refactor(core): extract semantic vector synchronization

[phernandez]

Why

• SearchRepositoryBase still mixed backend contracts, semantic retrieval, and nearly 1,000 lines of vector synchronization orchestration.
• This is the next behavior-preserving split in Split core god files (search_repository_base, project_context, config, runtime protocols) #1108 and establishes a focused boundary before the vector indexing performance work in Indexing consolidation: unify vector planners, clarify reindex mechanisms, micro-batching #1110 changes the transaction and batching model.

What Changed

• Moved vector synchronization planning, prepare-window reads, per-entity preparation, embedding flushes, finalization, and telemetry into semantic_vector_sync.py.
• Reduced search_repository_base.py from 2,058 lines to 1,213 lines while keeping its public and backend override surface as thin delegates.
• Preserved compatibility exports for vector sync result/state types, telemetry/time monkeypatch seams, and configurable shard/window constants.
• Added focused edge-case tests for empty, deferred, failed, oversized, stale, and incomplete synchronization paths.

Implementation Details

• The new module owns the synchronization data classes and typed phase functions; SearchRepositoryBase continues to expose the existing methods so SQLite/Postgres overrides and caller imports do not change.
• Internal orchestration deliberately calls repository delegate methods for prepare, flush, finalize, logging, and persistence hooks. This retains backend specialization and existing test override behavior.
• Shard size and SQLite prepare-window limits remain re-exported through the base module and are passed into the extracted planners, preserving the old monkeypatch/configuration seam.
• This PR is intentionally behavior-preserving. It does not implement Indexing consolidation: unify vector planners, clarify reindex mechanisms, micro-batching #1110's planner unification or transaction batching; that work should build on this settled boundary.

Testing

Automated

• just fast-check: passed (ruff, formatting, and type checking; only existing Python 3.14 deprecation warnings).
• Focused semantic suite with extracted-module coverage: 135 passed, 23 skipped; semantic_vector_sync.py reached 100% line coverage.
• just fast-test: 552 passed, 2 skipped, 1,262 deselected.
• just package-check: passed all Claude Code, Codex, skill, Hermes, and OpenClaw package checks.

Manual

• just doctor: passed the end-to-end local write, index, semantic vector sync, search, and status flow.

Risks / Follow-ups

• The main risk is compatibility across the SQLite/Postgres override surface; delegates and regression tests retain those seams, and the repository-wide fast selection plus doctor flow passed.
• Indexing consolidation: unify vector planners, clarify reindex mechanisms, micro-batching #1110 should rebase on this PR before changing vector prepare transactions, planner consolidation, or embedding job coalescing.
• Further Split core god files (search_repository_base, project_context, config, runtime protocols) #1108 splits remain for semantic retrieval, project context, config, note-content protocols, and project-index workflow responsibilities.

Part of #1108.