZoneMapSeedWriter, and
ZoneMapSeedWriter::new with examples demonstrating configuration and writer
usage, including the related APIs around the seed writer implementation.
- Around line 1106-1125: Update test_zone_map_seeds_used_during_update to match
default_use_seeds: replace the default Int32 zone map with a default-enabled
type such as LargeBinary, or configure the index parameters to explicitly enable
seeds, while preserving the seed-write assertion.
- Around line 1106-1125: FixedSizeList seed support is inconsistent because
training rejects the type, the writer hook returns None, and default_use_seeds
excludes it. Update the relevant training validation, writer hook, and
default_use_seeds logic to accept FixedSizeList and enable seeds where
appropriate, while preserving existing nested-type handling for unsupported
types.
- Around line 1370-1476: Persist each zone’s actual length in the seed IPC:
update seed batch construction and deserialization around seed_batch_from_zones
and deserialize_seed to write/read a UInt64 zone_length column and use it for
ZoneBound.length instead of rows_per_zone. Update the round-trip test to cover a
partial final zone and assert its restored length matches the actual row count.
- Around line 1275-1294: In create_seed_writer, stop swallowing
ZoneMapSeedWriter::new construction errors: replace the match that converts Err
into Ok(None) with error propagation using ?. Keep Ok(None) only for nested
data, missing rows_per_zone, or disabled seeds, and return the successfully
constructed writer wrapped in Some.

In `@rust/lance/src/dataset/write.rs`:
- Around line 1383-1390: Move the inline imports from
create_seed_writers—DatasetIndexExt, IndexDetails, fetch_index_details, and
IndexMetadata—to the file-level use declarations, removing them from the
function body while preserving all existing references.
- Around line 1406-1424: Add debug or warning logs before each early continue in
the seed-writer setup loop, especially for failures from schema field-path
lookup, missing field metadata, fetch_index_details, and details.get_plugin.
Include the relevant index or field context and the underlying error where
available, while preserving the existing skip behavior.
- Around line 742-746: Update the cleanup call in the seed-flush error path of
the write flow to pass the required target-base argument: invoke
cleanup_data_fragments with cleanup_bases.as_deref() between base_dir and
fragments, matching the loop-error cleanup path and ensuring routed target-base
files are removed.
- Around line 4616-4675: The zone-map seed test currently uses Int32, which does
not enable seeds by default. In the test around gen_batch and the WriteParams
append, either change the val column to a seed-eligible wider type or set
use_seeds: Some(true) in WriteParams, keeping the metadata assertion valid.
- Around line 4590-4592: The test function `test_multi_base_target_all_bases` is
missing its closing brace; add `}` immediately after the `assert_eq!(file_bases,
vec![None, Some(1), Some(2)])` assertion and before the `#[tokio::test]`
declaration for `test_zone_map_seeds_used_during_update`.

In `@rust/lance/src/index/append.rs`:
- Around line 244-246: Replace the bare unwrap in the metadata parsing logic
with an explicit let-else or equivalent pattern that returns Ok(None) when
split(':') yields no value, then parse the extracted segment as u32 while
preserving the existing fallback behavior.
- Around line 206-231: Update the fragment seed lookup to inspect all entries in
`fragment.files` and select the data file containing the indexed column’s seed
metadata instead of always using `files.first()`. Preserve the fallback when no
suitable file is found, and add debug or warning logs for scheduler open and
FileReader initialization failures so I/O errors are observable. Use the
surrounding seed-reading logic and its indexed-column symbols to identify the
required metadata.

---

Outside diff comments:
In `@rust/lance-index/src/scalar/zonemap.rs`:
- Around line 457-507: Update every remaining direct call to the changed scalar
index loading or construction API in the test module, especially the callers
near the referenced locations, to pass the new required parameter consistently
with the updated signature. Use the surrounding test setup and existing
production call sites to determine the correct argument value, ensuring all
affected tests compile.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In `@rust/lance/src/dataset/write.rs`:
- Around line 601-611: Restrict do_write_fragments to pub(crate) because it is
an internal helper rather than intended public API. Keep its existing behavior
and signature otherwise unchanged, including the seed_writers parameter.

In `@rust/lance/src/index/append.rs`:
- Around line 211-218: Update the seed-file handling around scheduler.open_file
to resolve both the target store and file path from data_file.base_id, rather
than always joining dataset.base with DATA_DIR and data_file.path. Support base
IDs that refer to non-dataset-root bases, then open the resolved path with the
corresponding scheduler before retaining the existing Ok(None) fallback.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In `@rust/lance-index/src/scalar/zonemap.rs`:
- Around line 1414-1420: Move the arrow_ipc::reader::FileReader and
std::io::Cursor imports from inside deserialize_seed to the module-level import
section, then keep deserialize_seed unchanged otherwise.
- Around line 3505-3509: Move the test-only imports from the function bodies of
test_zone_map_seed_writer_round_trip and the other newly added tests into the
enclosing test module’s import block. Consolidate the referenced
IndexSeedWriter, ZoneMapSeedWriter, ArrayRef, and Int32Array imports at module
scope, removing the duplicate local use declarations.
- Around line 1469-1483: The zone reconstruction loop creating ZoneMapStatistics
must reject malformed persisted bounds before pushing them. In the loop around
zone_start and zone_length_col.value, use checked multiplication for the zone
start, a fallible conversion for the persisted length, and validate that length
is greater than zero and no larger than rows_per_zone; return contextual errors
for overflow, conversion failure, or invalid lengths.

In `@rust/lance/src/dataset/write.rs`:
- Around line 1407-1408: Update the seed-writer vector initialization in the
dataset index-writing flow to use Vec::with_capacity(indices.len()), leveraging
the count loaded by dataset.load_indices() and preserving the existing writer
type and behavior.
- Around line 601-611: Narrow do_write_fragments to pub(crate) if it is not
intended for external callers; otherwise add rustdoc documenting seed_writers
ownership, observation, and flush behavior, including a compiling example.
Preserve the existing write behavior and signature semantics.
- Around line 4684-4708: Strengthen the integration test around
Dataset::optimize_indices so it verifies seed harvesting was selected rather
than merely validating the resulting ZoneMap. Add an available seam, metric, or
failure-injection assertion before or during optimization, and assert that the
seed-update path is exercised while preserving the existing index correctness
checks.
- Around line 4599-4611: Move the imports currently declared inside
test_zone_map_seeds_used_during_update to the surrounding test module’s import
block, removing the redundant function-local use declarations while preserving
the test’s behavior.