- Implement Document::extractAllParallelized() using std::async
to distribute pages across threads with per-thread Document
instances, achieving true parallelism (not the old fake alias)
- zyp_plugin.cpp now uses extractAllParallelized() by default
for PDF extraction during ingestion
- Add zyp_extraction_bench standalone benchmark measuring
sequential vs 2/4/N-thread extraction throughput (pages/s, MB/s)
Build: ninja bench_zyp
Run: build/debug/plugins/zyp/zyp_extraction_bench <pdf>
- Add IContentExtractor::extractTextAndMetadata() to carry plugin metadata
- Add extractDocumentContent() in extraction_util for text+metadata
- PIQ uses plugin-provided title from PDF metadata, skipping GLiNER
entirely when a title already exists (e.g. from zyp PDF parsing)
- Fix misleading 'parallel' naming in zyp wrapper/comments
(zpdf_extract_all_parallel is actually sequential)
This removes the dominant slow-write source for paperbridge PDF
ingestion — title metadata batches that previously backed up the
WriteCoordinator with 34s queue waits.
