perf: speed up semantic index preparation

[thisisjun786]

Summary

• replace per-parent SQLite lookups with two ordered bulk queries
• reuse one parent-window snapshot across planning, document writing, and live indexing
• bulk-load stored vector metadata and parallelize only in-memory hashing and file preparation
• count candidate chunks with COUNT(*) instead of materializing every chunk through the N+1 hydration path
• preserve atomic semantic generations, sequential SQLite mutations, and deterministic parent and child ordering

Why

katok index previously called all_chunks() only to compute candidate_chunks. Each chunk was then hydrated through several point queries, so index startup scaled as multiple SQLite reads per micro-chunk before any semantic output was written. Parent windows and stored vector metadata also used point-query loops.

The implementation keeps SQLite single-connection and single-writer behavior, loads immutable snapshots in bulk, and uses Rayon only after the data is owned in memory.

Verification

• cargo +1.91.0 fmt --all -- --check - passed
• cargo +1.91.0 clippy --all-targets -- -D warnings - passed
• KATOK_EMBEDDER=local-test cargo +1.91.0 test --all-targets - passed, 170 tests
• cargo +1.91.0 publish --dry-run - passed; upload correctly aborted because this was a dry run
• python3 scripts/verify_release_config.py - passed
• GitHub Actions CI run 31572087100 - passed
• stale generation reuse, orphan removal, failed rebuild preservation, concurrent writer, bulk parent transaction, and candidate count regressions - passed

Synthetic benchmark

All measurements used generated rows and a temporary synthetic SQLite archive. No KakaoTalk archive or message content was used.

• 5,000 parent windows with one child each, release build, one warm-up and seven measured loads per run
• latest upstream/main medians across three runs: 51,461 / 54,340 / 55,572 us
• feature medians across three runs: 3,382 / 3,378 / 3,406 us
• observed improvement: 15.2x to 16.3x

Privacy boundary

Tests and benchmarks use synthetic fixtures only. No real KakaoTalk database, messages, semantic documents, embeddings, or derived private data are included in this branch.

[aldegad]

Thanks for the focused indexing work. The implementation and the pinned local checks look good. Before merge, please remove docs/superpowers/plans/2026-08-10-katok-index-bulk-loader.md, rebase onto the latest main, and rerun the repository CI. The submitted change should contain the implementation, tests, and user-facing documentation only.

[aldegad]

Requested changes are complete: the repository-only plan document is gone, the branch is rebuilt directly on the latest main, and the pinned CI workflow passes. I also verified the atomic semantic generation and stale/orphan recovery contracts against the new bulk preparation path using synthetic fixtures only.