…y comment

Review findings:

1. ParquetWriterConfig was hardcoded to Default in the executor. If ingest
   uses custom compression, merge output would differ. Now threaded from
   ParquetMergePipelineParams through to the executor.

2. Fixed misleading comment claiming "planner will eventually re-plan"
   on merge failure. In reality, input splits are drained by operations()
   and won't be re-planned until the pipeline restarts with metastore
   re-seeding (not yet implemented — TODO added).

3. Added TODO for fetch_immature_parquet_splits() on pipeline respawn,
   matching the Tantivy MergePipeline pattern.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…riter_config

Review findings addressed:

1. fetch_immature_splits(): on pipeline respawn after crash, queries the
   metastore for published Parquet splits so the planner can re-plan
   merges that were in-flight during the crash. On first spawn, uses the
   initial splits from the IndexingService (same as Tantivy pattern).

2. ParquetWriterConfig threaded from pipeline params to executor so merge
   output uses the same compression as ingest.

3. Fixed misleading "planner will eventually re-plan" comment on merge
   failure — honest about the limitation that failed splits wait for
   respawn re-seeding.

4. Added index_uid to ParquetMergePipelineParams for metastore queries.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Matches the Tantivy MergePipeline pattern:
- ObservableState is now MergeStatistics (was unit type)
- perform_observe() collects counters from uploader + publisher handles
- Tracks generation, num_spawn_attempts, num_ongoing_merges,
  num_uploaded_splits, num_published_splits
- previous_generations_statistics preserved across respawns

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Both Tantivy and Parquet merge scheduling used identical score logic
(prefer merges that reduce more splits for less total bytes). Extracted
the core arithmetic into score_merge(num_splits, total_bytes) and have
both score_merge_operation() and score_parquet_merge_operation() call it.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Per CODE_STYLE.md: comments should convey intent, not implementation.
Added explanations for num_merge_ops lineage, known_split_ids rebuild
heuristic, output dir isolation, empty merge handling, scratch dir
lifetime, permit Drop safety, publisher setter ordering, and feedback
loop guard conditions.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…line

Reads parquet_indexing.sort_fields and parquet_indexing.window_duration_secs
from IndexingSettings when constructing the ingest pipeline's TableConfig
(was hardcoded to defaults).

Adds parquet_merge_policy_from_settings() that converts the config-layer
ParquetMergePolicyConfig to an Arc<dyn ParquetMergePolicy> runtime policy,
paralleling merge_policy_from_settings() for Tantivy.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>